Open-source platform for 3D AI agents. Turn text or a photo into a rigged, animated GLB avatar, give it an LLM brain, memory and a wallet, and embed it anywhere with one web component. Remote MCP server, x402 payments, Three.js, browser-native.
JavaScript 3d-generationai-agentsanimationavatardigital-humangenerative-aiglbgltfimage-to-3dllmmcpmcp-servermodel-context-protocolsolanatext-to-3dthreejsweb-componentswebglwebxrx402
docs(research): add the 2026-10 open-source landscape survey and its source notes
nirholas9a53dcf9e504
diff --git a/docs/research/oss-landscape-2026-10-notes/3d-generation.md b/docs/research/oss-landscape-2026-10-notes/3d-generation.mdnew file mode 100644index 000000000..dda05c0eb--- /dev/null+++ b/docs/research/oss-landscape-2026-10-notes/3d-generation.md@@ -0,0 +1,220 @@+# Open-source 3D generation, auto-rigging, motion, and world generation vs three.ws Forge (as of October 2026)++Scope note: all GitHub star counts below were read from the GitHub API on 2026-10-09 (`gh api repos/{owner}/{repo}`). License claims cite the license text in the repo (`gh api repos/{owner}/{repo}/license`) or the Hugging Face model card. "three.ws today" facts come from the worker READMEs in `/home/user/projects/three.ws/workers/*/README.md` (Hunyuan3D 2.0 and 2.1, TRELLIS-image-large, TripoSG, TripoSR for shape; Make-It-Animatable for rigging; MDM for text-to-motion; three Apache-2.0 video-to-motion models on the Wolf3D skeleton; LingBot-Map for video-to-scene; SDXL+ControlNet texturing; QuadriFlow/xatlas/Manifold remeshing; build123d CAD Forge; garment-forge on gemini-2.5-flash-image + Hunyuan3D 2.1).++Legal shorthand used throughout:+- **Tencent community license** (every Tencent-Hunyuan repo): "THIS LICENSE AGREEMENT DOES NOT APPLY IN THE EUROPEAN UNION, UNITED KINGDOM AND SOUTH KOREA AND IS EXPRESSLY LIMITED TO THE TERRITORY"; Territory = "the worldwide territory, excluding the territory of the European Union, United Kingdom and South Korea"; if "on the ... release date, the monthly active users of all products or services made available by or for Licensee is greater than 1 million monthly active users in the preceding calendar month, You must request a license from Tencent, which Tencent may grant to You in its sole discretion"; you may not use outputs "to improve any other AI model"; and you "must not use, reproduce, modify, distribute, or display the ... Works, Output or results ... outside the Territory". Verified verbatim in Hunyuan3D 2.0 [Source](https://github.com/Tencent-Hunyuan/Hunyuan3D-2/blob/HEAD/LICENSE), Hunyuan3D 2.1 [Source](https://huggingface.co/tencent/Hunyuan3D-2.1/raw/main/LICENSE), Hunyuan3D-Omni [Source](https://github.com/Tencent-Hunyuan/Hunyuan3D-Omni/blob/HEAD/License.txt), HY-Motion 1.0 [Source](https://github.com/Tencent-Hunyuan/HY-Motion-1.0/blob/HEAD/License.txt), and the same template in Hunyuan3D-Part, HunyuanWorld-1.0, and HY-World 2.0 (links in their sections). Consequence for an Apache-2.0 platform serving users worldwide: self-hosting is allowed only if EU/UK/South Korea users are geo-blocked from the feature AND its outputs, and the 1M-MAU test is measured once, on each model's release date.+- **Meta SAM License** (SAM 3D Objects, SAM 3D Body), "Last Updated: November 19, 2025": worldwide, no MAU threshold, commercial use allowed; requires Trade Control Law compliance, publication acknowledgement, and carrying the Agreement on redistribution. [Source](https://github.com/facebookresearch/sam-3d-objects/blob/HEAD/LICENSE)+- **Stability AI Community License** (SF3D, SPAR3D), "Last Updated: July 5, 2024": free "Research & Non-Commercial or Commercial" use only for "people or organizations generating annual revenue of less than US $1,000,000"; above that an enterprise license is required. [Source](https://github.com/Stability-AI/stable-fast-3d/blob/HEAD/LICENSE.md)+- **NVIDIA OneWay Noncommercial** (GENMO/GEM): "non-commercially means for research or evaluation purposes only". [Source](https://github.com/NVlabs/GENMO/blob/HEAD/LICENSE)++## Q1. Which models top 2026 benchmarks for text-to-3D and image-to-3D, and what do they output?++### Takeaway+There is no neutral, independent 2026 leaderboard that ranks every open model; the closest things are Meta's SA-3DAO artist-object benchmark (June 2026), vendor self-reports, and third-party blog tests. On those, the open frontier for image-to-3D is TRELLIS.2 (MIT, PBR, up to 1536^3 voxels), SAM 3D Objects (SAM License, strongest on real-world photos, splat export), Hunyuan3D 2.1 (PBR, territorial license), and newer MIT entrants Pixal3D (SIGGRAPH 2026) and PartCrafter (part-level meshes). three.ws already runs Hunyuan3D 2.1 and TRELLIS 1 but not TRELLIS.2, SAM 3D Objects, Pixal3D, or any part-level generator.++### Cited Findings++**Tencent Hunyuan3D family (three.ws runs 2.0 and 2.1 today)**+- Hunyuan3D-2: 15,037 stars (2026-10-09), last push 2025-10-28, license "TENCENT HUNYUAN 3D 2.0 COMMUNITY LICENSE AGREEMENT", release date January 21, 2025, EU/UK/South Korea excluded, 1M MAU gate. [Source](https://github.com/Tencent-Hunyuan/Hunyuan3D-2/blob/HEAD/LICENSE)+- Hunyuan3D-2.1: 4,147 stars (2026-10-09); "Tencent Hunyuan 3D 2.1 Community License Agreement", release date June 13, 2025, same Territory and MAU terms, Notice file required for non-hosted distribution, outputs may not be used to improve other models. [Source](https://huggingface.co/tencent/Hunyuan3D-2.1/raw/main/LICENSE)+- Hunyuan3D-2.1 outputs PBR (base color, metallic-roughness, normal) and the README states VRAM of 10 GB shape-only, 21 GB texture-only, 29 GB for both; self-reported Hunyuan3D-Paint-2.1 CLIP-FID 24.78 vs TEXGen 28.24. [Source](https://github.com/Tencent-Hunyuan/Hunyuan3D-2.1)+- Conflict to flag: source headers under `hy3dpaint/` state the module is for non-commercial use while the repository LICENSE is the community license permitting commercial use inside the Territory. [Source](https://github.com/Tencent-Hunyuan/Hunyuan3D-2.1/tree/main/hy3dpaint); contradicted by [Source](https://huggingface.co/tencent/Hunyuan3D-2.1/raw/main/LICENSE)+- Hunyuan3D-Part: 546 stars (2026-10-09), "Tencent Hunyuan 3D-Part Community License", release date September 23, 2025, same Territory exclusion; contains P3-SAM (part segmentation) and X-Part (part generation, light version released). [Source](https://github.com/Tencent-Hunyuan/Hunyuan3D-Part)+- Hunyuan3D-Omni: 639 stars (2026-10-09), "TENCENT HUNYUAN 3D Omni COMMUNITY LICENSE AGREEMENT", release date September 26, 2025, EU/UK/South Korea excluded, 1M MAU gate; adds controllable generation (point cloud, voxel, bounding box, pose conditions). [Source](https://github.com/Tencent-Hunyuan/Hunyuan3D-Omni/blob/HEAD/License.txt)+- Hunyuan3D 2.5, 3.0 and 3.1 exist only as hosted services (3d.hunyuan.tencent.com); no official open repo was found for them. [Source](https://3d.hunyuan.tencent.com/)++**Microsoft TRELLIS and TRELLIS.2 (three.ws runs TRELLIS-image-large today)**+- microsoft/TRELLIS: 13,778 stars (2026-10-09), MIT. [Source](https://github.com/microsoft/TRELLIS/blob/HEAD/LICENSE)+- microsoft/TRELLIS.2: 11,509 stars (2026-10-09), MIT code; weights `microsoft/TRELLIS.2-4B` are MIT on Hugging Face. [Source](https://github.com/microsoft/TRELLIS.2/blob/HEAD/LICENSE); [Source](https://huggingface.co/microsoft/TRELLIS.2-4B)+- TRELLIS.2 is a 4B-parameter image-to-3D model using the O-Voxel representation, outputs PBR Base Color / Roughness / Metallic / Opacity, supports 512^3, 1024^3 and 1536^3 resolutions with H100 timings of about 3 s / 17 s / 60 s, and needs a minimum 24 GB VRAM GPU; it is image-to-3D only (no text conditioning). [Source](https://github.com/microsoft/TRELLIS.2); paper [Source](https://arxiv.org/abs/2512.14692)+- A third-party test on AMD hardware measured TRELLIS.2 peak VRAM at 8.65 GiB versus Hunyuan3D-2.1 at 28.8 GiB when generating a 40k-face mesh. [Source](https://hawkymisc.github.io/)+- Meta's AssetGen paper notes TRELLIS.2 is "less robust on complex inputs" than their system (vendor comparison, not independent). [Source](https://ai.meta.com/research/)++**Meta SAM 3D Objects and SAM 3D Body**+- facebookresearch/sam-3d-objects: 7,496 stars (2026-10-09), SAM License (worldwide commercial, no MAU gate; Trade Control and attribution clauses). [Source](https://github.com/facebookresearch/sam-3d-objects/blob/HEAD/LICENSE)+- SAM 3D Objects reconstructs shape, texture and scene layout from a single real-world photo and exports Gaussian splats; Meta reports a 5:1 human-preference win rate over prior methods; the SA-3DAO benchmark of 1,000 artist-made objects and a Hugging Face leaderboard were published June 2, 2026, and the encoder weights on June 1, 2026. [Source](https://github.com/facebookresearch/sam-3d-objects); [Source](https://huggingface.co/facebook/sam-3d-objects)+- facebookresearch/sam-3d-body: 3,602 stars (2026-10-09), SAM License; outputs the Meta Human Rig (MHR) parametric body from a single image; CVPR 2026. [Source](https://github.com/facebookresearch/sam-3d-body)++**Other MIT / Apache image-to-3D models**+- VAST-AI-Research/TripoSG: 1,822 stars (2026-10-09), MIT (already used by three.ws). [Source](https://github.com/VAST-AI-Research/TripoSG/blob/HEAD/LICENSE)+- VAST-AI-Research/TripoSR: 7,032 stars (2026-10-09), MIT (already used). [Source](https://github.com/VAST-AI-Research/TripoSR/blob/HEAD/LICENSE)+- stepfun-ai/Step1X-3D: 896 stars (2026-10-09), Apache-2.0; two-stage geometry + texture, README lists 27 to 29 GB VRAM across both stages. [Source](https://github.com/stepfun-ai/Step1X-3D/blob/HEAD/LICENSE); [Source](https://github.com/stepfun-ai/Step1X-3D)+- Stable-X/Stable3DGen (Hi3DGen, repo moved from Stable-X/Hi3DGen): 1,279 stars (2026-10-09), MIT (copyright Bytedance Inc.); normal-bridged image-to-3D fine-tuned from TRELLIS with the StableNormal estimator. [Source](https://github.com/Stable-X/Stable3DGen/blob/HEAD/LICENSE); [Source](https://github.com/Stable-X/Stable3DGen)+- DreamTechAI/Direct3D-S2: 1,285 stars (2026-10-09), MIT (copyright DreamTechAI); NeurIPS 2025; README lists 10 GB VRAM at 512 resolution and 24 GB at 1024. [Source](https://github.com/DreamTechAI/Direct3D-S2/blob/HEAD/LICENSE.txt); [Source](https://github.com/DreamTechAI/Direct3D-S2)+- wgsxm/PartCrafter: 2,488 stars (2026-10-09), MIT, last push 2026-04-16; single image to multiple semantically separate part meshes (structured, decomposable output). [Source](https://github.com/wgsxm/PartCrafter/blob/HEAD/LICENSE); [Source](https://github.com/wgsxm/PartCrafter)+- TencentARC/Pixal3D: 2,435 stars (2026-10-09), MIT (copyright 2026 Tencent), last push 2026-09-01, SIGGRAPH 2026; pixel-aligned image-to-3D with textured GLB/OBJ output. [Source](https://github.com/TencentARC/Pixal3D/blob/HEAD/LICENSE); [Source](https://github.com/TencentARC/Pixal3D)+- PKU-YuanGroup/UltraShape-1.0: 859 stars (2026-10-09), GitHub reports NOASSERTION; a refinement DiT initialized from Hunyuan3D-2.1 (so it inherits Tencent terms through the base weights); weights at `infinith/UltraShape`. [Source](https://github.com/PKU-YuanGroup/UltraShape-1.0); [Source](https://huggingface.co/infinith/UltraShape)+- lizhihao6/Sparc3D: 1,358 stars (2026-10-09), no license file, no weights in the README; the commercial successor (Hitem3D, 2048^3) is closed. [Source](https://github.com/lizhihao6/Sparc3D)+- Stability-AI/stable-fast-3d: 1,843 stars (2026-10-09) and Stability-AI/stable-point-aware-3d: 1,078 stars (2026-10-09), both under the Stability AI Community License (free commercial use only under US $1M annual revenue). [Source](https://github.com/Stability-AI/stable-fast-3d/blob/HEAD/LICENSE.md); [Source](https://github.com/Stability-AI/stable-point-aware-3d/blob/HEAD/LICENSE.md)++**Texturing**+- huanngzh/MV-Adapter: 1,295 stars (2026-10-09), Apache-2.0; multi-view consistent image generation for texturing existing meshes. [Source](https://github.com/huanngzh/MV-Adapter/blob/HEAD/LICENSE)+- CVMI-Lab/TEXGen: 340 stars (2026-10-09), no license file in the repository. [Source](https://github.com/CVMI-Lab/TEXGen)++### Inferences+- For an image-to-3D upgrade with zero legal friction, TRELLIS.2 (MIT, PBR, 24 GB class GPU, 3 to 60 s on H100) is the obvious add: it beats the current TRELLIS 1 lane on resolution and PBR, and a 1024^3 run at 17 s is competitive with the current Hunyuan3D 2.1 lane while using far less VRAM in third-party tests. Verdict: adopt code and weights.+- SAM 3D Objects fills a gap three.ws does not have (photo-of-a-real-object to textured mesh plus layout plus splats) under a worldwide commercial license. Verdict: adopt code and weights; note the SAM License redistribution and Trade Control clauses in the worker README.+- Hunyuan3D 2.1 remains the best open PBR texture painter by its own numbers, but it is the lane that forces EU/UK/SK geo-blocking. Verdict: already covered, but the platform needs a documented geo-gate on that lane (see Q2).+- PartCrafter and Hunyuan3D-Part give a "parts" capability (separable, editable sub-meshes) that commercial rivals (Tripo segmentation, Rodin) sell and three.ws lacks. PartCrafter is MIT; Hunyuan3D-Part is territorial. Verdict: adopt PartCrafter code and weights first.+- Pixal3D (MIT, 2026-09 push, SIGGRAPH 2026) is the freshest MIT image-to-3D; worth a bake-off against TRELLIS.2 before adopting.+- Text-to-3D (not image-to-3D) remains a two-step problem for every open model listed: TRELLIS.2, SAM 3D, Pixal3D, PartCrafter and Direct3D-S2 are image-conditioned, so a text prompt still passes through an image generator first (three.ws already does this with gemini-2.5-flash-image in garment-forge).++### Gaps+- No independent, peer-reviewed 2026 leaderboard ranking TRELLIS.2, Hunyuan3D 2.1, SAM 3D Objects, Pixal3D and Direct3D-S2 head to head was found; SA-3DAO is Meta-run and the other numbers are vendor self-reports.+- Pixal3D VRAM (about 8 to 12 GB) and speed (30 to 60 s) and a claim that it is "academic only" came from secondary blog sources that conflict with the MIT LICENSE file; treat the license file as authoritative and the perf numbers as unverified.+- UltraShape's license: GitHub reports NOASSERTION; a wiki-style page claims Apache-2.0; not verified against a license file.+- Roblox Cube is reported as research-only and QuadGPT (arXiv 2509.21420, quad-mesh generation) has no confirmed code release; neither was verified from a license file.+- Whether the `hy3dpaint` non-commercial header in Hunyuan3D-2.1 is a stale artifact or a real carve-out needs a question to Tencent or a legal read; it was not resolvable from public text.++## Q2. Which can three.ws legally self-host commercially worldwide, and which carry territorial or MAU restrictions?++### Takeaway+Worldwide-commercial without conditions: every MIT/Apache repo (TRELLIS, TRELLIS.2, TripoSG, TripoSR, Step1X-3D, Stable3DGen, Direct3D-S2, PartCrafter, Pixal3D, MV-Adapter, UniRig, SkinTokens, Make-It-Animatable, MagicArticulate, Puppeteer, MoMask, MotionGPT, Matrix-3D, Matrix-Game, WorldGen code, gsplat, Spark, GaussianSplats3D, lingbot-map) plus Meta's SAM License models. Every Tencent model (Hunyuan3D 2.0/2.1/Part/Omni, HY-Motion 1.0, HunyuanWorld 1.0, HY-World 2.0) excludes EU, UK and South Korea and carries a 1M-MAU gate. Non-commercial and therefore unusable in production: GVHMR, GENMO/GEM, Lyra 2.0 weights, original Inria 3DGS, SceneScript, and Stability models above US $1M revenue.++### Cited Findings+- Tencent community license Territory clause and 1M MAU clause (verbatim quotes in the legal shorthand above). [Source](https://github.com/Tencent-Hunyuan/Hunyuan3D-2/blob/HEAD/LICENSE); [Source](https://huggingface.co/tencent/Hunyuan3D-2.1/raw/main/LICENSE); [Source](https://github.com/Tencent-Hunyuan/Hunyuan3D-Omni/blob/HEAD/License.txt); [Source](https://github.com/Tencent-Hunyuan/HY-Motion-1.0/blob/HEAD/License.txt)+- Tencent licenses also forbid using outputs "to improve any other AI model" and forbid displaying outputs outside the Territory, which binds the platform's asset gallery and marketplace, not just the generation request. [Source](https://github.com/Tencent-Hunyuan/Hunyuan3D-2/blob/HEAD/LICENSE)+- SAM License: Meta grants a worldwide, royalty-free license; conditions include Trade Control Law compliance and research-publication acknowledgement; no MAU threshold appears in the text. [Source](https://github.com/facebookresearch/sam-3d-objects/blob/HEAD/LICENSE)+- Stability AI Community License: free commercial use only for organizations under US $1,000,000 annual revenue. [Source](https://github.com/Stability-AI/stable-fast-3d/blob/HEAD/LICENSE.md)+- NVIDIA OneWay Noncommercial (GENMO): research or evaluation only. [Source](https://github.com/NVlabs/GENMO/blob/HEAD/LICENSE)+- zju3dv/GVHMR: 2,087 stars (2026-10-09), license text restricts use to non-commercial research; commercial use requires contacting the authors. [Source](https://github.com/zju3dv/GVHMR/blob/HEAD/LICENSE)+- graphdeco-inria/gaussian-splatting: 24,145 stars (2026-10-09), Inria/MPII license for research and evaluation only. [Source](https://github.com/graphdeco-inria/gaussian-splatting/blob/HEAD/LICENSE.md)+- facebookresearch/scenescript: 385 stars (2026-10-09), CC-BY-NC-4.0. [Source](https://github.com/facebookresearch/scenescript/blob/HEAD/LICENSE)+- nv-tlabs/lyra: 2,647 stars (2026-10-09); README: "Lyra source code is released under the Apache 2.0 License. Please refer to Lyra-1 and Lyra-2 for their respective model licenses." The Lyra-2.0 model card places the weights under the NVIDIA Internal Scientific Research and Development Model License (no production or commercial use). [Source](https://github.com/nv-tlabs/lyra); [Source](https://huggingface.co/nvidia/Lyra-2.0)+- SMPL/SMPL-X body models (used by HY-Motion, GVHMR, MoMask, MotionGPT) are distributed through MPI registration and cannot be redistributed; HY-Motion's README credits SMPL/SMPL-H. [Source](https://smpl.is.tue.mpg.de/); [Source](https://github.com/Tencent-Hunyuan/HY-Motion-1.0)+- MIT license texts verified for TRELLIS.2, Direct3D-S2, Pixal3D, Stable3DGen, PartCrafter, UniRig, SkinTokens, Make-It-Animatable, Matrix-Game (README: "This project is licensed under the MIT License"), momask-codes, MotionGPT, Spark, GaussianSplats3D. [Source](https://github.com/microsoft/TRELLIS.2/blob/HEAD/LICENSE); [Source](https://github.com/DreamTechAI/Direct3D-S2/blob/HEAD/LICENSE.txt); [Source](https://github.com/TencentARC/Pixal3D/blob/HEAD/LICENSE); [Source](https://github.com/Stable-X/Stable3DGen/blob/HEAD/LICENSE); [Source](https://github.com/wgsxm/PartCrafter/blob/HEAD/LICENSE); [Source](https://github.com/VAST-AI-Research/UniRig/blob/HEAD/LICENSE); [Source](https://github.com/VAST-AI-Research/SkinTokens/blob/HEAD/LICENSE); [Source](https://github.com/jasongzy/Make-It-Animatable/blob/HEAD/LICENSE); [Source](https://github.com/SkyworkAI/Matrix-Game/blob/HEAD/LICENSE); [Source](https://github.com/EricGuo5513/momask-codes/blob/HEAD/LICENSE); [Source](https://github.com/OpenMotionLab/MotionGPT/blob/HEAD/LICENSE); [Source](https://github.com/sparkjsdev/spark/blob/HEAD/LICENSE); [Source](https://github.com/mkkellogg/GaussianSplats3D/blob/HEAD/LICENSE)+- Apache-2.0 verified for Step1X-3D, MV-Adapter, MagicArticulate, Puppeteer, WorldGen, gsplat, lingbot-map, Lyra code. [Source](https://github.com/stepfun-ai/Step1X-3D/blob/HEAD/LICENSE); [Source](https://github.com/huanngzh/MV-Adapter/blob/HEAD/LICENSE); [Source](https://github.com/Seed3D/MagicArticulate/blob/HEAD/LICENSE); [Source](https://github.com/Seed3D/Puppeteer/blob/HEAD/LICENSE); [Source](https://github.com/ZiYang-xie/WorldGen/blob/HEAD/LICENSE); [Source](https://github.com/nerfstudio-project/gsplat/blob/HEAD/LICENSE); [Source](https://github.com/Robbyant/lingbot-map/blob/HEAD/LICENSE); [Source](https://github.com/nv-tlabs/lyra/blob/HEAD/LICENSE)+- WorldGen depends on the gated FLUX.1-dev model ("You should also accept the license of the gated model (FLUX.1-dev)"), whose own license is non-commercial for the weights. [Source](https://github.com/ZiYang-xie/WorldGen); [Source](https://huggingface.co/black-forest-labs/FLUX.1-dev)++### Inferences+- three.ws's live Hunyuan3D 2.0/2.1 lanes are out of compliance for any EU, UK or South Korean user unless a geo-gate exists; the license also forbids displaying those outputs there, so galleries and marketplace listings of Hunyuan-generated assets need region handling or a lane switch to TRELLIS.2 / SAM 3D for those users. This is the single most consequential legal finding of this sheet.+- The 1M-MAU clause is measured on each model's release date (Jan 21, 2025 for 2.0; June 13, 2025 for 2.1), so a platform that was under 1M MAU then stays licensed for those versions even after it grows; each new Tencent release re-tests.+- A fully MIT/Apache/SAM stack covering shape (TRELLIS.2, SAM 3D Objects, Pixal3D, PartCrafter), rigging (SkinTokens, Make-It-Animatable), motion (MDM already, MoMask, MotionGPT), and worlds (Matrix-3D, WorldGen code with a non-FLUX panorama model, lingbot-map) is achievable today with no territorial gating.+- Any motion lane built on SMPL-family skeletons must retarget to the platform's Wolf3D/Mixamo skeleton inside the worker and never redistribute SMPL assets; three.ws's existing video2motion workers already follow this pattern.++### Gaps+- Whether three.ws currently geo-blocks EU/UK/SK on the Hunyuan lanes was not checked (workers directory read-only; routing code not inspected for this task).+- The Stability license's revenue test applies to "You" (the licensee); whether three.ws's revenue is under US $1M is a business fact not in scope.+- Hunyuan3D-Part and HunyuanWorld-1.0 license texts were confirmed to use the same community template via README and license headers, but their full texts were not read line by line; MAU wording assumed identical.++## Q3. Best open auto-rigger and best open motion generator, and what skeleton formats they emit++### Takeaway+Rigging: VAST's SkinTokens/TokenRig (MIT, Feb 2026) is the state of the art for arbitrary (including non-humanoid) meshes and is the successor to UniRig, which three.ws retired in favor of Make-It-Animatable; Puppeteer and MagicArticulate (Apache-2.0) are the ByteDance Seed alternatives. Motion: HY-Motion 1.0 (Dec 2025) is the largest open text-to-motion model (1.0B, SMPL-H skeleton) but carries Tencent's territorial license; MIT choices remain MDM (already used), MoMask and MotionGPT.++### Cited Findings++**Auto-rigging**+- VAST-AI-Research/SkinTokens: 463 stars (2026-10-09), MIT, last push 2026-05-12, arXiv 2602.04805 (Feb 2026). TokenRig "models the entire rig, i.e., skeleton and skinning weights, as a single token sequence" on a Qwen3-0.6B transformer with GRPO refinement; README claims "98%-133%" improvement in skinning accuracy over UniRig; needs "at least 14 GB" GPU memory; demo input/output is GLB (`demo.py --input examples/giraffe.glb --output results/giraffe.glb`), with `--use_skeleton` to skin an existing skeleton; trained on ArticulationXL 2.0 + VRoid Hub + ModelsResource. [Source](https://github.com/VAST-AI-Research/SkinTokens); weights [Source](https://huggingface.co/VAST-AI/SkinTokens)+- VAST-AI-Research/UniRig: 1,804 stars (2026-10-09), MIT; accepts obj/fbx/glb/vrm, outputs FBX skeleton and skinning plus merged GLB, 8 GB VRAM, Articulation-XL2.0 checkpoint. [Source](https://github.com/VAST-AI-Research/UniRig)+- Seed3D/Puppeteer: 427 stars (2026-10-09), Apache-2.0, NeurIPS 2025 spotlight; skeleton + skinning + video-guided animation; FBX export requires `bpy` 4.2.0. [Source](https://github.com/Seed3D/Puppeteer)+- Seed3D/MagicArticulate: 417 stars (2026-10-09), Apache-2.0; skeleton generation (no skinning) from meshes. [Source](https://github.com/Seed3D/MagicArticulate)+- jasongzy/Make-It-Animatable: 463 stars (2026-10-09), MIT; humanoid-focused, already three.ws's rig worker (Mixamo skeleton + ARKit-52 blendshapes via ICT-FaceKit per the worker README). [Source](https://github.com/jasongzy/Make-It-Animatable)+- The Stroke3D paper reports that UniRig and MagicArticulate are weaker on animals and plants than on humanoids. [Source](https://arxiv.org/abs/2509.00000)+- SkinTokens' README shows a qualitative comparison "TokenRig vs. Puppeteer vs. UniRig" where baselines "suffer from bleeding artifacts across disconnected mesh parts". [Source](https://github.com/VAST-AI-Research/SkinTokens)++**Motion generation**+- Tencent-Hunyuan/HY-Motion-1.0: 2,586 stars (2026-10-09); "TENCENT HY-MOTION 1.0 COMMUNITY LICENSE AGREEMENT", release date December 30, 2025, EU/UK/SK excluded, 1M MAU gate. [Source](https://github.com/Tencent-Hunyuan/HY-Motion-1.0/blob/HEAD/License.txt)+- HY-Motion 1.0 is "a series of text-to-3D human motion generation models based on Diffusion Transformer (DiT) and Flow Matching" producing "skeleton-based 3D character animations"; the README credits SMPL/SMPL-H, i.e. the output skeleton is SMPL-H; 1.0B and 0.46B-Lite variants; Jan 29, 2026 release of the SSAE evaluation code. [Source](https://github.com/Tencent-Hunyuan/HY-Motion-1.0)+- EricGuo5513/momask-codes: 1,327 stars (2026-10-09), MIT; HumanML3D 22-joint representation. [Source](https://github.com/EricGuo5513/momask-codes)+- OpenMotionLab/MotionGPT: 1,975 stars (2026-10-09), MIT. [Source](https://github.com/OpenMotionLab/MotionGPT)+- NVlabs/GENMO (renamed GEM): 522 stars (2026-10-09), NVIDIA OneWay Noncommercial. [Source](https://github.com/NVlabs/GENMO/blob/HEAD/LICENSE)+- zju3dv/GVHMR: 2,087 stars (2026-10-09), non-commercial research license. [Source](https://github.com/zju3dv/GVHMR/blob/HEAD/LICENSE)+- facebookresearch/sam-3d-body emits the MHR (Meta Human Rig) parametric body, an alternative to SMPL that Meta distributes under the SAM License rather than MPI's registration terms. [Source](https://github.com/facebookresearch/sam-3d-body)++### Inferences+- SkinTokens is the right upgrade for the "rig anything" gap (quadrupeds, creatures, props with articulation) that Meshy and Tripo advertise and three.ws's humanoid-only Make-It-Animatable lane cannot serve; MIT, GLB in/out, 14 GB VRAM fits existing L4/A10-class workers. Verdict: adopt code and weights as a second rig lane behind a "humanoid vs non-humanoid" classifier, keep Make-It-Animatable for humanoids (it gives Mixamo bones + blendshapes, which SkinTokens does not).+- HY-Motion 1.0 would be a quality jump over MDM but it brings the Tencent geo-gate plus SMPL-H retargeting; the 0.46B-Lite still needs about 24 GB VRAM per the README. Verdict: adopt design (DiT + flow matching, large pretrain) only; keep MDM/MoMask for a worldwide lane, or run HY-Motion only for non-EU/UK/SK users.+- Skeleton formats: HY-Motion, MoMask, MotionGPT and GVHMR all emit SMPL-family joint rotations; UniRig/Puppeteer emit FBX or GLB with arbitrary predicted skeletons; SkinTokens emits GLB with a predicted skeleton; Make-It-Animatable emits a Mixamo-compatible rig. Any lane added needs a retarget step to three.ws's canonical Wolf3D/Mixamo skeleton in `src/glb-canonicalize.js` terms.++### Gaps+- Independent benchmark numbers (e.g. on Articulation-XL test split) comparing SkinTokens vs Puppeteer vs Make-It-Animatable were not found outside the SkinTokens paper.+- HY-Motion's exact export format (BVH vs SMPL npz vs FBX) was not confirmed from the README grep; only the SMPL/SMPL-H credit and "skeleton-based" wording were captured.+- MotionGPT and MoMask have had no 2026 releases found; their current state may be unmaintained.++## Q4. Open world and scene generation models with usable licenses vs a text-to-3D-world feature++### Takeaway+Tencent HY-World 2.0 (April to May 2026) is the most capable open text/image-to-3D-world system, emitting meshes, 3DGS and point clouds importable into engines, but it is territorial and its full pipeline is about 100B parameters across stages. Worldwide-commercial alternatives are Skywork Matrix-3D (MIT, panorama video to scene) and WorldGen (Apache-2.0 code, but FLUX.1-dev dependency); NVIDIA Lyra 2.0 is Apache code with non-commercial weights. three.ws's Diorama (text to 3D world) and LingBot-Map (video to scene) already cover part of this.++### Cited Findings+- Tencent-Hunyuan/HY-World-2.0: 2,705 stars (2026-10-09), Tencent community license (release date April 15, 2026, same Territory and MAU terms). README news: April 16, 2026 tech report and WorldMirror 2.0 weights; May 11, 2026 HY-Pano 2.0 code and weights; May 18, 2026 World Generation inference code and WorldStereo 2.0 weights; "July, 2026: Update HY World 2.1! Try our product" (hosted only). [Source](https://github.com/Tencent-Hunyuan/HY-World-2.0); license [Source](https://github.com/Tencent-Hunyuan/HY-World-2.0/blob/HEAD/License.txt)+- HY-World 2.0 "produces 3D world representations (meshes / Gaussian Splattings)" from "text, single-view images, multi-view images, and videos"; outputs are "3DGS, meshes, and point clouds ... directly importable into Unity / Unreal Engine / Isaac"; pipeline is HY-Pano-2.0 (text/image to 360 panorama, ~80B, plus a ~425M Qwen LoRA variant) to WorldNav trajectory planning to WorldStereo 2.0 (panorama to 3DGS, ~17B) to WorldMirror 2.0 (multi-view to 3D, ~1.2B); README recommends FlashAttention-3 on Hopper and supports multi-GPU rendering. [Source](https://github.com/Tencent-Hunyuan/HY-World-2.0); weights [Source](https://huggingface.co/tencent/HY-World-2.0)+- Tencent-Hunyuan/HunyuanWorld-1.0: 2,943 stars (2026-10-09), community license, release date July 27, 2025, EU/UK/SK excluded. [Source](https://github.com/Tencent-Hunyuan/HunyuanWorld-1.0/blob/HEAD/LICENSE)+- Tencent-Hunyuan/HunyuanWorld-Voyager: 1,600 stars (2026-10-09), NOASSERTION (Tencent community). [Source](https://github.com/Tencent-Hunyuan/HunyuanWorld-Voyager)+- SkyworkAI/Matrix-3D: 844 stars (2026-10-09), MIT; text or image to panorama video to 3D scene. [Source](https://github.com/SkyworkAI/Matrix-3D/blob/HEAD/LICENSE); [Source](https://github.com/SkyworkAI/Matrix-3D)+- SkyworkAI/Matrix-Game: 2,346 stars (2026-10-09), MIT ("This project is licensed under the MIT License"), last push 2026-09-29; Matrix-Game-3.0 released March 27, 2026 as "a real-time and streaming interactive world model with long-horizon Memory" (video world model, not a mesh generator). [Source](https://github.com/SkyworkAI/Matrix-Game)+- ZiYang-xie/WorldGen: 2,154 stars (2026-10-09), Apache-2.0; text-to-scene and image-to-scene in seconds, splat or mesh output (`--return_mesh`), low-VRAM mode about 10 GB (Oct 5, 2025), ml-sharp support for better GS (Jan 10, 2026); requires accepting the gated FLUX.1-dev license. [Source](https://github.com/ZiYang-xie/WorldGen); license [Source](https://github.com/ZiYang-xie/WorldGen/blob/HEAD/LICENSE)+- nv-tlabs/lyra: 2,647 stars (2026-10-09), Apache-2.0 code; Lyra 2.0 released April 15, 2026 ("Explorable generative 3D worlds with long-horizon, 3D-consistent generation"), GUI and training code July 20, 2026, arXiv 2604.13036, ICLR 2026 and SIGGRAPH Asia 2026; weights under NVIDIA's non-production research license; outputs 3DGS .ply, built on Wan2.1-14B. [Source](https://github.com/nv-tlabs/lyra); [Source](https://huggingface.co/nvidia/Lyra-2.0)+- facebookresearch/scenescript: 385 stars (2026-10-09), CC-BY-NC-4.0 (structured scene language, not usable commercially). [Source](https://github.com/facebookresearch/scenescript/blob/HEAD/LICENSE)+- Robbyant/lingbot-map: 17,505 stars (2026-10-09), Apache-2.0 (already three.ws's video-to-scene worker). [Source](https://github.com/Robbyant/lingbot-map/blob/HEAD/LICENSE)+- World Labs Marble is proprietary; commercial rights are sold on a paid tier. [Source](https://marble.worldlabs.ai/)+- Splat and glTF tooling: nerfstudio-project/gsplat 5,762 stars (2026-10-09), Apache-2.0 [Source](https://github.com/nerfstudio-project/gsplat/blob/HEAD/LICENSE); sparkjsdev/spark 3,705 stars (2026-10-09), MIT, Spark 2.0 (April 14, 2026) renders 100M+ splats on WebGL2 and loads PLY/SPZ/SPLAT/KSPLAT/SOG [Source](https://github.com/sparkjsdev/spark); mkkellogg/GaussianSplats3D 2,905 stars (2026-10-09), MIT [Source](https://github.com/mkkellogg/GaussianSplats3D/blob/HEAD/LICENSE); donmccurdy/glTF-Transform 1,983 stars (2026-10-09), MIT [Source](https://github.com/donmccurdy/glTF-Transform/blob/HEAD/LICENSE); zeux/meshoptimizer (gltfpack) 8,524 stars (2026-10-09), MIT [Source](https://github.com/zeux/meshoptimizer/blob/HEAD/LICENSE.md); ahujasid/blender-mcp 30,273 stars (2026-10-09), MIT [Source](https://github.com/ahujasid/blender-mcp/blob/HEAD/LICENSE).++### Inferences+- HY-World 2.0 is the only open system that outputs engine-ready meshes plus 3DGS plus navmesh-planned trajectories from a single prompt; it is what a "text to explorable 3D world" feature should be benchmarked against. But the ~80B HY-Pano-2.0 stage and the ~17B WorldStereo 2.0 stage need multi-GPU Hopper hardware, and the territorial license applies to the outputs' display. Verdict: adopt design, rebuild on worldwide-licensed parts (a permissive panorama model, then Matrix-3D or WorldGen-style lifting, then lingbot-map reconstruction).+- WorldGen's Apache-2.0 code is usable but its FLUX.1-dev dependency is non-commercial; swapping the panorama generator for a GCP-hosted Imagen or an Apache-licensed diffusion model is required before production. Verdict: adopt code, replace the panorama model.+- Matrix-3D (MIT) is the cleanest fully open text/image to 3D scene path; quality claims are unverified, so run a bake-off against the Diorama lane.+- Lyra 2.0 is a reference for long-horizon 3D-consistent generation but cannot ship (weights non-commercial). Verdict: adopt design only.+- Spark 2.0 (MIT) matches or exceeds what the three.ws Splat Viewer needs (100M splats, SOG/SPZ formats); if the viewer is still on GaussianSplats3D, upgrading is a low-risk win. Verdict: adopt code.++### Gaps+- HY-World 2.0's exact per-stage VRAM and wall-clock were not in the README excerpt captured; "multi-GPU recommended" came from a secondary source.+- Matrix-3D output quality, resolution and runtime were not captured from primary text.+- Which splat library the three.ws Splat Viewer currently uses was not checked in this task.+- WorldGen's splat quality vs mesh quality and its panorama-model swap feasibility were not tested.++## Q5. Commercial-only competitors (Meshy, Tripo, Rodin/Hyper3D, Luma, Kaedim, Mixamo): features three.ws lacks++### Takeaway+The commercial leaders have converged on the same bundle: sub-15-second textured generation, quad retopology at generation time, PBR up to 8K, built-in auto-rig for humanoids AND non-humanoids with hundreds of animation presets, segmentation/decimate/remesh/UV API endpoints, multiple-image conditioning, and MCP servers plus engine plugins. three.ws has pieces (rig, remesh, PBR via Hunyuan3D 2.1) but not the non-humanoid rig, animation preset library, quad topology at generation, part segmentation, or multi-image input.++### Cited Findings+- Meshy: the Meshy-6 generation (January 2026) adds quad or triangle topology selection at generation time, A/T-pose output, auto-rigging for humanoid and quadruped characters with a library of several hundred animation presets, FBX/GLB export, Remesh/Convert/Resize/UV-unwrap API endpoints (2026), 8K base color textures (July 2026), PBR including emission, printability checks, and an MCP server. [Source](https://docs.meshy.ai/en/changelog)+- Tripo 3.1: about 10 s to a textured model; Rigging 1.0 for bipeds and Rigging 2.5 for non-bipeds (quadruped, hexapod, avian, serpentine, aquatic), retarget presets, segmentation v1/v2, Smart Mesh P1.0 quad retopology in about 2 s; API endpoints `/v3/mesh/segment`, `/v3/mesh/decimate`, `/v3/animations/rig`, rig-check and retarget; CLI (June 2026), MCP server and engine plugins. [Source](https://platform.tripo3d.ai/docs)+- Rodin Gen-2 (Hyper3D): 10B-parameter BANG model, up to 5 input images, quad output (1k to 200k faces) or raw triangle (500 to 1M), PBR or shaded textures, bounding-box control, T/A pose, GLB/FBX/OBJ/USDZ/STL export; Gen-2.5 tiers. [Source](https://hyper3d.ai/)+- Luma Genie: four quad-mesh candidates per prompt, FBX/GLTF/USDZ/Blend/STL/OBJ export. [Source](https://lumalabs.ai/genie)+- Mixamo: free, Adobe-hosted humanoid auto-rig and animation library; no feature updates found in 2026 (competitors position against it as stagnant). [Source](https://www.mixamo.com/)++### Inferences+- The highest-leverage gaps for three.ws, in order: (1) non-humanoid auto-rig plus a retarget preset library (SkinTokens + the existing clip library retargeter in `src/animation-retarget.js`), (2) part segmentation and part-level generation (PartCrafter, Hunyuan3D-Part P3-SAM), (3) quad topology at generation time (QuadriFlow exists in the remesh worker; expose it as a generation option and benchmark against Tripo's 2 s Smart Mesh), (4) multi-image conditioning (TRELLIS.2 and Hunyuan3D 2.1 support multi-view inputs; Rodin accepts 5), (5) a public mesh-ops API surface (segment, decimate, remesh, UV) that mirrors Tripo and Meshy endpoints, since three.ws already has the workers.+- Speed: Tripo's ~10 s textured and Rodin's quad output are set by proprietary models; the open path to similar latency is TRELLIS.2 at 512^3 (3 s on H100) plus the existing PBR lane, which three.ws can host on GCP credits.++### Gaps+- Kaedim: no reliable primary source describing 2026 features was found; only marketing copy for image-to-3D exists, so it is not characterized here.+- Meshy and Tripo pricing, rate limits and output license terms were not captured.+- Luma Genie's 2026 status (whether still maintained after Luma's video focus) was not confirmed.+- Rodin Gen-2.5 tier specifics were not captured from a primary page.++## Per-repo verdict table++| Repo | Stars (2026-10-09) | Code license | Weights license | Worldwide commercial? | Standout vs three.ws | Verdict |+|---|---|---|---|---|---|---|+| microsoft/TRELLIS.2 | 11,509 | MIT | MIT (HF) | Yes | 4B O-Voxel, PBR, 1536^3, 3-60 s H100 | Adopt code+weights |+| facebookresearch/sam-3d-objects | 7,496 | SAM License | SAM License | Yes (Trade Control clauses) | Real-photo shape+texture+layout, splats, SA-3DAO | Adopt code+weights |+| facebookresearch/sam-3d-body | 3,602 | SAM License | SAM License | Yes | MHR body from one image | Adopt for avatar reconstruction |+| TencentARC/Pixal3D | 2,435 | MIT | MIT | Yes | Pixel-aligned, SIGGRAPH 2026, fresh | Bake-off then adopt |+| wgsxm/PartCrafter | 2,488 | MIT | MIT | Yes | Part-level meshes from one image | Adopt code+weights |+| Tencent-Hunyuan/Hunyuan3D-2 / 2.1 | 15,037 / 4,147 | Tencent community | Tencent community | No (EU/UK/SK excluded, 1M MAU) | Best open PBR painter | Already covered; add geo-gate |+| Tencent-Hunyuan/Hunyuan3D-Part | 546 | Tencent community | Tencent community | No | P3-SAM segmentation, X-Part | Adopt design, rebuild (PartCrafter) |+| Tencent-Hunyuan/Hunyuan3D-Omni | 639 | Tencent community | Tencent community | No | Pose/bbox/voxel control | Adopt design |+| DreamTechAI/Direct3D-S2 | 1,285 | MIT | MIT | Yes | 1024^3 sparse SDF at 24 GB | Bake-off |+| Stable-X/Stable3DGen | 1,279 | MIT | MIT | Yes | Normal-bridged geometry fidelity | Bake-off |+| stepfun-ai/Step1X-3D | 896 | Apache-2.0 | Apache-2.0 | Yes | Two-stage, 27-29 GB | Already covered class |+| Stability SF3D / SPAR3D | 1,843 / 1,078 | Stability community | Stability community | Only under US $1M revenue | Fast, point-aware | Skip |+| VAST-AI-Research/SkinTokens | 463 | MIT | MIT (HF) | Yes | Unified rig for any mesh, 14 GB | Adopt code+weights (non-humanoid lane) |+| VAST-AI-Research/UniRig | 1,804 | MIT | MIT | Yes | FBX/GLB rig, 8 GB | Superseded by SkinTokens |+| Seed3D/Puppeteer | 427 | Apache-2.0 | Apache-2.0 | Yes | Rig + video-driven animation | Bake-off vs SkinTokens |+| jasongzy/Make-It-Animatable | 463 | MIT | MIT | Yes | Mixamo rig + blendshapes | Already covered |+| Tencent-Hunyuan/HY-Motion-1.0 | 2,586 | Tencent community | Tencent community | No | 1.0B DiT text-to-motion, SMPL-H | Adopt design, rebuild |+| EricGuo5513/momask-codes, OpenMotionLab/MotionGPT | 1,327 / 1,975 | MIT | MIT | Yes | Masked / LLM motion | Already covered by MDM; optional |+| NVlabs/GENMO, zju3dv/GVHMR | 522 / 2,087 | Noncommercial | Noncommercial | No | | Skip |+| Tencent-Hunyuan/HY-World-2.0 | 2,705 | Tencent community | Tencent community | No | Mesh+3DGS+pointcloud worlds, navmesh | Adopt design, rebuild |+| SkyworkAI/Matrix-3D | 844 | MIT | MIT | Yes | Text/image to 3D scene | Bake-off vs Diorama |+| SkyworkAI/Matrix-Game | 2,346 | MIT | MIT | Yes | Real-time interactive video world model | Adopt design (not a mesh lane) |+| ZiYang-xie/WorldGen | 2,154 | Apache-2.0 | FLUX.1-dev gated (non-commercial) | Code yes, panorama model no | Seconds-fast scene, 10 GB low-VRAM | Adopt code, swap panorama model |+| nv-tlabs/lyra | 2,647 | Apache-2.0 | NVIDIA research-only | No (weights) | Long-horizon 3DGS worlds | Adopt design only |+| Robbyant/lingbot-map | 17,505 | Apache-2.0 | Apache-2.0 | Yes | Video to scene | Already covered |+| sparkjsdev/spark | 3,705 | MIT | n/a | Yes | 100M+ splats, SOG/SPZ | Adopt code (viewer) |+| graphdeco-inria/gaussian-splatting | 24,145 | Inria/MPII research | n/a | No | Reference 3DGS | Use gsplat (Apache-2.0) instead |diff --git a/docs/research/oss-landscape-2026-10-notes/agent-commerce.md b/docs/research/oss-landscape-2026-10-notes/agent-commerce.mdnew file mode 100644index 000000000..067649eaa--- /dev/null+++ b/docs/research/oss-landscape-2026-10-notes/agent-commerce.md@@ -0,0 +1,228 @@+# OSS agent-commerce, payments, identity, wallet and A2A repos vs three.ws (as of 2026-10-09)++Scope note: three.ws is Apache-2.0, Solana-first, promoted coin $THREE. Its surfaces were read from `three-ws-surfaces.txt` and its docs (`docs/a2a-payments.md`, `docs/erc8004.md`, `docs/solana-reputation.md`, `docs/agent-reputation.md`, `docs/payment-sessions.md`, `docs/recurring-payments.md`, `docs/x402.md`, `docs/bnb-payments.md`, `docs/agora.md`). Star counts below were read from the GitHub REST API through an authenticated `gh` session on **2026-10-09** unless a different date is stated. "License" is the SPDX id GitHub reports, corrected from the LICENSE file text where GitHub reports NOASSERTION. All commercial-reuse statements assume an Apache-2.0 platform redistributing code (Apache-2.0, MIT, BSD, CC0 code: compatible with attribution/notice retention; AGPL-3.0: network copyleft, do not vendor; proprietary developer-agreement code: do not copy).++**Crypto-project flag legend (commit gate):** `[TOKEN]` = the project is itself a crypto token/project or is inseparable from one; naming it in a committed file needs owner approval. `[chain-infra]` = a chain, L2 or stablecoin issuer (also a crypto project, same gate). Unflagged = not a crypto project.++---++## Key Question 0: Which repos lead each category, what do their licenses permit, and what is the verdict for three.ws?++### Takeaway+No single open repo "beats" three.ws across the board: three.ws already ships the AP2-shaped Intent/Cart mandate chain, x402 on Solana and EVM, a Metaplex registry mirror of ERC-8004, memo-based Solana reputation, escrowed tasks (AgenC), receipts and prepaid credits. What the leaders have that three.ws lacks is concentrated in six places: (1) standardized on-chain job escrow with an evaluator role and hookable lifecycle (ERC-8183 / Virtuals ACP v2), (2) a payments standard with built-in session deposits, subscriptions, refunds and multi-rail (MPP), (3) verifiable delegation chains with selective disclosure (Verifiable Intent, SD-JWT), (4) signed-agent HTTP identity accepted by CDNs (Web Bot Auth), (5) Solana-native ERC-8004 registries that the Solana Foundation now points to (QuantuLabs 8004-solana), and (6) policy-enforced agent wallets with session caps and compliance screening (Coinbase Agentic Wallets, MetaMask Agent Wallet, Crossmint smart-contract policies).++### Cited Findings++#### A. Payment protocols for agents++| Repo | Stars (2026-10-09) | License | Commercial reuse by an Apache-2.0 platform | Standout features three.ws lacks | Verdict |+|---|---|---|---|---|---|+| **x402** (canonical repo moved to the Linux Foundation org) [x402-foundation/x402](https://github.com/x402-foundation/x402); the old [coinbase/x402](https://github.com/coinbase/x402) URL now holds a 163-star shim | 6,689 (2,146 forks) | Apache-2.0 | Full reuse, vendorable | v2 extensions three.ws has not documented adopting: SIWx (CAIP-122 wallet sign-in bound to prior payments), wallet-controlled reusable sessions, lifecycle hooks, dynamic `payTo` routing + input-based dynamic pricing, legacy-rail facilitators (ACH/SEPA/card), `@x402/paywall` split; Foundation working groups for identity, domain discovery, tax, card acceptance | adopt code (three.ws already consumes `@x402/*`; pull v2 SIWx, sessions, hooks, dynamic payTo) |+| **AP2 (Agent Payments Protocol)** [google-agentic-commerce/AP2](https://github.com/google-agentic-commerce/AP2) | 3,211 (509 forks) | Apache-2.0 | Full reuse | Intent/Cart/Payment mandate triad as **verifiable digital credentials**, "Human Not Present" flow (v0.2), 60+ partner list, FIDO governance, passkey binding; three.ws has an Intent Mandate and a cart mandate but not the Payment Mandate handed to the issuer/network nor a VC/passkey signing scheme | adopt design, rebuild (mandate schema already mirrored; add Payment Mandate + VC envelope) |+| **Agentic Commerce Protocol (ACP, OpenAI+Stripe)** [agentic-commerce-protocol/agentic-commerce-protocol](https://github.com/agentic-commerce-protocol/agentic-commerce-protocol) | 1,568 | Apache-2.0 | Full reuse | Merchant-side checkout_sessions state machine (create/update/complete/cancel/get), HMAC-signed order lifecycle webhooks incl. refunds, Delegated Payment Spec with single-use Shared Payment Tokens scoped to merchant+cart total, fiat cards | adopt design, rebuild (fiat card rail is the gap; the state machine maps onto payment sessions) |+| **Stripe agent toolkit** (repo renamed to [stripe/ai](https://github.com/stripe/ai), `stripe/agent-toolkit` redirects) | 1,864 | MIT | Full reuse | Stripe-hosted fiat rails for agents (MCP server, SPT issuance, metered billing) | adopt design only (needs Stripe account; EVM/fiat secondary) |+| **MPP (Machine Payments Protocol, Tempo+Stripe)** [tempoxyz/mpp-specs](https://github.com/tempoxyz/mpp-specs) | 95 | Specs CC0-1.0; tooling Apache-2.0 OR MIT (LICENSE text, GitHub reports NOASSERTION) | Full reuse of spec and SDK | Three intents (charge, **session** with deposit + automatic refund of unused deposit, **subscription**), `Payment-Receipt` header, refunds extension, relays, identity extension, OpenAPI discovery, x402 interop, **Solana SOL/SPL method**, Lightning, cards, Stripe SPT; SDKs in TS (mppx), Python, Rust, Go, Ruby; adopters: Cloudflare Agents SDK, official MCP SDK, Vercel AI SDK | adopt code (three.ws already bridges x402 to MPP on BNB; extend to Solana SPL method + sessions/subscriptions/refunds) |+| **L402 / Aperture** [lightninglabs/aperture](https://github.com/lightninglabs/aperture), [lightninglabs/lnget](https://github.com/lightninglabs/lnget) | 271 / 15 | MIT / none stated | Aperture: full reuse; lnget: no license file, do not vendor | Macaroon-scoped credentials, `--max-cost` client spend cap, Lightning rail, seven agent "skills" | already covered conceptually (not a Solana rail); adopt design for scoped credential caveats only |+| **Skyfire KYAPay** [skyfire-xyz/kyapay](https://github.com/skyfire-xyz/kyapay), [kyapay_a2a](https://github.com/skyfire-xyz/kyapay_a2a) | 13 / 2 | Spec: Community Specification License 1.0; code MIT; A2A ext Apache-2.0 | Code reusable; spec license is non-OSI, treat as read-only | KYA ("Know Your Agent") identity token bound to pay tokens via OAuth-style JWT (IETF drafts), A2A payment extension | adopt design, rebuild (the JWT token-exchange pattern) |+| **Nevermined payments** [nevermined-io/payments](https://github.com/nevermined-io/payments), [payments-py](https://github.com/nevermined-io/payments-py), [contracts](https://github.com/nevermined-io/contracts) | 16 / 10 / 1 | Apache-2.0 | Full reuse | Delegated "card"-style scoped API keys, credit plans, x402 support; service is commercial SaaS | already covered (prepaid credits, payment sessions) |+| **Payman** [PaymanAI/sigilum](https://github.com/PaymanAI/sigilum) | 9 | Apache-2.0 | Full reuse | Only "Auditable Identity for AI Agents" is open; payments SDK not on GitHub; product is agent-pays-human payouts with policies | already covered |+| **Locus (YC F25)** [locus-technologies](https://github.com/locus-technologies) (only `locus-pro-plugin`, `agent-access-bench`, both 0 stars, MIT) | 0 | MIT | Reusable but trivial | USDC-on-Base policy wallets with per-tx limits and approval thresholds; SDK not open | already covered |+| **Visa Trusted Agent Protocol** [visa/trusted-agent-protocol](https://github.com/visa/trusted-agent-protocol) | 212 (last push 2025-10-28) | Visa Developer agreement (NOASSERTION) | Do not copy | Merchant-side signed-intent verification, agent registry sample | adopt design only |+| **Verifiable Intent (Mastercard+Google, donated to FIDO)** [agent-intent/verifiable-intent](https://github.com/agent-intent/verifiable-intent) | 97 | Apache-2.0 | Full reuse | SD-JWT delegation chain issuer→user→agent with key-confirmation claims, 8 constraint types (amount bounds, merchant allowlists, budget caps, recurrence), selective disclosure, mappings to AP2/ACP/UCP | adopt code (wrap three.ws Intent Mandate in a VI delegation chain) |++#### B. Agent-to-agent and tool protocols++| Repo | Stars | License | Reuse | Features three.ws lacks | Verdict |+|---|---|---|---|---|---|+| **A2A** [a2aproject/A2A](https://github.com/a2aproject/A2A) | 26,078 | Apache-2.0 | Full | v1.0 (March 2026): signed Agent Cards (crypto-verified identity + metadata), multi-tenancy per endpoint, protobuf schema, 150+ orgs; three.ws has no A2A Agent Card for its agents | adopt code (publish `/.well-known/agent-card.json` per three.ws agent; KYAPay and x402 both have A2A payment extensions) |+| **MCP spec** [modelcontextprotocol/modelcontextprotocol](https://github.com/modelcontextprotocol/modelcontextprotocol) | 9,411 | Transitioning MIT→Apache-2.0 (LICENSE text) | Full | 2026-07-28 revision: stateless MRTR for sampling/elicitation (`InputRequiredResult`), Tasks moved to official extension SEP-2663 (`tasks/get`, `tasks/update`), extensions framework (MCP Apps, Tasks), auth hardening (RFC 9207 `iss`, CIMD instead of DCR, Enterprise-Managed Authorization) | adopt code (three.ws MCP servers should implement the 07-28 revision; its long-running tools map to the Tasks extension) |+| **MCP Registry** [modelcontextprotocol/registry](https://github.com/modelcontextprotocol/registry) | 7,331 | Transitioning MIT→Apache-2.0 | Full | Official upstream feed consumed by Smithery, Glama, PulseMCP | already covered if three.ws servers are listed; verify listing |++#### C. Identity and reputation++| Repo | Stars | License | Reuse | Features three.ws lacks | Verdict |+|---|---|---|---|---|---|+| **ERC-8004 contracts** [erc-8004/erc-8004-contracts](https://github.com/erc-8004/erc-8004-contracts) | 238 (114 forks) | CC0-1.0 | Full | Mainnet since 2026-01-29; same addresses on 25 mainnets; Reputation Registry stores signed int128 `value` + `valueDecimals` + two tags with `revokeFeedback`/`appendResponse`; Validation Registry with TEE `supportedTrust`; `setAgentWallet` needs EIP-712/ERC-1271 proof | already covered on EVM (three.ws `src/erc8004/*`); verify it tracks the current ABI (`appendResponse`, `revokeFeedback`) |+| **Solana Agent Registry (QuantuLabs port of ERC-8004)** [QuantuLabs/8004-solana](https://github.com/QuantuLabs/8004-solana), [8004-solana-ts](https://github.com/QuantuLabs/8004-solana-ts) | 13 / 9 | MIT | Full | Identity+reputation+validation programs on Solana, fee schedule 0.009 SOL register, 0.00001 SOL feedback, linked from solana.com/agent-registry, indexed by 8004scan since 2026-03-04 | adopt code (register three.ws agents there alongside the Metaplex mint; see KQ3) |+| **ERC-8183 Agentic Commerce (jobs/escrow)** [EIP-8183](https://eips.ethereum.org/EIPS/eip-8183), reference impl Virtuals ACP v2 | Draft EIP | EIP text CC0 | Full for the spec | Job primitive with client/provider/evaluator, six states, escrow pulled on fund, evaluator-only complete/reject, anyone-can-trigger expiry refund, before/after hooks, fees on completion only, x402 compat via ERC-2771 | adopt design, rebuild on Solana (see KQ2) |+| **Virtuals ACP** [Virtual-Protocol/agent-commerce-protocol](https://github.com/Virtual-Protocol/agent-commerce-protocol) (11 stars, MIT), [acp-node-v2](https://github.com/Virtual-Protocol/acp-node-v2) (1 star, no license), [acp-cli](https://github.com/Virtual-Protocol/acp-cli) (37, no license), [acp-python](https://github.com/Virtual-Protocol/acp-python) (28, no license), [protocol-contracts](https://github.com/Virtual-Protocol/protocol-contracts) (101, no license) `[TOKEN]` | see row | Mostly **no license file** | Only the 11-star MIT spec repo is reusable; SDKs/contracts are all-rights-reserved by default | Evaluator agents (human or AI) scoring deliverables into reputation; 12.3M ACP memos on Base | adopt design, rebuild; never vendor |+| **Web Bot Auth** [cloudflare/web-bot-auth](https://github.com/cloudflare/web-bot-auth) | 165 | Apache-2.0 | Full | RFC 9421 HTTP Message Signatures with Ed25519 per-agent keys, JWKS at `/.well-known/http-message-signatures-directory`, `Signature-Agent` header; IETF `webbotauth` WG draft (Sep 2026); used as the auth layer by Visa TAP and Mastercard Agent Pay | adopt code (sign outbound three.ws agent HTTP calls; verify inbound at the x402 server) |+| **Trust lists** [sudeepb02/awesome-erc8004](https://github.com/sudeepb02/awesome-erc8004) (81), [CodeSigils/awesome-agent-trust](https://github.com/CodeSigils/awesome-agent-trust) (1), [Pink-Agentic-Payments/awesome-agentic-payments](https://github.com/Pink-Agentic-Payments/awesome-agentic-payments) (0, CC0) | | | | reference only | n/a |++#### D. Agent wallets and on-chain action kits++| Repo | Stars | License | Reuse | Features three.ws lacks | Verdict |+|---|---|---|---|---|---|+| **Solana Agent Kit (SendAI)** [sendaifun/solana-agent-kit](https://github.com/sendaifun/solana-agent-kit) `[TOKEN]` (SendAI has a token per Solana Compass) | 1,719 (last push 2026-05-14) | Apache-2.0 | Full | v2 core+plugins, wallet-interface signing (no private key required), MCP server (~164 stars); 5,435 npm downloads/30d to 2026-10-04 | already covered (three.ws `solana-agent-sdk/`); mine plugins for protocol actions |+| **Coinbase AgentKit** [coinbase/agentkit](https://github.com/coinbase/agentkit) | 1,322 | Apache-2.0 (LICENSE text; GitHub NOASSERTION) | Full | README admits no human approval gate, no spend caps, no allowlists; EVM + Solana | already covered (three.ws allowances are stricter) |+| **Coinbase Agentic Wallets** (product, Feb 2026) + [coinbase/agentic-wallet-skills](https://github.com/coinbase/agentic-wallet-skills), [coinbase/cb-mpc](https://github.com/coinbase/cb-mpc) | 127 / 502 | MIT / MIT | Full | Session caps, transaction limits, OFAC screening, KYT, `awal` CLI + MCP | adopt design (compliance screening + session caps); three.ws already uses awal skills |+| **GOAT SDK** [goat-sdk/goat](https://github.com/goat-sdk/goat) | 1,008 | MIT | Full but **archived** (read-only, MCP server last human commit 2025-08-19) | wallet-agnostic plugin model | already covered; do not adopt |+| **elizaOS** [elizaOS/eliza](https://github.com/elizaOS/eliza) `[TOKEN, wound down]` | 19,565 | MIT | Full | Largest plugin library (Solana, x402, wallets); token and foundation wound down Aug 2026 after class action; code continues | adopt code selectively (plugin patterns), never the token references |+| **Turnkey** [tkhq/sdk](https://github.com/tkhq/sdk) | 102 | Apache-2.0 | Full (client; signing is a hosted service) | Enclave-enforced policies: sub-orgs, scoped API keys, quorum approvals, tx limits, allowlists, time-bound sessions | adopt design (policy DSL) |+| **Crossmint** [Crossmint/agent-commerce-sample-app](https://github.com/Crossmint/agent-commerce-sample-app) | 9 | MIT | Full | Smart-contract-wallet policies enforced on-chain: per-tx limits, rolling caps, recipient allowlists; agent signer in TEE | adopt design, rebuild on Solana (on-chain allowance program) |+| **Privy (Stripe)** | no agent repo found | n/a | n/a | Off-chain policies (allowlisted contracts/recipients, max amounts, MFA, quorum); no per-day budget on top of x402 signature screening | already covered |+| **MetaMask Agent Wallet** | no public repo found | n/a | n/a | Announced 2026-06-08, "live for everyone" 2026-08-06 per MetaMask on X; Guard/Beast modes, spend limits, Transaction Shield + Blockaid, TEE keys | adopt design (guard modes) |+| **Phantom** | MCP server + CLI v1.0.0 (April 2026), no agent SDK repo found | n/a | n/a | simulate-before-broadcast default, preview status | already covered |+| **Lightspark MPP SDK** [buildonspark/lightning-mpp-sdk](https://github.com/buildonspark/lightning-mpp-sdk) | 9 | MIT | Full | Lightning method for MPP | n/a (not Solana) |++#### E. Marketplaces and bazaars++| Repo / surface | Stars | License | Reuse | Features three.ws lacks | Verdict |+|---|---|---|---|---|---|+| **x402 Bazaar** (Coinbase discovery index; public storefront Agentic.Market; x402 v2 Discovery extension puts JSON Schema in the 402) | part of x402 repo | Apache-2.0 | Full | v2 discovery schema inside the 402 body; Agent402 ranks sellers by settled USDC volume | already covered (three.ws unified catalog, `search-for-service`); verify v2 discovery extension emitted on every three.ws 402 |+| **Smithery CLI** [arcadeai-labs/smithery-cli](https://github.com/arcadeai-labs/smithery-cli) (`smithery-ai/cli` redirects; Smithery now under Arcade) | 840 | **AGPL-3.0** | Do not vendor | hosted remote MCP endpoints, Toolbox meta-MCP routing; 6k-11k servers | adopt design only |+| **mcp.run** (dylibso) | no public repo found | n/a | n/a | Wasm "servlet" registry | n/a |+| **Glama** | no public repo found | n/a | n/a | quality scorecard (security, licensing, docs, ops readiness), ~37k-48k servers | adopt design (scorecard for three.ws skill listings) |+| **Virtuals ACP marketplace** `[TOKEN]` | see C | no license | no | evaluator-scored jobs, 2,000+ agents, 12.3M memos | adopt design |+| **8004 Market** (8004market.io, linked from solana.com/agent-registry) | no repo found | n/a | n/a | marketplace over registered Solana+EVM agents | adopt design |++Sources for the table rows: GitHub API reads on 2026-10-09 (stars, license, pushed_at, archived flag); x402 v2 extension list from [x402.org V2 launch post](https://x402.org/x402-v2-launch/); x402 Foundation membership and 165M tx / ~$50M from the [InfoQ July 2026 report](https://www.infoq.com/news/2026/07/cloudflare-aws-x402-micropayment/) and the [majormatters tracker](https://majormatters.co/x402); AP2 FIDO donation and v0.2 from [FIDO Alliance](https://fidoalliance.org/google-donates-agent-payments-protocol-to-fido-alliance/) and [Google blog](https://blog.google/products-and-platforms/platforms/google-pay/agent-payments-protocol-fido-alliance/); AP2 mandate triad from the [Google Cloud AP2 announcement](https://cloud.google.com/blog/products/ai-machine-learning/announcing-agents-to-payments-ap2-protocol); ACP mechanics from [Stripe ACP docs](https://docs.stripe.com/agentic-commerce/acp) and [OpenAI Delegated Payment Spec](https://developers.openai.com/commerce/specs/payment); MPP intents/methods/extensions from [mpp.dev](https://mpp.dev/) and licensing from [tempoxyz/mpp-specs](https://github.com/tempoxyz/mpp-specs); L402 from [Lightning Labs 2026-03-11 post](https://lightning.engineering/posts/2026-03-11-L402-for-agents/) and [2026-02-11 tools post](https://lightning.engineering/posts/2026-02-11-ln-agent-tools/); ERC-8004 registry details from the [contracts README](https://github.com/erc-8004/erc-8004-contracts); ERC-8183 from the [EIP](https://eips.ethereum.org/EIPS/eip-8183); Solana Agent Registry fees and links from [solana.com/agent-registry](https://solana.com/agent-registry); Web Bot Auth architecture from the [IETF draft](https://datatracker.ietf.org/doc/html/draft-meunier-web-bot-auth-architecture) and [Cloudflare's agentic commerce post](https://blog.cloudflare.com/secure-agentic-commerce/); Verifiable Intent from [verifiableintent.dev](https://verifiableintent.dev/) and [Mastercard](https://www.mastercard.com/global/en/news-and-trends/stories/2026/verifiable-intent.html); A2A v1.0 from [a2a-protocol.org](https://a2a-protocol.org/latest/announcing-1.0/) and [Google open source blog](https://opensource.googleblog.com/2026/04/a-year-of-open-collaboration-celebrating-the-anniversary-of-a2a.html); MCP 2026-07-28 from the [MCP blog](https://blog.modelcontextprotocol.io/posts/2026-07-28/) and [WorkOS](https://workos.com/blog/mcp-2026-spec-agent-authentication); Coinbase Agentic Wallets from [eco explainer](https://eco.com/support/en/articles/14845485-coinbase-agentic-wallets-explained); AgentKit limitations from the [AgentKit README](https://github.com/coinbase/agentkit); MetaMask Agent Wallet from [Crowdfund Insider](https://www.crowdfundinsider.com/2026/06/284472-metamask-introduces-agent-wallet-enabling-ai-agents-with-autonomous-defi-access/) and [MetaMask on X](https://x.com/MetaMask/status/2085365793441685614); Phantom from [Phantom docs updates](https://docs.phantom.com/updates); Crossmint/Privy/Turnkey policy comparison from [Crossmint](https://www.crossmint.com/learn/agent-wallets-compared), [Fireblocks](https://www.fireblocks.com/report/compare-agent-wallets), [Privy docs](https://docs.privy.io/wallets/overview/embedded); Solana Agent Kit npm downloads from [stablecoininsider](https://stablecoininsider.org/how-to-send-usdc-with-solana-agent-kit/); Virtuals ACP from the [whitepaper](https://whitepaper.virtuals.io/about-virtuals/commerce-layer), [ACP changelog](https://whitepaper.virtuals.io/acp/acp-changelogs) and [agenteconomy.to](https://agenteconomy.to/virtuals-acp); elizaOS token wind-down from [hoge.gg](https://hoge.gg/eliza-token-dead-framework-survives-2026/) and [Crypto Briefing](https://cryptobriefing.com/elizaos-token-migration-ecosystem/); Smithery/Glama counts from [TrueFoundry](https://www.truefoundry.com/blog/best-mcp-registries) and [Ginger Labs](https://gingerlabs.ai/blog/find-evaluate-mcp-servers-smithery-glama-mcp-so); x402 Bazaar from [Coinbase launch page](https://www.coinbase.com/developer-platform/discover/launches/x402-bazaar); Skyfire from [stellagent](https://stellagent.ai/insights/skyfire-kyapay-know-your-agent) and [kyapay repo](https://github.com/skyfire-xyz/kyapay); Locus from [paywithlocus.com](https://paywithlocus.com/) and the [org page](https://github.com/locus-technologies); Payman from [docs](https://docs.paymanai.com/overview/introduction) and [org page](https://github.com/paymanai); Nevermined from [payments repo](https://github.com/nevermined-io/payments).++### Inferences+- The highest-leverage "adopt code" items for a Solana-first Apache-2.0 platform are all permissively licensed: x402 v2 extensions (Apache-2.0), MPP spec + mppx (CC0 / Apache-or-MIT), Verifiable Intent (Apache-2.0), Web Bot Auth (Apache-2.0), QuantuLabs 8004-solana (MIT), A2A (Apache-2.0), MCP 07-28 (Apache-2.0 transition). None requires naming a token in committed code.+- Every repo that is "big by stars" (A2A 26k, elizaOS 19.5k, MCP 9.4k, x402 6.7k, AP2 3.2k) is a protocol or framework, not a marketplace. The marketplace layer (Virtuals ACP, 8004 Market, Agentic.Market, Smithery) is either unlicensed, AGPL, or closed, so three.ws's open marketplace code is itself differentiated.+- Virtuals' SDKs lacking a license file means their ACP v2 code is not legally reusable even though the EIP it implements (ERC-8183) is CC0; rebuilding the ERC-8183 state machine as a Solana program is the clean path.++### Gaps+- No public repo was found for MetaMask Agent Wallet, Phantom's agent MCP server, Coinbase Agentic Wallets core, Privy agent policies, mcp.run, Glama, 8004 Market, Dexter or PayAI facilitators; their features are documented only in vendor pages.+- AP2's spec page at ap2-protocol.org/specification/ and the repo's docs/specification.md both returned 404 on 2026-10-09, so the exact v0.2 field-level mandate schema was not read from a primary source; the mandate triad is cited from Google's announcement.+- Mastercard Verifiable Intent's GitHub home is [agent-intent/verifiable-intent](https://github.com/agent-intent/verifiable-intent), credited only to "Verifiable Intent Contributors"; the Mastercard/Google authorship is from Mastercard's own newsroom, not the repo.++---++## Key Question 1: What changed in the agent-payments landscape between mid-2025 and October 2026?++### Takeaway+The period saw protocol consolidation into foundations (x402 to the Linux Foundation, AP2 and Verifiable Intent to FIDO, MCP toward Apache-2.0, A2A v1.0), the launch of a second HTTP-402 standard with sessions and refunds (MPP, March 2026) backed by Stripe/Visa/Tempo, ERC-8004 going to mainnet with 20k+ registered agents and a Solana port the Solana Foundation now links to, and Solana overtaking Base in monthly x402 transaction share while fiat-side checkout (ChatGPT Instant Checkout) was rolled back.++### Cited Findings+- x402 v2 released after a two-week community period; "over 100M payments" since May 2025; headers PAYMENT-SIGNATURE / PAYMENT-REQUIRED / PAYMENT-RESPONSE replace X-* ; Discovery extension, SIWx (CAIP-122), reusable sessions, multi-facilitator, legacy-rail facilitators, dynamic payTo and pricing, lifecycle hooks, `@x402` npm org — [x402.org](https://x402.org/x402-v2-launch/)+- x402 spec v2 "Released December 11, 2025"; "Batch Settlement extension added May 11, 2026"; Foundation founding members listed as Coinbase, Cloudflare, Stripe, Visa, Mastercard, American Express, Google, Microsoft, AWS — [majormatters x402 tracker](https://majormatters.co/x402) (tracker, secondary)+- x402 Foundation under the Linux Foundation with AWS, Cloudflare, Anthropic, Circle and 20+ members; AWS CloudFront x402 GA, Cloudflare Monetization Gateway waitlist; 165M transactions and ~$50M cumulative cited at Foundation launch — [InfoQ, July 2026](https://www.infoq.com/news/2026/07/cloudflare-aws-x402-micropayment/); Foundation launch date reported as 2 April 2026 with 22 institutions incl. Solana Foundation — [Crypto Briefing](https://cryptobriefing.com/solana-x402-market-share-dominance/); InfoQ gives 14 July 2026 as operational date (sources conflict)+- Canonical x402 repo now lives at x402-foundation/x402 (6,689 stars, Apache-2.0, pushed 2026-10-07); working groups `wg-identity`, `wg-domain-discovery`, `wg-tax`, `wg-card-acceptance` exist under the org — [GitHub org](https://github.com/x402-foundation) (API read 2026-10-09)+- AP2 donated to the FIDO Alliance 2026-04-29 together with v0.2, which adds "Human Not Present" payments on pre-authorized instructions; Verifiable Intent co-developed with Mastercard also donated — [FIDO Alliance](https://fidoalliance.org/google-donates-agent-payments-protocol-to-fido-alliance/); [Google](https://blog.google/products-and-platforms/platforms/google-pay/agent-payments-protocol-fido-alliance/)+- FIDO formed an Agentic Authentication TWG and a Payments TWG (chaired by Mastercard and Visa members) on 2026-04-28 — [FIDO](https://fidoalliance.org/fido-alliance-to-develop-standards-for-trusted-ai-agent-interactions/); [Help Net Security](https://www.helpnetsecurity.com/2026/04/29/fido-alliance-ai-agents-authentication-payments-standards/)+- Verifiable Intent unveiled by Mastercard March 2026, open source at verifiableintent.dev, protocol-agnostic across AP2 and UCP — [Mastercard](https://www.mastercard.com/global/en/news-and-trends/stories/2026/verifiable-intent.html)+- ACP (OpenAI+Stripe, Apache-2.0, released 2025-09-29); Stripe credits Meta as co-creator; latest stable spec 2026-04-17; ChatGPT Instant Checkout retired/rolled back toward discovery in March 2026 after roughly a dozen Shopify merchants shipped — [eco ACP explainer](https://eco.com/support/en/articles/14845478-acp-agentic-commerce-protocol-explained); [Stripe docs](https://docs.stripe.com/agentic-commerce/acp); [Applied Technology Index](https://appliedtechnologyindex.com/research/2026-comparative-analysis-agentic-commerce-payment-protocols/); [signalcommerce](https://signalcommerce.substack.com/p/the-agentic-commerce-specs-stayed)+- MPP announced 2026-03-18 with Tempo mainnet; Visa adding a card-based MPP spec; Lightspark adapted MPP to Lightning — [Stripe blog](https://stripe.com/blog/machine-payments-protocol); [Forbes](https://www.forbes.com/sites/ninabambysheva/2026/03/18/paradigm-and-stripe-roll-out-new-payment-standard-for-ai-agents-with-visas-support/); [The Defiant](https://thedefiant.io/news/blockchains/tempo-launches-mainnet-unveils-machine-payments-protocol-with-stripe)+- MPP claims 50+ services incl. OpenAI, Anthropic, Google Gemini, Dune (secondary) and "$14M in 60 days" (secondary blog) — [getblock](https://getblock.io/blog/what-is-a-machine-payments-protocol-mpp/); [signb.ee](https://signb.ee/blog/mpp-is-moving-faster-than-anyone-expected)+- UCP (Google+Shopify) is a separate commerce standard compatible with AP2 mandates; snapshot v2026-08-25; Salesforce Commerce Cloud GA slated October 2026 (unverified) — [awesome-agent-payments-protocol](https://github.com/0x-starter/awesome-agent-payments-protocol)+- Visa TAP (2025-10-14) and Mastercard Agent Pay (2025-04-29) both use Web Bot Auth as the agent authentication layer — [Cloudflare](https://blog.cloudflare.com/secure-agentic-commerce/); [Visa investor release](https://investor.visa.com/news/news-details/2025/Visa-Introduces-Trusted-Agent-Protocol-An-Ecosystem-Led-Framework-for-AI-Commerce/default.aspx)+- Web Bot Auth adopted as an IETF `webbotauth` working-group draft, as at 2026-09-12, co-authored by Cloudflare and Google; chatgpt.com and browserbase serve live JWKS directories — [stellagent](https://stellagent.ai/insights/web-bot-auth-cloudflare-ietf); [IETF architecture draft](https://datatracker.ietf.org/doc/html/draft-meunier-web-bot-auth-architecture)+- A2A v1.0 released March 2026: signed Agent Cards, multi-tenancy, breaking interaction changes, 150+ organizations — [a2a-protocol.org](https://a2a-protocol.org/latest/announcing-1.0/); [PR Newswire](https://www.prnewswire.com/news-releases/a2a-protocol-surpasses-150-organizations-lands-in-major-cloud-platforms-and-sees-enterprise-production-use-in-first-year-302737641.html)+- MCP 2026-07-28 revision: MRTR replaces held-open streams for sampling/elicitation; sessions removed; Tasks became official extension SEP-2663; extensions framework; RFC 9207 `iss`; CIMD over Dynamic Client Registration; Enterprise-Managed Authorization — [MCP blog](https://blog.modelcontextprotocol.io/posts/2026-07-28/); [Appwrite](https://appwrite.io/blog/post/mcp-goes-stateless-in-the-2026-07-28-specification); [WorkOS](https://workos.com/blog/mcp-2026-spec-agent-authentication); [Stacktree](https://stacktr.ee/blog/mcp-2026-spec-changes)+- ERC-8004 mainnet 2026-01-29; 21,562 registered agents per 8004scan, Agent Arena lists 22,000+ across 17 chains (16 EVM + Solana) — [news.bitcoin.com](https://news.bitcoin.com/what-is-erc-8004-ethereums-new-agent-standard-powers-thousands-of-onchain-ai-identities/); [Agent Arena](https://agentarena.site/); [Bankless Times](https://www.banklesstimes.com/articles/2026/01/28/ethereum-erc-8004-set-for-mainnet-launch-to-enable-trustless-ai-agents/)+- ERC-8183 created 2026-02-25 (Draft) by Davide Crapis, Bryan Lim, Tay Weixiong, Chooi Zuhwa; Virtuals ACP v2.0 (April 2026) is its reference implementation after "18 months in production and over 2,000 agents" — [EIP-8183](https://eips.ethereum.org/EIPS/eip-8183); [Virtuals ACP changelog](https://whitepaper.virtuals.io/acp/acp-changelogs); Arc (Circle) adopting ERC-8183 with USDC escrow — [Arc on X](https://x.com/arc/status/2041525873816580365)+- Coinbase Agentic Wallets launched 2026-02-11 with session caps, tx limits, OFAC screening, KYT (secondary) — [eco](https://eco.com/support/en/articles/14845485-coinbase-agentic-wallets-explained)+- MetaMask Agent Wallet early access 2026-06-08, "live for everyone" 2026-08-06 — [Crowdfund Insider](https://www.crowdfundinsider.com/2026/06/284472-metamask-introduces-agent-wallet-enabling-ai-agents-with-autonomous-defi-access/); [MetaMask on X](https://x.com/MetaMask/status/2085365793441685614)+- GOAT SDK archived (read-only) — [goat-sdk/goat](https://github.com/goat-sdk/goat) (API `archived` flag not set but description reads "[Archived] Read-only historical snapshot", pushed 2026-07-02)+- Skyfire: no primary evidence of shutdown; skyfire.xyz returned HTTP 200 on 2026-10-09 and the org pushed demos 2026-10-08; one blog claims a shutdown, one directory says "private or recently acquired" — [archtools blog](https://archtools.dev/blog-skyfire-migration.html) (unverified) vs [skyfire-xyz org](https://github.com/skyfire-xyz)+- Lightning Labs released seven agent skills (2026-02-12) and the L402 site (2026-07-29) — [The Block](https://www.theblock.co/news/ecosystems/2026-02-12-lightning-labs-releases-ai-agent-tools-for-native-bitcoin-lightning-payments-389584); [The Defiant](https://thedefiant.io/news/infrastructure/lightning-labs-launches-site-for-l402-bitcoin-agent-payments)+- elizaOS token migrated from AI16Z (window ended 2026-02-04); token and foundation wound down August 2026 after class-action settlement, code continues under MIT — [Crypto Briefing](https://cryptobriefing.com/elizaos-token-migration-ecosystem/); [hoge.gg](https://hoge.gg/eliza-token-dead-framework-survives-2026/)++### Inferences+- Two 402 standards now coexist (x402 v2 and MPP) and both have Solana methods; MPP is the one with sessions/subscriptions/refunds in-spec, x402 is the one with Linux Foundation governance and the larger install base. A platform that speaks both (three.ws already bridges them on BNB) is ahead of either camp.+- Fiat-side agent checkout (ACP) stalled on merchant adoption while stablecoin 402 rails grew; the card networks responded by moving under FIDO and Web Bot Auth rather than open repos. The open, reusable fiat artifacts are therefore the specs (ACP Apache-2.0, Verifiable Intent Apache-2.0), not processors.++### Gaps+- Protocol-wide volume figures conflict by source and window (x402 165M tx / $50M vs weekly volume collapse below $220k); no primary dashboard was fetched.+- No independent measurement of MPP volume beyond vendor blogs.++---++## Key Question 2: Which protocol features (mandates, intent signatures, delegated budgets, multi-party settlement, refunds) do leading repos implement that three.ws's economy lacks?++### Takeaway+three.ws already has the delegated-budget core (signed Intent Mandate with per-call/per-day/per-counterparty ceilings, kill switch, custody ledger, portable signed cart mandate, escrowed AgenC tasks, receipts). The documented gaps are: a standard evaluator-arbitrated escrow with expiry refunds and hooks (ERC-8183), in-protocol refunds and session-deposit auto-refund (MPP), a Payment Mandate handed to the issuer with VC/passkey binding and selective-disclosure delegation chains (AP2 v0.2 + Verifiable Intent), fiat Shared Payment Tokens (ACP/Stripe), signed Agent Cards (A2A), Web-Bot-Auth signed requests, x402 v2 SIWx/sessions/hooks/dynamic payTo, and on-chain-enforced wallet policies with compliance screening.++### Cited Findings+**What three.ws already has (internal docs, read 2026-10-09):**+- Intent Mandate: signed credential for a named agent, total budget, per-call cap, named networks, up to 90 days; seven ordered gates (signature/ownership, subject agent, freeze switch, per-call policy, optional peer reputation bar, lifetime budget, agent spend policy); custody ledger rows pending/confirmed/released; portable signed cart mandate verifiable at a public endpoint — three.ws `docs/a2a-payments.md`, `docs/start-here.md`+- ERC-8004 identity/reputation/validation helpers on EVM, same-address deployments, Solana mirror via Metaplex mint, memo attestations (feedback, stakes, tasks, disputes) aggregated by `/api/agents/solana-reputation` — three.ws `docs/agent-reputation.md`, `docs/solana.md`+- AgenC task escrow in $THREE: `createAgenCTask` locks reward, `completeAgenCTask` releases — three.ws `docs/agora.md`+- Multi-rail x402 v2 header names already used on X Layer; x402↔MPP bridge on BNB Chain — three.ws `docs/x402.md`, surfaces list line 285++**What the leaders implement that three.ws lacks:**+- ERC-8183 Job: client/provider/evaluator, six states (Open, Funded, Submitted, Completed, Rejected, Expired), evaluator-only complete/reject after submission, anyone-can-trigger refund after expiry that is never hookable, before/after hooks per job, platform fee only on completion, optional evaluator fee, reason hash emitted for audit, explicit "No dispute resolution or arbitration; reject/expire is final", x402 compat via ERC-2771 meta-tx + ERC-2612 permits — [EIP-8183](https://eips.ethereum.org/EIPS/eip-8183)+- Virtuals ACP: four phases (Request, Negotiation, Transaction, Evaluation), evaluators human or AI, evaluation results feed reputation; fee split reported 95/5 or 90/5/5 provider/evaluator/protocol (secondary, unverified) — [Virtuals whitepaper](https://whitepaper.virtuals.io/about-virtuals/commerce-layer); [dev.to comparison](https://dev.to/meshledger/agent-payment-protocols-compared-x402-vs-acp-vs-escrow-based-settlement-56l4)+- MPP: charge / session / subscription intents; `Payment-Receipt` header; refunds extension incl. automatic refund of unused session deposits; relays; identity extension; OpenAPI discovery; Solana SOL/SPL method — [mpp.dev](https://mpp.dev/)+- AP2 mandates: Intent Mandate (user's instructions/constraints), Cart Mandate (exact items/price signed), Payment Mandate (signal to issuer/network that an agent was involved), as verifiable digital credentials; v0.2 Human-Not-Present; passkeys central per FIDO — [Google Cloud announcement](https://cloud.google.com/blog/products/ai-machine-learning/announcing-agents-to-payments-ap2-protocol); [FIDO passkeys post](https://fidoalliance.org/mastercard-ai-payments-passkeys-fido-alliance/)+- Verifiable Intent: SD-JWT delegation chain issuer→user→agent with key-confirmation claims; Identity/Intent/Action layers; eight constraint types (amount bounds, merchant allowlists, budget caps, recurrence terms, others); selective disclosure; mappings to AP2/ACP/UCP; Apache-2.0 — [verifiableintent.dev](https://verifiableintent.dev/)+- ACP: checkout_sessions state machine not_ready_for_payment → ready_for_payment → completed; HMAC-signed order webhooks covering confirmation, shipping, delivery, refunds; Shared Payment Token single-use scoped to merchant + cart total; refunds/chargebacks stay with merchant PSP — [Stripe docs](https://docs.stripe.com/agentic-commerce/acp); [OpenAI Delegated Payment Spec](https://developers.openai.com/commerce/specs/payment); [signalcommerce](https://signalcommerce.substack.com/p/the-agentic-commerce-specs-stayed)+- x402 v2: SIWx (CAIP-122) reusing paid access, wallet-controlled sessions, lifecycle hooks, dynamic payTo per request, input-based pricing, multi-facilitator, legacy-rail facilitators (ACH/SEPA/cards) — [x402.org](https://x402.org/x402-v2-launch/); batch settlement extension (2026-05-11) — [majormatters](https://majormatters.co/x402)+- A2A v1.0: signed Agent Cards, multi-tenant endpoints — [a2a-protocol.org whats-new](https://a2a-protocol.org/latest/whats-new-v1/)+- KYAPay: JWT-based KYA+PAY token exchange, A2A payment extension (Apache-2.0), IETF drafts for OAuth KYAPay tokens — [kyapay](https://github.com/skyfire-xyz/kyapay); [kyapay_a2a](https://github.com/skyfire-xyz/kyapay_a2a); [draft](https://github.com/skyfire-xyz/draft-skyfire-oauth-kyapay-token)+- Web Bot Auth: Ed25519 per-agent key, JWKS at `/.well-known/http-message-signatures-directory`, `Signature-Agent` header, RFC 9421 — [IETF draft](https://datatracker.ietf.org/doc/html/draft-meunier-web-bot-auth-architecture); [cloudflare/web-bot-auth](https://github.com/cloudflare/web-bot-auth)+- Wallet policy enforcement: Crossmint smart-contract wallets enforce per-tx limits, rolling caps, recipient allowlists on-chain with the agent signer in a TEE; Turnkey enforces tx limits, allowlists, quorum, time-bound sessions in enclaves; Privy policies are off-chain with no per-day budget over x402 signatures — [Crossmint](https://www.crossmint.com/learn/agent-wallets-compared); [Fireblocks](https://www.fireblocks.com/report/compare-agent-wallets); [dcpagent](https://dcpagent.com/blog/agentic-wallets-compared); Coinbase Agentic Wallets add OFAC screening + KYT — [eco](https://eco.com/support/en/articles/14845485-coinbase-agentic-wallets-explained); MetaMask Guard/Beast modes + Blockaid — [Crowdfund Insider](https://www.crowdfundinsider.com/2026/06/284472-metamask-introduces-agent-wallet-enabling-ai-agents-with-autonomous-defi-access/)+- L402: macaroon caveats scope credentials; `lnget --max-cost` client-side cap — [Lightning Labs](https://lightning.engineering/posts/2026-02-11-ln-agent-tools/)+- MCP Tasks extension (SEP-2663) with `tasks/get` polling and `tasks/update` client input — [Stacktree](https://stacktr.ee/blog/mcp-2026-spec-changes); [MCP blog](https://blog.modelcontextprotocol.io/posts/2026-07-28/)++### Inferences+- three.ws's AgenC escrow is bilateral (create/complete); ERC-8183 adds the third party (evaluator), expiry-driven refunds and hooks. A Solana port of the ERC-8183 state machine as an Anchor program, with the three.ws reputation feed as the "evaluator attestation → reputation" mapping the EIP recommends, would out-build Virtuals on Solana where it has no presence (Virtuals is Base/Arbitrum/Ethereum/Polygon/BNB).+- three.ws's receipts and prepaid credits lack an in-protocol refund primitive; MPP's refund extension and session-deposit auto-refund are the closest open design and already carry a Solana SPL method.+- Wrapping the existing Intent Mandate in a Verifiable Intent SD-JWT chain gives three.ws FIDO-track compatibility (AP2/ACP/UCP mappings) without changing its gate logic.+- Publishing an A2A Agent Card and a Web Bot Auth JWKS per three.ws agent is low-cost and makes agents discoverable/verifiable by Google, Microsoft, Visa and Cloudflare stacks.++### Gaps+- Whether three.ws's x402 server already emits the v2 Discovery extension schema and honors SIWx/sessions could not be confirmed from docs grep alone (only header names and MPP bridge are documented).+- The Virtuals evaluator fee split and the AP2 v0.2 field-level schema were not confirmed from primary sources.++---++## Key Question 3: Which identity/reputation standards are gaining real adoption on Solana?++### Takeaway+On Solana the only standard with institutional pull is the ERC-8004 port: the Solana Foundation's solana.com/agent-registry page points developers to QuantuLabs' MIT-licensed `8004-solana` programs and TypeScript SDK, 8004scan has indexed Solana since 2026-03-04, and 8004market / SATI dashboards sit on top. Everything else (openSVM aeamcp, s8004 POC, Virtuals ACP) is either a proof of concept or absent from Solana.++### Cited Findings+- Solana Foundation page "Agent Registry": identity, reputation and validation registries "natively to Solana", "interoperable with ERC-8004 on Ethereum"; fees Register 0.009 SOL (~$0.81), Leave Feedback 0.00001 SOL (~$0.001), Set Agent Wallet 0.00001 SOL; links Quickstart to QuantuLabs/8004-solana-ts, explorer 8004scan.io, marketplace 8004market.io, SATI dashboard sati.cascade.fyi, spec at 8004.qnt.sh — [solana.com/agent-registry](https://solana.com/agent-registry)+- Registry "developed by Quantu AI" as the Solana implementation of ERC-8004 — [ChainCatcher](https://www.chaincatcher.com/en/article/2249074); [KuCoin news](https://www.kucoin.com/news/flash/solana-launches-ai-agent-trust-layer-agent-registry)+- QuantuLabs/8004-solana: 13 stars, MIT, pushed 2026-09-10, programs `agent-registry-8004` and `atom-engine`, mainnet/devnet program ids beginning `8oo4`; 8004-solana-ts: 9 stars, MIT — [8004-solana](https://github.com/QuantuLabs/8004-solana); [8004-solana-ts](https://github.com/QuantuLabs/8004-solana-ts) (API read 2026-10-09)+- QuantuLabs/agent-registry-program (MIT) lists a different program id `25wEsSLdsmZUisXuciyUXZqbpocsk5CJ7Uf6Eq553N8r`; the two repos disagree, verify on-chain — [agent-registry-program](https://github.com/QuantuLabs/agent-registry-program)+- 8004scan launched Solana support 2026-03-04; Agent Arena indexes 22,000+ agents across 16 EVM chains + Solana — [8004scan via eco](https://eco.com/support/en/articles/14730445-erc-8004-trustless-agent-identity); [Agent Arena](https://agentarena.site/)+- RelAI claims Solana Foundation "officially adopted" the registry (vendor claim) — [RelAI](https://relai.fi/blog/solana-agent-registry)+- Registry Broker indexes ERC-8004 agents on Solana **devnet**, suggesting part of the activity is still devnet — [hol.org](https://hol.org/docs/registry-broker/erc-8004-solana/)+- Competing Solana designs: openSVM/aeamcp "Solana Protocol Design for Agent and MCP Server Registries" (12 stars, no license, pushed 2026-02-16); Woody4618/s8004 POC (4 stars, no license, last push 2025-11-06) — [aeamcp](https://github.com/openSVM/aeamcp); [s8004](https://github.com/Woody4618/s8004)+- Solana x402 share: ~70% of monthly x402 transactions in August 2026 (~$24.24M/30d); Solana passed Base in daily tx count 2026-08-25; share swung 49.7%→88% earlier in 2026; Base still leads cumulatively (Chainalysis 100M+ agentic payments on Base by mid-2026) — [Crypto Briefing](https://cryptobriefing.com/solana-x402-market-share-dominance/); [SolanaFloor](https://solanafloor.com/news/solana-commands-49-of-x402-market-share-as-the-race-for-micropayment-dominance-intensifies); [Pluang](https://pluang.com/en/news-feed/solana-kuasai-pangsa-pasar-x402-saat-volume-transaksi-menurun)+- Solana facilitators: Dexter ~68.7% of Solana x402 tx and >75% of volume; PayAI 30.5% (PayAI self-reports ~90% of "real" tx) — [SolanaFloor](https://solanafloor.com/news/solana-commands-49-of-x402-market-share-as-the-race-for-micropayment-dominance-intensifies); [PayAI blog](https://blog.payai.network/largest-x402-facilitator-solana/)+- Solana Foundation is among the 22 x402 Foundation backers — [Crypto Briefing](https://cryptobriefing.com/solana-x402-market-share-dominance/)+- MPP lists a Solana SOL/SPL payment method — [mpp.dev](https://mpp.dev/)+- Solana Agent Kit (SendAI): 1,719 stars, Apache-2.0, last push 2026-05-14; MCP server ~164 stars — [GitHub](https://github.com/sendaifun/solana-agent-kit); [SendAI org](https://github.com/sendaifun)+- Phantom's agent surface is an MCP server + CLI v1.0.0 (April 2026), preview — [Phantom docs](https://docs.phantom.com/updates)+- three.ws internal: "A full ERC-8004-equivalent on Solana (Identity / Reputation / Validation as Anchor programs with a parallel JS client) remains a separate, larger piece of work" — three.ws `docs/solana.md`++### Inferences+- three.ws's own docs name the exact gap that QuantuLabs has now filled under MIT. Registering every three.ws agent in the Foundation-endorsed registry (in addition to the Metaplex Genesis mint) and writing feedback there makes three.ws reputation visible on 8004scan/8004market/SATI at ~$0.001 per attestation.+- Because three.ws already runs its own x402 facilitator on Solana, the Dexter/PayAI duopoly is a competitive reference, not a dependency; listing the three.ws facilitator in the x402 Foundation's facilitator set is the adoption lever.+- Solana Agent Kit's last push (May 2026) and GOAT's archive suggest on-chain action kits have plateaued; the growth is in registries and payment standards.++### Gaps+- No primary count of agents registered on the Solana registry (solana.com page states none; 8004scan not fetched).+- No Solana Foundation primary source for the "~65% of 2026 tx volume" figure; only a secondary roundup.+- Program-id discrepancy between QuantuLabs repos unresolved.++---++## Key Question 4: Which of these repos are themselves crypto projects or tokens (commit-gate flags)?++### Takeaway+Flag before committing any reference: Virtuals Protocol (VIRTUAL token), elizaOS (ELIZAOS/AI16Z token, wound down but still a token), SendAI/Solana Agent Kit (SendAI token per Solana Compass), Tempo (Stripe/Paradigm L1 behind MPP), Arc (Circle chain), Coinbase/Base, Circle/USDC, Solana itself, Lightning/Bitcoin, Nevermined (has issued a token historically), Skyfire (stablecoin settlement company, no token found), plus any chain named in a registry deployment list. Not crypto projects: Google AP2, OpenAI/Stripe ACP, A2A, MCP, Cloudflare Web Bot Auth, Visa TAP, Mastercard Verifiable Intent, Turnkey, Privy, Locus, Payman, Smithery, Glama, mcp.run.++### Cited Findings+- Virtuals Protocol: VIRTUAL token, 1B cap, ~658M circulating (secondary) — [ownyourmind](https://ownyourmind.ai/projects/virtuals/); ACP on Base/Arbitrum — [crypto.news](https://crypto.news/virtuals-protocol-brings-ai-agent-commerce-to-arbitrum-in-new-integration/)+- elizaOS: AI16Z→ELIZAOS migration ended 2026-02-04; token "dead", foundation winding down Aug 2026; founder will not allow a future token on the IP — [Crypto Briefing](https://cryptobriefing.com/elizaos-token-migration-ecosystem/); [hoge.gg](https://hoge.gg/eliza-token-dead-framework-survives-2026/)+- SendAI listed as a Solana project with programs and token — [Solana Compass](https://solanacompass.com/projects/sendai)+- Tempo: Stripe/Paradigm-backed blockchain, mainnet 2026-03-18 — [Finextra](https://www.finextra.com/newsarticle/47462/stripe-backed-tempo-blockchain-goes-live-with-machine-payments-protocol)+- Arc adopting ERC-8183 with Circle Wallets and USDC — [Arc on X](https://x.com/arc/status/2041525873816580365)+- x402 Foundation members include Circle, Coinbase, Solana Foundation — [InfoQ](https://www.infoq.com/news/2026/07/cloudflare-aws-x402-micropayment/); [Crypto Briefing](https://cryptobriefing.com/solana-x402-market-share-dominance/)+- Skyfire: KYAPay settles in USDC; no token found — [stellagent](https://stellagent.ai/insights/skyfire-kyapay-know-your-agent)+- Lightning Labs / L402: Bitcoin Lightning — [Lightning Labs](https://lightning.engineering/posts/2026-03-11-L402-for-agents/)+- ERC-8004 contracts list 25 mainnets (Ethereum, Base, Arbitrum, Optimism, Polygon, Avalanche, BSC, Linea, ...) — [contracts README](https://github.com/erc-8004/erc-8004-contracts)+- Virtuals SDK repos carry no license file (acp-node-v2, acp-cli, acp-python, protocol-contracts) — GitHub API read 2026-10-09 ([org](https://github.com/Virtual-Protocol))++### Inferences+- The safe pattern for committed code: implement the CC0/Apache specs (ERC-8183 state machine, MPP, Verifiable Intent, Web Bot Auth, ERC-8004 ABI) and name only the standard (EIP number, protocol name), never the token project that popularized it. "ERC-8183" and "ERC-8004" are EIP identifiers, not coins; "Virtuals" and "VIRTUAL" are. Chain names already present in three.ws (Solana, Base, BNB Chain, X Layer, Robinhood Chain) are existing references, not new ones, but the gate still applies to any new one (Tempo, Arc, Arbitrum, Polygon).++### Gaps+- Nevermined's current token status was not researched in this pass; treat as `[TOKEN]` until checked.+- Whether SendAI's token is live or defunct was not verified beyond the Solana Compass listing.diff --git a/docs/research/oss-landscape-2026-10-notes/agent-frameworks.md b/docs/research/oss-landscape-2026-10-notes/agent-frameworks.mdnew file mode 100644index 000000000..6e38a23ff--- /dev/null+++ b/docs/research/oss-landscape-2026-10-notes/agent-frameworks.md@@ -0,0 +1,268 @@+# Open-source AI agent frameworks and personal-assistant platforms vs three.ws (state as of 2026-10-09)++Method notes for the writer:+- Every star / fork / license SPDX figure below was read from the GitHub REST API (`gh api repos/<owner>/<repo>`, authenticated) on **2026-10-09 05:14 UTC**. The cited URL is the repo page; the API returns the same numbers the page shows. "NOASSERTION" means GitHub could not classify the LICENSE file, and the exact license was then read from the raw LICENSE file (cited).+- The three.ws baseline (what it already ships) was read read-only from the repo: the surfaces list, `STRUCTURE.md`, `workers/agent-gateway/README.md`, `api/_lib/agents-v1/runs.js`, `docs/agent-runtime.md`, `docs/memory.md`, `docs/agent-skills.md`, `docs/multi-agent.md`, `docs/agent-vitals.md`, `package.json`. Repo paths are given as `repo:<path>` because they have no URL.+- Verdict vocabulary: **adopt code** (license lets an Apache-2.0 commercial platform vendor or depend on it; module named), **adopt design, rebuild** (license or language makes copying impractical or forbidden), **already covered**.+- CRYPTO FLAG marks repos tied to a crypto project. three.ws has a commit gate on naming any coin other than $THREE, so anything from these repos must not be named in committed files without owner approval.++## three.ws baseline used to judge "lacks" (from the repo, read-only)++- Chat channels: a shared gateway core serves **Telegram and Discord only** (webhook receivers `api/gateway/telegram.js`, `api/gateway/discord.js`; adapters `workers/agent-gateway/src/adapters/{telegram,discord}.js`), with voice notes, photos, slash commands, and trade previews with Approve / Cancel buttons. No Slack, WhatsApp, Signal, iMessage, Teams or Google Chat adapter exists. `repo:workers/agent-gateway/README.md`+- Durable runs: agents-v1 runs are DB-backed ("A run never lives in one process"; every model call, tool call and result is a step row; `schedule_cron` / `scheduled_for` fields). No graph/DAG API, no state editing or time travel. `repo:api/_lib/agents-v1/runs.js`+- Human-in-the-loop: a seven-layer transaction guard with a per-tool approval policy (`never` / `required` / `always`), dollar caps, rolling windows and an auto-execute ceiling, scoped to fund-moving tools. `repo:docs/agent-runtime.md`+- Memory: two layers, a salience-ranked in-memory store plus Markdown files with YAML frontmatter, one file per topic, four memory types. `repo:docs/memory.md`; packaged as `@three-ws/agent-memory` (surfaces list).+- Skills: an in-app skill bundle format (`manifest.json` + `tools.json` + `handlers.js`) with a marketplace, skill bundles, trials and x402 per-call pricing; plus an external Agent Skills pack that follows the agentskills.io standard. `repo:docs/agent-skills.md`, surfaces list.+- Multi-agent: a frontend `<agent-stage>` that hosts several `<agent-3d>` characters with turn-taking and orchestration patterns; no server-side crew/handoff abstraction. `repo:docs/multi-agent.md`+- Observability: `agent-vitals` answers "can it act right now" (liveness of credentials, models, workers); it is not per-run LLM tracing. `repo:docs/agent-vitals.md`. `package.json` has `ai`, `@ai-sdk/anthropic`, `@ai-sdk/openai` and none of langfuse / opentelemetry / langgraph / mastra / inngest.+- Agent-to-agent: a native a2a payment mandate path (`api/agents/a2a-mandate.js`, `a2a-call.js`), not the Google A2A protocol. `repo:docs/a2a-payments.md`+- Visual builders: Scene Studio, Tour Builder and Diorama are 3D/tour builders; there is no node-graph builder for agent logic (STRUCTURE.md rows, surfaces list).+- Already strong (do not recommend): MCP servers, x402 paid APIs, marketplace, custodial Solana wallets, coin launchpad, embeddable avatars, multiplayer world, voice cloning, smart-home, iOS / Mac / desktop companion apps (surfaces list).++## Key question 1: Which agent frameworks have the most traction in 2026 and what do users praise?++### Takeaway+Two personal-assistant harnesses born in late 2025 now dwarf every developer framework: OpenClaw (391k stars, the most-starred project on GitHub) and Hermes Agent (252k). Among frameworks, the visual/low-code platforms (n8n 207k, AutoGPT 187k, Dify 158k, Langflow 155k) lead on stars, while the code-first libraries cluster at 20k to 60k (AutoGen 61k in maintenance, CrewAI 59k, Goose 55k, LangGraph 43k, Agno 43k, OpenAI Agents 30k, smolagents 30k, Mastra 29k, Letta 25k). What users praise is consistent: talk-to-it setup in the chat app they already use, memory they can read in plain files, proactive scheduled behavior, and an installable skill ecosystem.++### Cited Findings++Star table, all observed 2026-10-09 05:14 UTC via the GitHub API (stars | forks | SPDX per API | last push | created):+- openclaw/openclaw: 391,499 | 82,300 | MIT | 2026-10-09 | created 2025-11-24. Latest release v2026.9.9 on 2026-10-08. [Source](https://github.com/openclaw/openclaw)+- NousResearch/hermes-agent: 252,089 | 54,397 | MIT | 2026-10-09 | created 2025-07-22. Latest release v0.21.6 on 2026-10-08. [Source](https://github.com/NousResearch/hermes-agent)+- n8n-io/n8n: 206,752 | 61,027 | NOASSERTION (Sustainable Use License, see KQ3) | 2026-10-09. [Source](https://github.com/n8n-io/n8n)+- Significant-Gravitas/AutoGPT: 187,488 | 45,933 | NOASSERTION (PolyForm Shield + MIT split, see KQ3) | 2026-10-09. [Source](https://github.com/Significant-Gravitas/AutoGPT)+- langgenius/dify: 157,953 | 24,946 | NOASSERTION (modified Apache-2.0, see KQ3) | 2026-10-09. [Source](https://github.com/langgenius/dify)+- langflow-ai/langflow: 155,408 | 10,191 | MIT | 2026-10-09. [Source](https://github.com/langflow-ai/langflow)+- browser-use/browser-use: 117,335 | 12,986 | MIT | 2026-10-07. [Source](https://github.com/browser-use/browser-use)+- OpenHands/OpenHands (redirect target of All-Hands-AI/OpenHands): 90,319 | 11,952 | MIT | 2026-10-09. [Source](https://github.com/OpenHands/OpenHands)+- FoundationAgents/MetaGPT: 70,782 | 9,000 | MIT | last push 2026-01-21 (stale). [Source](https://github.com/FoundationAgents/MetaGPT)+- mem0ai/mem0: 66,860 | 7,872 | Apache-2.0 | 2026-10-08. [Source](https://github.com/mem0ai/mem0)+- microsoft/autogen: 61,313 | 9,276 | API says CC-BY-4.0 because the root LICENSE is Creative Commons Attribution 4.0; the code is under LICENSE-CODE, MIT. Last push 2026-04-15 (maintenance). [Source LICENSE](https://raw.githubusercontent.com/microsoft/autogen/main/LICENSE); [Source LICENSE-CODE](https://raw.githubusercontent.com/microsoft/autogen/main/LICENSE-CODE); [repo](https://github.com/microsoft/autogen)+- crewAIInc/crewAI: 59,478 | 8,663 | MIT | 2026-10-08. [Source](https://github.com/crewAIInc/crewAI)+- FlowiseAI/Flowise: 55,495 | 25,066 | NOASSERTION (Apache-2.0 + commercial enterprise dir, see KQ3) | last push 2026-08-13. [Source](https://github.com/FlowiseAI/Flowise)+- aaif-goose/goose (redirect target of block/goose): 55,093 | 6,390 | Apache-2.0 | 2026-10-09. Org `aaif-goose` describes itself as "The goose ai agent platform", homepage goose-docs.ai. [Source](https://github.com/aaif-goose/goose)+- langchain-ai/langgraph: 42,925 | 7,295 | MIT | 2026-10-08. Latest release 1.2.14 on 2026-10-06. [Source](https://github.com/langchain-ai/langgraph)+- agno-agi/agno: 42,627 | 6,113 | Apache-2.0 | 2026-10-09. [Source](https://github.com/agno-agi/agno)+- langfuse/langfuse: 35,546 | 3,953 | NOASSERTION (MIT core + ee/, see KQ3) | 2026-10-09. [Source](https://github.com/langfuse/langfuse)+- openai/openai-agents-python: 29,924 | 4,855 | MIT | 2026-10-08. [Source](https://github.com/openai/openai-agents-python)+- huggingface/smolagents: 29,743 | 3,051 | Apache-2.0 | 2026-10-06. [Source](https://github.com/huggingface/smolagents)+- mastra-ai/mastra: 28,654 | 2,946 | NOASSERTION (Apache-2.0 core + ee/, see KQ3) | 2026-10-09. [Source](https://github.com/mastra-ai/mastra)+- a2aproject/A2A: 26,078 | 2,659 | Apache-2.0 | 2026-10-08. [Source](https://github.com/a2aproject/A2A)+- letta-ai/letta: 25,082 | 2,644 | Apache-2.0 | last push 2026-09-10; README says current source moved to letta-ai/letta-code. [Source](https://github.com/letta-ai/letta)+- google/adk-python: 21,753 | 4,120 | Apache-2.0 | 2026-10-09. [Source](https://github.com/google/adk-python)+- pydantic/pydantic-ai: 20,495 | 2,885 | MIT | 2026-10-09. [Source](https://github.com/pydantic/pydantic-ai)+- camel-ai/owl: 20,153 | 2,296 | no LICENSE file; README says "The source code is licensed under Apache 2.0." [Source](https://raw.githubusercontent.com/camel-ai/owl/main/README.md)+- elizaOS/eliza: 19,565 | 5,787 | MIT | 2026-10-09 | created 2024-07-09. CRYPTO FLAG. [Source](https://github.com/elizaOS/eliza)+- camel-ai/camel: 17,836 | 2,108 | Apache-2.0 | 2026-10-05. [Source](https://github.com/camel-ai/camel)+- microsoft/agent-framework: 14,023 | 2,448 | MIT | 2026-10-08 | created 2025-04-28. Releases python-1.21.0 (2026-10-08), dotnet-1.24.0 (2026-10-07). [Source](https://github.com/microsoft/agent-framework)+- HKUDS/AutoAgent: 9,805 | 1,359 | MIT | last push 2025-10-16 (stale). [Source](https://github.com/HKUDS/AutoAgent)+- anthropics/claude-agent-sdk-python: 8,229 | 1,330 | MIT | 2026-10-09. [Source](https://github.com/anthropics/claude-agent-sdk-python)+- inngest/inngest: 5,927 | 373 | NOASSERTION (SSPL + "Apache 2.0 Future License", see KQ3) | 2026-10-08. [Source](https://github.com/inngest/inngest)+- openai/openai-agents-js: 3,902 | 988 | MIT | 2026-10-08. [Source](https://github.com/openai/openai-agents-js)+- game-by-virtuals/game-node: 93 | 91 | MIT | last push 2026-03-06. CRYPTO FLAG. [Source](https://github.com/game-by-virtuals/game-node)++Traction context:+- Hacker News thread "OpenClaw surpasses React to become the most-starred software project on GitHub". [Source](https://news.ycombinator.com/item?id=47217812)+- HN thread "OpenClaw (ClawdBot) joins OpenAI" (dated "7 months ago" on the page, so roughly Feb/Mar 2026), linking a Sam Altman post; commenters say the MIT license stays and the project moves to a foundation that, at the time, had no board or governance documents. [Source](https://news.ycombinator.com/item?id=47027907). The current LICENSE reads "Copyright (c) 2026 OpenClaw Foundation". [Source](https://raw.githubusercontent.com/openclaw/openclaw/main/LICENSE)+- OpenClaw grew from roughly 9,000 stars to over 382,000 by mid-2026 (secondary roundup). [Source](https://www.aimagicx.com/blog/best-open-source-ai-agent-frameworks-2026)+- LangGraph is credited with 34.5M monthly downloads and reached 1.0 in October 2025 (secondary roundup). [Source](https://www.firecrawl.dev/blog/best-open-source-agent-frameworks)+- CrewAI README: "over 100,000 developers certified through our community courses at learn.crewai.com". [Source](https://raw.githubusercontent.com/crewAIInc/crewAI/main/README.md)+- Microsoft Agent Framework: secondary sources place the 1.0 GA on April 3, 2026 (one says March 2026), merging AutoGen and Semantic Kernel, with new feature work going only to Agent Framework and v1.x of the old projects getting security fixes. [Source](https://www.digitalapplied.com/blog/microsoft-agent-framework-1-0-dotnet-python-guide); [Source](https://atlan.com/know/ai-agent/microsoft/agent-framework/); [Source](https://www.theagentecosystem.com/blog/agent-framework-consolidation-2026)+- Goose is "part of the Agentic AI Foundation (AAIF) at the Linux Foundation" (README). [Source](https://raw.githubusercontent.com/aaif-goose/goose/main/README.md)+- Flowise was acquired by Workday in August 2025 (secondary). [Source](https://futureagi.com/blog/dify-vs-flowise-vs-langflow-2026/)+- DataStax Langflow on Astra was retired in April 2026; Langflow now sits inside IBM's orbit after IBM's acquisition of DataStax (secondary). [Source](https://futureagi.com/blog/dify-vs-flowise-vs-langflow-2026/)++What users praise (primary HN threads):+- "n8n but you create workflows by talking to it" and "it made it easy to attach it to whatever 'channel' you're comfortable with" are the recurring explanations of OpenClaw's appeal; non-programmers feel "empowered to automate things that they could not otherwise automate". [Source](https://news.ycombinator.com/item?id=47219250)+- "Ask HN: Who is using OpenClaw?" reports use through WhatsApp/Telegram/Discord for daily tasks, an Obsidian second brain, calorie/workout logging, bill and birthday reminders, morning calendar briefings, support-email triage, Jira tickets, PR review, and family story archiving through a group chat. Praised: natural-language input, memory "stored in version control that I can read and edit", proactive behavior (chasing overdue tasks, flagging declined meetings), phone access with minimal setup. Complaints: about $100 a month in tokens for one user; "the morning debriefer worked maybe once or twice a week and broke every other morning"; broad access and API keys are risky; "the whole thing seems designed around getting people to burn more tokens". [Source](https://news.ycombinator.com/item?id=47783940)+- HN also carries sustained skepticism ("just a toy", spam agents posting to public forums, lack of restriction by default). [Source](https://news.ycombinator.com/item?id=47219250)+- Goose: multiple users praise the recipes system as "the feature that separates Goose from everything else"; MCP-first extension design; 15+ providers configurable per session; tool errors are returned to the model instead of aborting (secondary review summarizing Reddit/GitHub/HN). [Source](https://aitoolanalysis.com/goose-ai-review/)+- Agno: a developer choosing a framework praised "the clarity and modularity of Agno's Python components"; complaints are a smaller ecosystem and Python-only. [Source](https://huggingface.co/blog/aovabo/choosing-an-ai-agent-framework)+- Letta: praise centers on self-editing memory and the Agent Development Environment that lets you inspect memory tiers and tool calls live; "steeper learning curve". [Source](https://www.promptquorum.com/power-local-llm/letta-review)+- Mastra: almost no direct developer praise found; one developer called its workflow engine "great for serverless apps in TypeScript". [Source](https://huggingface.co/blog/aovabo/choosing-an-ai-agent-framework)++### Inferences+- The star gap (OpenClaw + Hermes at 640k combined vs under 60k for any code-first framework) says the market rewards a finished personal assistant that lives in chat apps, not a library. three.ws's agent is closer to this category than to LangGraph, so OpenClaw and Hermes are the real benchmarks.+- The praised features map almost one to one onto gaps in three.ws: more channels than Telegram/Discord, user-readable memory in git, proactive cron output pushed into chat, and a self-growing skill library.+- Roundup star figures are weeks to months stale (several quote 346k for OpenClaw, 14k for Agent Framework); the API numbers above should be used instead.++### Gaps+- No primary-source download or active-user figures for any project; only secondary "34.5M downloads" for LangGraph.+- The Microsoft Agent Framework GA date is sourced only from blogs (April 3 vs March 2026); no Microsoft release note was fetched.+- Developer sentiment for Mastra, smolagents, pydantic-ai and ADK was not found in primary threads.++## Key question 2: Which features do these repos have that three.ws lacks or does weakly, with a verdict per repo?++### Takeaway+The gaps are concentrated in six areas: messaging breadth (Slack, WhatsApp, Signal, iMessage, Teams, Google Chat), a self-improving skill loop plus cross-session conversation search, graph workflows with suspend/resume and time travel, first-class guardrails/handoffs/tracing, an OpenTelemetry-grade run tracer with evals, and a node-graph builder for agent logic. The best license-and-language fit for vendoring is Mastra (Apache-2.0, TypeScript, built on the same Vercel AI SDK three.ws already uses); OpenClaw's MIT channel adapters are the fastest route to channel breadth; Dify, n8n and the AutoGPT platform are design-only.++### Cited Findings++**OpenClaw** (https://github.com/openclaw/openclaw, 391,499 stars on 2026-10-09, MIT)+- Runs on your own computer and "meets you in the channels you already use: Discord, iMessage, Slack, Teams, Telegram, WhatsApp, and 20+ more, plus native apps for macOS, iOS, Android, Windows, and Linux"; one Gateway runs as a personal assistant or a shared team deployment, "configuration is the only difference". [Source](https://raw.githubusercontent.com/openclaw/openclaw/main/README.md)+- Architecture: the Gateway is "the local control plane for sessions, tools, events, and channel connections"; Control UI, CLI and TUI connect to it; Channels cover "WhatsApp, Telegram, Slack, Discord, Google Chat, Signal, iMessage, and other messaging services"; companion apps and nodes "add voice, Canvas, camera, screen, and device-local actions". Models and agent harnesses (Claude, Codex, local models) are swappable plugins; by default it phones home only for a daily version check. [Source](https://raw.githubusercontent.com/openclaw/openclaw/main/README.md)+- ClawHub is the official skill store, described as "700+ skills" by the awesome list; the agent can search ClawHub and pull in skills as needed. [Source](https://github.com/SamurAIGPT/awesome-openclaw). Koi Security reviewed "all 2,857 skills on ClawHub" and found 341 malicious ones (late January 2026, mostly macOS Atomic Stealer); ClawHub then integrated VirusTotal and ClawScan. [Source](https://thehackernews.com/2026/02/researchers-find-341-malicious-clawhub.html); [Source](https://thehackernews.com/2026/02/openclaw-integrates-virustotal-scanning.html)+- Security record: CVE-2026-25253 (CVSS 8.8) one-click RCE in versions before 2026.1.29; the gateway bound to 0.0.0.0:18789 by default and Censys counted over 30,000 internet-exposed instances on 2026-02-08; Unit 42 calls the setup a "lethal trifecta" where persistent memory "acts as an accelerant". [Source](https://aviatrix.ai/threat-research-center/openclaw-2026-clawhub-malicious-skills/); [Source](https://unit42.paloaltonetworks.com/openclaw-ai-supply-chain-risk/)+- three.ws lacks: Slack / WhatsApp / Signal / iMessage / Teams / Google Chat channels (has Telegram + Discord, `repo:workers/agent-gateway/README.md`); team (shared) deployment mode of one gateway; device nodes exposing camera/screen/Canvas to the agent from the companion apps; skill auto-discovery and install by the agent itself from a registry.+- Verdict: **adopt code** (MIT). Vendor or depend on the channel adapters for Slack, WhatsApp, Signal, iMessage, Teams and Google Chat and plug them into the existing `gateway_inbox` worker; adopt the design of the security hardening (skill scanning, bind-to-localhost) rather than the defaults that produced CVE-2026-25253.++**Hermes Agent** (https://github.com/NousResearch/hermes-agent, 252,089 stars on 2026-10-09, MIT, (c) 2025 Nous Research)+- README: "the only agent with a built-in learning loop: it creates skills from experience, improves them during use, nudges itself to persist knowledge, searches its own past conversations, and builds a deepening model of who you are across sessions"; runs "on a $5 VPS, a GPU cluster, or serverless"; any model via Nous Portal, OpenRouter, OpenAI or your own endpoint; Nous Portal offers 300+ models and a Tool Gateway (web search, FAL image generation, OpenAI TTS, Browser Use cloud browser). [Source](https://raw.githubusercontent.com/NousResearch/hermes-agent/main/README.md)+- It ships a migration path from OpenClaw that imports SOUL.md (persona), MEMORY.md and USER.md, user skills, the command allowlist, messaging settings and API keys. [Source](https://raw.githubusercontent.com/NousResearch/hermes-agent/main/README.md)+- Secondary sources: memory uses SQLite FTS5 full-text search with LLM summarization for cross-session recall rather than a vector DB; roughly 160 bundled skills compatible with the agentskills.io standard; gateway to Telegram, Discord, Slack, WhatsApp, Signal, Teams (one source says 20+ platforms); built-in cron that delivers output to any connected platform; isolated subagents; Python scripts that call tools over RPC to collapse multi-step pipelines; seven terminal backends (local, Docker, SSH, Singularity, Modal, Daytona, Vercel Sandbox). [Source](https://dev.to/wonderlab/one-open-source-project-a-day-no40-hermes-agent-nous-researchs-self-improving-ai-agent-4ale); [Source](https://hermes-agent.nousresearch.com/); [Source](https://www.width.ai/post/what-is-hermes-ai-agent); [Source](https://lucaberton.com/blog/what-is-hermes-agent-nous-research/)+- three.ws lacks: an agent that authors and refines its own skills from repeated tasks; full-text search over the agent's own past conversations; pluggable sandboxed execution backends; subagent spawning; a cron whose output is delivered into the chat channel (three.ws has `schedule_cron` on runs but delivery to a channel is a separate notification lane).+- Verdict: **adopt design, rebuild** for the learning loop and FTS conversation memory (Python codebase; three.ws is JS); **adopt code** is possible for the skill format since both already follow agentskills.io (`repo:docs/agent-skills.md`).++**LangGraph** (https://github.com/langchain-ai/langgraph, 42,925 stars on 2026-10-09, MIT)+- Official overview: "durable execution, streaming, human-in-the-loop, and more"; agents "persist through failures and can run for extended periods"; humans can review and change agent state at any point during a run; short-term working memory and long-term memory across sessions; LangSmith for tracing and a deployment platform "for long running, stateful workflows". [Source](https://docs.langchain.com/oss/python/langgraph/overview)+- three.ws lacks: a graph API over its DB-backed runs; editing a paused run's state; replay / time travel. It already has persistence of steps and cron scheduling (`repo:api/_lib/agents-v1/runs.js`).+- Verdict: **adopt design, rebuild** on the existing runs engine (adding interrupt / resume / state edit semantics); the JS port could be depended on under MIT, but the existing step-row engine already covers persistence, so vendoring a second runtime adds little.++**CrewAI** (https://github.com/crewAIInc/crewAI, 59,478 stars on 2026-10-09, MIT)+- "Crews" (role-based autonomous agent collaboration) and "Flows" (event-driven automations combining precise workflow control, single LLM calls, and native Crews); the commercial AMP Suite adds a Crew Control Plane with tracing, metrics, logs and traces. [Source](https://raw.githubusercontent.com/crewAIInc/crewAI/main/README.md)+- three.ws lacks: server-side role-based crews and event-driven flows (multi-agent is a frontend stage, `repo:docs/multi-agent.md`).+- Verdict: **adopt design, rebuild** (Python).++**AutoGen / Microsoft Agent Framework** (https://github.com/microsoft/autogen 61,313 stars, code MIT, maintenance since 2026-04-15; https://github.com/microsoft/agent-framework 14,023 stars, MIT)+- Agent Framework README: production-grade agents in .NET and Python (Go in a separate repo); "graph-based workflows supporting sequential, concurrent, handoff, and group collaboration patterns; includes checkpointing, streaming, human-in-the-loop, and time-travel"; a middleware pipeline for request/response processing. [Source](https://raw.githubusercontent.com/microsoft/agent-framework/main/README.md)+- three.ws lacks: handoff and group-collaboration orchestration, checkpoint time travel, a middleware pipeline around agent calls.+- Verdict: **adopt design, rebuild** (Python/.NET).++**OpenAI Agents SDK** (https://github.com/openai/openai-agents-python 29,924 stars, MIT; https://github.com/openai/openai-agents-js 3,902 stars, MIT)+- Core concepts: Agents; Sandbox agents "preconfigured to work with a container to perform work over long time horizons"; Realtime agents (voice with gpt-realtime-2.1); Voice pipelines (STT, agent, TTS); Agents as tools / Handoffs; Tools (functions, MCP, hosted); Guardrails for input and output validation; Human in the loop; Sessions; Tracing. Provider-agnostic across 100+ LLMs. [Source](https://raw.githubusercontent.com/openai/openai-agents-python/main/README.md)+- three.ws lacks: first-class guardrails as validators on input/output, handoffs between agents, built-in run tracing, sandbox agents.+- Verdict: **adopt code** (MIT, the JS SDK `@openai/agents` is TypeScript) for guardrails / handoffs / tracing if a second agent runtime is acceptable; otherwise adopt the design onto the AI SDK loop already in `repo:api/_lib/agent-loop.js`.++**Dify** (https://github.com/langgenius/dify, 157,953 stars on 2026-10-09, modified Apache-2.0)+- "Build Agentic workflows, RAG pipelines, with rich AI model and tool support on one collaborative workspace" (repo description). [Source](https://github.com/langgenius/dify)+- three.ws lacks: a collaborative visual builder for agent workflows and RAG pipelines.+- Verdict: **adopt design, rebuild**. The license forbids operating a multi-tenant environment without written authorization and three.ws is multi-tenant by construction (see KQ3).++**n8n** (https://github.com/n8n-io/n8n, 206,752 stars on 2026-10-09, Sustainable Use License)+- "Fair-code workflow automation platform with native AI capabilities. Combine visual building with custom code" (repo description). [Source](https://github.com/n8n-io/n8n)+- HN commenters describe OpenClaw's value as "n8n but you talk to it", which positions n8n's visual integration graph as the benchmark for business glue. [Source](https://news.ycombinator.com/item?id=47027907)+- three.ws lacks: a visual integration/automation graph with triggers.+- Verdict: **adopt design, rebuild** (embedding requires a separate n8n agreement, see KQ3).++**Flowise** (https://github.com/FlowiseAI/Flowise, 55,495 stars on 2026-10-09, Apache-2.0 outside the enterprise dir)+- "Build AI Agents, Visually" (repo description); last push 2026-08-13. [Source](https://github.com/FlowiseAI/Flowise)+- Verdict: **adopt code** (Apache-2.0 portions, TypeScript) for the visual node canvas and agent-flow node types; exclude `packages/server/src/enterprise` and `IdentityManager.ts` (see KQ3). Slower cadence after the Workday acquisition is a maintenance risk.++**Langflow** (https://github.com/langflow-ai/langflow, 155,408 stars on 2026-10-09, MIT)+- "a powerful tool for building and deploying AI-powered agents and workflows" (repo description), drag-and-drop builder. [Source](https://github.com/langflow-ai/langflow)+- Verdict: **adopt code** (MIT) for the React flow canvas if a visual agent builder is pursued; the runtime is Python.++**AutoGPT** (https://github.com/Significant-Gravitas/AutoGPT, 187,488 stars on 2026-10-09, PolyForm Shield for `autogpt_platform`, MIT elsewhere)+- README: "Describe an outcome in plain English or shape every step in the visual builder, then run the agent on demand, on a schedule, or from a trigger"; four surfaces including AutoPilot ("Describe the job in plain English and turn the conversation into a working agent") and an Agents dashboard "showing statuses, runs, and costs". [Source](https://raw.githubusercontent.com/Significant-Gravitas/AutoGPT/master/README.md)+- three.ws lacks: conversation-to-agent authoring (AutoPilot), trigger-based runs, a per-agent run/cost dashboard for end users (three.ws has an ops Agent Monitor and an Earned page, not a per-run cost ledger per the surfaces list).+- Verdict: **adopt design, rebuild** (PolyForm Shield forbids competing use, see KQ3).++**OpenHands** (https://github.com/OpenHands/OpenHands, 90,319 stars on 2026-10-09, MIT)+- "AI-Driven Development" coding agent (repo description). [Source](https://github.com/OpenHands/OpenHands)+- Verdict: **already covered / adopt code** only if a sandboxed coding-agent runtime is wanted; three.ws already has "Take the wheel" and the agent screen caster (surfaces list). Feature detail not researched beyond the description.++**Letta** (https://github.com/letta-ai/letta, 25,082 stars on 2026-10-09, Apache-2.0; active code in letta-ai/letta-code)+- README: install `@letta-ai/letta-code` from npm; terminal UI, App Server for self-hosting, desktop app, chat.letta.com, "Slack, Telegram, Discord, and custom channels", a TypeScript Letta Agent SDK, Letta Cloud for memory across computers; the `archive` branch holds the retired V1 server. [Source](https://raw.githubusercontent.com/letta-ai/letta/main/README.md)+- "Our next phase": memory moves "from specialized memory tools that edit memory in a database to generalized computer use tools like bash that operate over memory projected into git-backed files (context repositories)"; `core_memory_replace` and other legacy server memory tools removed; server-side sleep-time agents replaced by a client-side subagent system; server-side MCP replaced by client-side skills; tool rules deprecated; templates replaced by the Letta Code SDK. [Source](https://www.letta.com/blog/our-next-phase)+- three.ws lacks: git-backed "context repositories" the agent edits with ordinary file tools; a visual memory debugger (ADE, secondary). Its Markdown-file memory layer (`repo:docs/memory.md`) is already close.+- Verdict: **adopt design** (the memory layer is near-covered); the Apache-2.0 TypeScript SDK could be depended on, but the license of letta-ai/letta-code was not verified in this session.++**Mastra** (https://github.com/mastra-ai/mastra, 28,654 stars on 2026-10-09, Apache-2.0 core)+- README: TypeScript; model routing to 40+ providers; agents; graph-based workflow engine with `.then()`, `.branch()`, `.parallel()`; human-in-the-loop via suspend and resume ("pause indefinitely and resume where you left off" using storage); conversation history, RAG, "Observational Memory"; author MCP servers; built-in evals and observability; integrates with Vercel AI SDK UI and CopilotKit. [Source](https://raw.githubusercontent.com/mastra-ai/mastra/main/README.md)+- three.ws lacks: a typed workflow DSL with suspend/resume, built-in evals, an observability layer; it already uses the AI SDK (`repo:package.json`).+- Verdict: **adopt code** (Apache-2.0, same language and same AI SDK base). Candidate modules: `@mastra/core` workflows + suspend/resume, `@mastra/core` evals, Observational Memory. Avoid every `ee/` directory (`@mastra/core/auth/ee`, `@mastra/core/agent-builder/ee`, `@mastra/editor/ee`).++**Agno** (https://github.com/agno-agi/agno, 42,627 stars on 2026-10-09, Apache-2.0)+- README: AgentOS runtime + UI; "50+ endpoints with SSE and websockets"; sessions, memory, knowledge and traces stored in your own DB; 100+ toolkits; Context Providers (Slack, Drive, wikis, MCP); "Human approval: Pause runs for user confirmation. Block tools that require admin approval."; OpenTelemetry tracing, run history, audit logs; JWT-based RBAC; "turn your agent platform into a learning loop with simulations and usage data"; a coding-agent prompt bootstraps the platform from `agentos-railway` / `agentos-gcp` templates. [Source](https://raw.githubusercontent.com/agno-agi/agno/main/README.md)+- three.ws lacks: OTel traces per run with an audit log UI; admin-approval gating for arbitrary (non-financial) tools; simulations.+- Verdict: **adopt design, rebuild** (Python).++**smolagents** (https://github.com/huggingface/smolagents, 29,743 stars on 2026-10-09, Apache-2.0)+- "a barebones library for agents that think in code" (repo description). [Source](https://github.com/huggingface/smolagents)+- Verdict: **adopt design** (the code-agent pattern, where the model writes code that calls tools, which Hermes also uses to collapse turns). Not researched further.++**Goose** (https://github.com/aaif-goose/goose, 55,093 stars on 2026-10-09, Apache-2.0)+- README: general-purpose agent, native desktop app for macOS/Linux/Windows, CLI, and an API to embed anywhere; built in Rust; 15+ providers; "Use API keys or your existing Claude, ChatGPT, or Gemini subscriptions via ACP"; 70+ MCP extensions; "Custom Distributions: build your own goose distro with preconfigured providers, extensions, and branding"; governance under the Agentic AI Foundation at the Linux Foundation. [Source](https://raw.githubusercontent.com/aaif-goose/goose/main/README.md)+- Users praise recipes (shareable, parameterized agent task definitions). [Source](https://aitoolanalysis.com/goose-ai-review/)+- three.ws lacks: shareable recipes as a first-class object; white-label custom distributions of its agent; ACP subscription auth.+- Verdict: **adopt design** for recipes and custom distros (Rust codebase); **adopt code** is license-clean (Apache-2.0) if a Rust component is ever wanted.++**ElizaOS** (https://github.com/elizaOS/eliza, 19,565 stars on 2026-10-09, MIT) CRYPTO FLAG+- README: TypeScript framework plus product stack; the Eliza app covers "chat, voice, memory, knowledge, and document workflows; messaging and workspace connectors; calendar, reminders, inbox, goals, health; browser and desktop automation; camera, phone, messages, contacts, location, and other native device bridges; non-custodial EVM and Solana wallet operations with approval boundaries; scheduled workflows, coding-agent orchestration, and installable app views". Plugins register "actions, providers, evaluators, services, model handlers, routes, events, tests, and app views". Eliza-1 local inference ships Gemma 4 based 2B/4B/9B/27B tiers plus embeddings, speech, vision and image generation for offline use. The project "no longer accept[s] third-party plugins or registry items". [Source](https://raw.githubusercontent.com/elizaOS/eliza/develop/README.md); [Source](https://github.com/elizaos/eliza)+- Founder Shaw Walters declared the ELIZAOS token "dead" on 2026-08-04 and wound down the foundation after settling a Burwick Law class action; he said Eliza Labs keeps building the software. [Source](https://www.theblock.co/post/410774/eliza-labs-native-token-dead); [Source](https://decrypt.co/374958/eliza-ai-token-dead-shuts-down-foundation-lawsuit)+- three.ws lacks: native device bridges (contacts, location, phone) exposed to the agent; offline local-inference tier; installable app views inside the agent app.+- Verdict: **adopt design, rebuild**. MIT would permit code, but any committed reference to ElizaOS trips the commit gate; also three.ws is custodial, ElizaOS is non-custodial.++**Virtuals GAME** (https://github.com/game-by-virtuals/game-node, 93 stars on 2026-10-09, MIT) CRYPTO FLAG+- Agents are created by staking 100 VIRTUAL and launched on a bonding curve, graduating to a Uniswap pool at 42,000 VIRTUAL (secondary). [Source](https://www.koinx.com/blog/virtuals-protocol-explained)+- Verdict: **already covered** (three.ws ships a native bonding-curve launchpad and agent token plans per the surfaces list). Do not name in commits.++**Google A2A protocol / ADK** (https://github.com/a2aproject/A2A 26,078 stars; https://github.com/google/adk-python 21,753 stars; both Apache-2.0)+- A2A is "an open protocol enabling communication and interoperability between opaque agentic applications" (repo description). [Source](https://github.com/a2aproject/A2A)+- three.ws lacks: an A2A agent card / task endpoint, so external A2A agents cannot discover or hire three.ws agents; its a2a payments are a native mandate scheme (`repo:docs/a2a-payments.md`).+- Verdict: **adopt code** (Apache-2.0 spec and SDKs): publish an A2A agent card per three.ws agent and front the existing hire/x402 path with the A2A task API.++**Supporting infrastructure worth noting**+- mem0 (Apache-2.0, 66,860 stars) is a drop-in memory layer; Langfuse core is MIT (35,546 stars) for traces and evals; Inngest (5,927 stars) is SSPL now with an Apache-2.0 future license; browser-use (MIT, 117,335 stars) for browser automation. Verdicts: mem0 and Langfuse are **adopt code** candidates for observability and memory retrieval; Inngest is **adopt design** until its SSPL term converts (see KQ3). Sources in the star table above.++### Inferences+- The single highest-leverage gap is channel breadth: the two most-starred projects on GitHub both win on "the agent in the chat app you already use", and three.ws's gateway worker is already structured for additional adapters.+- Mastra is the only top-tier framework that matches three.ws on language, license and base library, so it is the natural vendoring target for workflows, suspend/resume and evals.+- Three.ws's financial HITL guard is stronger than any framework's generic approval; the gap is generalizing it to non-financial tool approval and to pausing a whole run.+- The ClawHub malware wave (341 of 2,857 skills) is a cautionary design input for the three.ws skill marketplace: scanning before listing is now table stakes.++### Gaps+- Feature lists for OpenHands, smolagents, pydantic-ai, ADK, CAMEL/OWL and MetaGPT were not read beyond repo descriptions.+- Dify and n8n feature depth (number of integrations, node types) is from descriptions and secondary roundups only; the primary README bodies were not fetched past the banner.+- Hermes feature counts (platforms, backends, skills) come from secondary blogs; the README section with the feature bullets was not captured.+- Letta's exact deprecation dates and the letta-code license were not verified.+- Mastra's `ee/LICENSE` text was not read (only the LICENSE.md pointer).++## Key question 3: What does each license allow and forbid for commercial reuse by an Apache-2.0 project?++### Takeaway+Permissive (MIT / Apache-2.0) covers OpenClaw, Hermes, LangGraph, CrewAI, AutoGen code, Agent Framework, OpenAI Agents, Langflow, OpenHands, Letta, Mastra core, Agno, smolagents, Goose, ElizaOS, ADK, A2A, mem0, Langfuse core, browser-use, Flowise outside its enterprise dir and AutoGPT outside `autogpt_platform`. Three repos are off-limits for code: Dify (multi-tenant needs written authorization; frontend logo must stay), n8n (internal business use only, embedding needs a separate agreement), AutoGPT's platform (PolyForm Shield, no competing use). Inngest is SSPL today.++### Cited Findings+- **Dify**: "licensed under a modified version of the Apache License 2.0, with the following additional conditions": (a) "Unless explicitly authorized by Dify in writing, you may not use the Dify source code to operate a multi-tenant environment" where "one tenant corresponds to one workspace"; (b) "you may not remove or modify the LOGO or copyright information in the Dify console or applications", inapplicable to uses that do not involve the `web/` frontend; contributors agree the producer "can adjust the open-source agreement to be more strict or relaxed" and that contributions may be used commercially "including but not limited to its cloud business operations"; "The interactive design of this product is protected by appearance patent." Everything else follows Apache-2.0. [Source](https://raw.githubusercontent.com/langgenius/dify/main/LICENSE). An October 2025 OSI license-discuss thread argued the terms conflict with the Open Source Definition. [Source](https://lists.opensource.org/pipermail/license-discuss_lists.opensource.org/2025-October/022425.html). For an Apache-2.0 commercial platform: backend-only reuse avoids the logo clause, but any multi-tenant deployment (every three.ws user is a workspace) needs written authorization from Dify; the appearance patent also argues against copying the UI.+- **n8n**: files with `.ee.` in the name or `.ee` in the dirname are under the n8n Enterprise License ("may only be used in production, if you ... hold a valid n8n Enterprise license"); everything else is under the Sustainable Use License 1.0: "You may use or modify the software only for your own internal business purposes or for non-commercial or personal use. You may distribute the software or provide it to others only if you do so free of charge for non-commercial purposes. You may not alter, remove, or obscure any licensing, copyright, or other notices." [Source](https://raw.githubusercontent.com/n8n-io/n8n/master/LICENSE.md); [Source](https://raw.githubusercontent.com/n8n-io/n8n/master/LICENSE_EE.md). n8n's own docs say embedding n8n in a product requires signing a separate agreement ([email protected]), and the license is source-available, not OSI open source. [Source](https://docs.n8n.io/faircode-license). For three.ws: no vendoring of n8n code into an Apache-2.0 product; design only.+- **AutoGPT**: "Everything inside the autogpt_platform folder is under the Polyform Shield License. Everything outside the autogpt_platform folder is under the MIT License" (the classic agent, Forge, AG Benchmark, classic GUI). [Source](https://raw.githubusercontent.com/Significant-Gravitas/AutoGPT/master/LICENSE). PolyForm Shield 1.0.0 is the noncompete variant (the license text is reproduced in the file, with the canonical text at polyformproject.org/licenses/shield/1.0.0); a third-party review notes it "restricts competing commercial use". [Source](https://www.promptquorum.com/power-local-llm/autogpt-local-review-2026). For three.ws, which competes as an agent platform: platform code is off-limits; the MIT classic code is stale.+- **Flowise**: content under `packages/server/src/enterprise` and files with explicit notices such as `IdentityManager.ts` are under a Commercial License; "Content outside of the above mentioned directories or restrictions above is available under the Apache 2.0 license". [Source](https://raw.githubusercontent.com/FlowiseAI/Flowise/main/LICENSE.md). Apache-2.0 portions can be vendored into an Apache-2.0 project with attribution and NOTICE retention.+- **Mastra**: `LICENSE.md` places every `ee/` directory (`@mastra/core/auth/ee`, `@mastra/core/agent-builder/ee`, `@mastra/editor/ee`) under `ee/LICENSE`; `packages/core/package.json` declares `"license": "Apache-2.0"`. [Source](https://raw.githubusercontent.com/mastra-ai/mastra/main/LICENSE.md); [Source](https://raw.githubusercontent.com/mastra-ai/mastra/main/packages/core/package.json). Apache-2.0 core is fully compatible; exclude `ee/`.+- **Langfuse**: "ee/", "web/src/ee/" and "worker/src/ee/" are under `ee/LICENSE` (an Enterprise License: "Langfuse is an open core project. Langfuse's core is permissively licensed (MIT license)"); everything else is MIT Expat. Copyright is now "ClickHouse, Inc." [Source](https://raw.githubusercontent.com/langfuse/langfuse/main/LICENSE); [Source](https://raw.githubusercontent.com/langfuse/langfuse/main/ee/LICENSE)+- **Inngest**: `LICENSE.md` is the "Server Side Public License, Version 1.0" with an "Apache 2.0 Future License" clause; the file states the Apache-2.0 grant becomes "effective on the third" anniversary (the clause was only partially read, line 564 of the file). [Source](https://raw.githubusercontent.com/inngest/inngest/main/LICENSE.md). SSPL requires offering the service source if the software is offered as a service, so vendoring into an Apache-2.0 product is not clean until the conversion date for a given version.+- **AutoGen**: root LICENSE is Creative Commons Attribution 4.0 (docs/content), `LICENSE-CODE` is MIT. [Source](https://raw.githubusercontent.com/microsoft/autogen/main/LICENSE); [Source](https://raw.githubusercontent.com/microsoft/autogen/main/LICENSE-CODE). Code is reusable under MIT; documentation needs CC-BY attribution.+- **OWL (camel-ai)**: no LICENSE file in the repo root; README states "The source code is licensed under Apache 2.0." [Source](https://raw.githubusercontent.com/camel-ai/owl/main/README.md). Treat as Apache-2.0 by declaration only; a missing LICENSE file weakens the grant's clarity.+- **OpenClaw, Hermes, ElizaOS**: plain MIT, copyright holders "OpenClaw Foundation" (2026), "Nous Research" (2025), "Shaw Walters and elizaOS Contributors" (2026). [Source](https://raw.githubusercontent.com/openclaw/openclaw/main/LICENSE); [Source](https://raw.githubusercontent.com/NousResearch/hermes-agent/main/LICENSE); [Source](https://raw.githubusercontent.com/elizaOS/eliza/develop/LICENSE)+- **Apache-2.0 repos** (Letta, Agno, smolagents, Goose, ADK, A2A, mem0, CAMEL): identical license to three.ws, so vendoring needs only the Apache NOTICE and attribution requirements. **MIT repos** (LangGraph, CrewAI, Agent Framework, OpenAI Agents, Langflow, OpenHands, browser-use, pydantic-ai, claude-agent-sdk, MetaGPT, AutoAgent, game-node): MIT text must accompany copied code. Star-table sources above.+- Model weights vs code: ElizaOS's Eliza-1 local models are "based on Gemma 4" and are downloaded at runtime; the repo's MIT license covers the code, not the Gemma weights, whose terms were not checked. [Source](https://raw.githubusercontent.com/elizaOS/eliza/develop/README.md). No other repo in this list ships weights; OpenClaw, Hermes and Goose are BYO-model.++### Inferences+- For a multi-tenant SaaS like three.ws, Dify is effectively proprietary and n8n is design-only; both are safe to study and to describe in internal docs.+- Open-core splits (Mastra, Langfuse, Flowise) are safe as long as vendoring excludes the named `ee/` or `enterprise/` directories; an import-path lint rule would enforce this mechanically.+- The Apache-2.0 TypeScript stack (Mastra core, A2A JS SDK, mem0 TS client, Langfuse core) composes cleanly with three.ws without any license notice beyond NOTICE files.++### Gaps+- PolyForm Shield's exact competing-use clause was not quoted from the file (only the file's license split and a secondary characterization).+- Inngest's conversion clause was read only in fragments; the exact mechanics (per-version date vs per-release) should be confirmed before relying on it.+- The Gemma 4 weights license referenced by ElizaOS was not retrieved.+- Mastra `ee/LICENSE` and letta-code's license file were not read.++## Key question 4: Which repos are crypto-related (commit-gate relevance)?++### Takeaway+Two repos in scope are crypto projects and must not be named in committed three.ws files without owner approval: ElizaOS (a token that the founder declared dead in August 2026, with the software continuing) and Virtuals Protocol's GAME framework (agents launched on a VIRTUAL bonding curve). Every other repo on the list is non-crypto.++### Cited Findings+- ElizaOS began as ai16z; its token reached a roughly $2.5 billion market cap in January 2025, was rebranded after Andreessen Horowitz objected to the name, and on 2026-08-04 founder Shaw Walters declared the ELIZAOS token "dead", handed the remaining treasury to token holders represented by Burwick Law under a class-action settlement, and wound down the foundation while saying Eliza Labs keeps building the software. [Source](https://www.theblock.co/post/410774/eliza-labs-native-token-dead); [Source](https://thedefiant.io/news/tokens/eliza-labs-shaw-walters-says-ai16z-token-is-dead-after-settling-burwick-law-class-action)+- The ElizaOS app includes "non-custodial EVM and Solana wallet operations with approval boundaries", and `@elizaos/plugin-solana` is published on npm. [Source](https://raw.githubusercontent.com/elizaOS/eliza/develop/README.md); [Source](https://www.npmjs.com/package/@elizaos/plugin-solana)+- Virtuals Protocol: agents are "created by staking 100 VIRTUAL tokens and launched on a bonding curve; once 42,000 VIRTUAL accumulates, the agent graduates to a Uniswap liquidity pool"; the maintained SDK is `game-by-virtuals/game-python` (the older `agentdao/virtuals-python` README points there); one aggregator claims 18,000+ agents and $75M+ cumulative revenue as of early 2026, unverified. [Source](https://www.koinx.com/blog/virtuals-protocol-explained); [Source](https://github.com/agentdao/virtuals-python); [Source](https://www.startuphub.ai/startups/virtuals-game.md)+- Non-crypto by their own descriptions: OpenClaw, Hermes Agent, LangGraph, CrewAI, AutoGen / Agent Framework, OpenAI Agents, Dify, n8n, Flowise, Langflow, AutoGPT, OpenHands, Letta, Mastra, Agno, smolagents, Goose, ADK, A2A, mem0, Langfuse, browser-use, Inngest, CAMEL/OWL, MetaGPT, AutoAgent (repo descriptions in the star table). Note that HN commenters flag OpenClaw agents being used to promote things on public forums and suspect "token sales or influencer marketing" behind some hype, but the project itself has no token. [Source](https://news.ycombinator.com/item?id=47783940)++### Inferences+- ElizaOS is the only direct feature competitor that is both an "agentic operating system" and Solana-native with wallet operations; it is the one whose design should be studied most carefully and whose name must never land in a commit.+- Virtuals' bonding-curve launch is already matched by three.ws's own launchpad, so there is nothing to adopt beyond awareness.++### Gaps+- Court documents for the Eliza settlement were not retrieved; only press reports.+- The `game-by-virtuals/game-python` repo (the maintained SDK) was not queried for stars or license; only the 93-star `game-node` repo was.diff --git a/docs/research/oss-landscape-2026-10-notes/avatars-voice.md b/docs/research/oss-landscape-2026-10-notes/avatars-voice.mdnew file mode 100644index 000000000..585165819--- /dev/null+++ b/docs/research/oss-landscape-2026-10-notes/avatars-voice.md@@ -0,0 +1,125 @@+# OSS talking-avatar / VTuber / lip-sync / voice-agent / TTS-STT repos vs three.ws embodied agents (as of 2026-10-09)++Method note: stars, SPDX license and pushed_at come from the GitHub REST API (https://api.github.com/repos/OWNER/NAME) fetched 2026-10-09. "NOASSERTION" means GitHub could not classify the LICENSE file. Claims not yet verified against a primary page are listed under Gaps. three.ws baseline (from surfaces list): `<agent-3d>` component, chat, voice lane (`@three-ws/voice`, agent voice cloning), gesture vocabulary, procedural IK, VRM/Mixamo rig support, ARKit-52 auto-rig worker, companion, /drive, /voice/home, assistant widget, terminal avatar, Motion Swap, Splat viewer.++## 1. Browser 3D talking heads and VTuber stacks++### Takeaway+No browser repo matches three.ws's breadth (embed, rigs, commerce), but TalkingHead (MIT, 3D, browser) and AITuberKit/Open-LLM-VTuber (huge feature sets, restrictive or mixed licenses) show lip-sync, multi-language visemes, streaming-TTS sync, and interruption/desktop-pet UX that three.ws should match. Only MIT/Apache code can be adopted directly.++### Cited Findings+- TalkingHead (met4citizen) https://github.com/met4citizen/TalkingHead : 1,599 stars (2026-10-09), MIT, last push 2026-09-25; "JavaScript class for real-time lip-sync using full-body 3D avatars" : [GitHub API](https://api.github.com/repos/met4citizen/TalkingHead). Closest direct competitor; MIT code so adoptable. Avatar assets (Ready Player Me / Avaturn GLBs) carry their own terms.+- three-vrm (pixiv) https://github.com/pixiv/three-vrm : 2,209 stars (2026-10-09), MIT, last push 2026-10-02 : [GitHub API](https://api.github.com/repos/pixiv/three-vrm). Already the VRM runtime layer; three.ws "VRM/Mixamo rig support" is covered; track for VRM 1.0 expression/lookAt/spring-bone features.+- AITuberKit (tegnike) https://github.com/tegnike/aituber-kit : 1,118 stars (2026-10-09), SPDX NOASSERTION, last push 2026-10-04 : [GitHub API](https://api.github.com/repos/tegnike/aituber-kit). License is a custom dual: free Non-Commercial license, commercial use needs a separate paid license (contact [email protected]) : [LICENSE](https://raw.githubusercontent.com/tegnike/aituber-kit/main/LICENSE). Cannot adopt code; adopt design only. Not Apache-compatible.+- Open-LLM-VTuber https://github.com/Open-LLM-VTuber/Open-LLM-VTuber : 14,025 stars (2026-10-09), SPDX NOASSERTION, last push 2026-05-15 (no push in ~5 months) : [GitHub API](https://api.github.com/repos/Open-LLM-VTuber/Open-LLM-VTuber). README says project is MIT-implied but bundled Live2D sample models are under Live2D Free Material License (commercial/enterprise needs permission); features: hands-free voice + voice interruption, Live2D expressions, desktop-pet transparent mode, camera/screen vision, touch feedback, proactive speech, inner thoughts display, offline local models; long-term memory "temporarily removed"; v2.0 complete rewrite is only in planning : [README](https://github.com/Open-LLM-VTuber/Open-LLM-VTuber). 2D Live2D, not 3D. Verdict: adopt design (proactive speech, inner thoughts, desktop pet; three.ws already has a desktop companion app).+- Amica (semperai) https://github.com/semperai/amica : ~1.6k stars (README page capture, 2026-10-09), MIT (3D assets under their authors' licenses); VRM display, voice chat (STT+VAD), multi-service/local TTS, Llama.cpp/Ollama/OpenAI-compatible/OpenRouter backends, Bakllava vision, emotional expressions, Tauri desktop : [GitHub](https://github.com/semperai/amica). Verdict: already covered; adopt design (emotion tags in LLM output).+- Utsuwa (The-Lab-by-Ordinary-Company) https://github.com/The-Lab-by-Ordinary-Company/utsuwa : privacy-first VRM companion, 7 LLM providers, lip-sync, relationship-progression system, local memory embeddings, MIT, macOS desktop only : [search summary of README](https://proxy-dev.blitzz.co/proxy/123456/github.com/The-Lab-by-Ordinary-Company/utsuwa) (proxy mirror; verify on GitHub). Relationship-progression is a feature three.ws /companion may lack.+- ARPAHLS/avatar https://github.com/arpahls/avatar : desktop VRM/VRMA companion, lip-sync to live/system audio, MIT; sample animations belong to pixiv VRoid and need credit for commercial use : [search summary](https://github.com/arpahls/avatar).+- Clawatar (Dongping-Chen): VRM viewer skill for AI agents over WebSocket, 162 animations, TTS lip-sync (ElevenLabs); its Mixamo-derived animations are non-commercial : [playbooks listing](https://playbooks.com/skills/openclaw/skills/clawatar). Overlaps with three.ws gestures + Spatial MCP; not adoptable.+- TEN Framework example "Lip Sync Avatars": Live2D character Kei with MotionSync lip sync, plus Trulience, HeyGen, Tavus avatar vendors : [TEN README](https://github.com/TEN-framework/ten-framework).+- ChatVRM (pixiv) and forks: not fetched (see Gaps).++### Inferences+- The strongest open 3D-in-browser rival is TalkingHead; none bundle three.ws's embed distribution, rig doctor, wallet/commerce, or in-car/house surfaces.+- Verdicts: TalkingHead = adopt design, rebuild (MIT, so code is allowed but three.ws has own viseme path; diff multi-language viseme tables and streaming-sync API); three-vrm = already covered; AITuberKit = adopt design, rebuild (license blocks code); Open-LLM-VTuber = adopt design, rebuild; Amica/Utsuwa/ARPAHLS = adopt design (relationship progression, system-audio lip-sync).++### Gaps+- ChatVRM forks, Amica license/stars, Utsuwa stars not API-verified.++## 2. Video talking-head / lip-sync models++### Takeaway+SOTA is Wan-based diffusion (InfiniteTalk, SoulX-FlashTalk-14B) with Apache-2.0 code; MuseTalk is the only truly real-time lightweight option with MIT code and commercial-use weights. All need GPU, so they fit three.ws's GPU worker fleet rather than the browser, and complement (not replace) the 3D avatars.++### Cited Findings+- LivePortrait (KwaiVGI) https://github.com/KwaiVGI/LivePortrait : 19,191 stars (2026-10-09), code MIT (LICENSE "Copyright (c) 2024 Kuaishou") : [LICENSE](https://raw.githubusercontent.com/KwaiVGI/LivePortrait/main/LICENSE); last push 2026-06-01 : [API](https://api.github.com/repos/KwaiVGI/LivePortrait). GitHub flags SPDX NOASSERTION. Weights on HF KlingTeam/LivePortrait; README does not state weight terms; depends on InsightFace (non-commercial weights per my background knowledge, UNVERIFIED) : [README](https://github.com/KwaiVGI/LivePortrait). Features: image/video-driven portrait animation, animals mode, video editing, motion templates (.pkl), regional control.+- MuseTalk (TMElyralab) https://github.com/TMElyralab/MuseTalk : 6,691 stars (2026-10-09), SPDX NOASSERTION in API but README states code MIT (academic and commercial) and weights "available for any purpose"; last push 2025-09-26 (stale); v1.5 released 2025-03-28; claims 30fps+ on V100; 256x256 face region; third-party parts (whisper, dwpose, S3FD) have own licenses : [README](https://github.com/TMElyralab/MuseTalk), [API](https://api.github.com/repos/TMElyralab/MuseTalk). Verdict: adopt code into a GPU worker (realtime lip-sync of a photo/video).+- LatentSync (ByteDance) https://github.com/bytedance/LatentSync : 6,123 stars, Apache-2.0, last push 2025-06-20 (stale) : [API](https://api.github.com/repos/bytedance/LatentSync). Weights license not verified.+- InfiniteTalk (MeiGen-AI) https://github.com/MeiGen-AI/InfiniteTalk : 8,026 stars, Apache-2.0, last push 2026-05-22 : [API](https://api.github.com/repos/MeiGen-AI/InfiniteTalk). Sparse-frame video dubbing / unlimited-length talking video on Wan.+- SoulX-FlashTalk-14B (Soul-AILab): real-time infinite-streaming audio-driven avatar via self-correcting bidirectional distillation; code and weights released 2026-01-08; HF metadata Apache-2.0; built on InfiniteTalk and Wan; 14B needs heavy GPU : [model card](https://huggingface.co/Soul-AILab/SoulX-FlashTalk-14B/raw/main/README.md). Newest SOTA found; stars not checked.+- TalkingAvatar TA2.0 (neosapience): Apache-2.0, image+audio+prompt to 480x832 25fps H.264 video, offline clips not streaming : [GitHub](https://github.com/neosapience/TA2.0).+- Hallo3 (fudan-generative-vision) https://github.com/fudan-generative-vision/hallo3 : 1,407 stars, MIT, last push 2025-03-13 (stale) : [API](https://api.github.com/repos/fudan-generative-vision/hallo3).+- EchoMimic V3 (antgroup) https://github.com/antgroup/echomimic_v3 : 1,079 stars, Apache-2.0, last push 2026-03-18 : [API](https://api.github.com/repos/antgroup/echomimic_v3).+- Others named by one blog, unchecked: OpenTalking (full STT-LLM-TTS-avatar WebRTC loop), SoulX-LiveAct, duix.ai, LiveTalking : [Atlas Cloud blog](https://www.atlascloud.ai/id/blog/guides/free-opensource-ai-avatar-heygen-alternative) (vendor blog, low reliability).++### Inferences+- three.ws's Motion Swap and GPU fleet could host MuseTalk (realtime) and InfiniteTalk/FlashTalk (quality) as "photoreal talking video" lanes. Verdict: MuseTalk adopt code (worker); InfiniteTalk/FlashTalk adopt code (heavy, Apache); LivePortrait adopt design only until weights/InsightFace terms confirmed; Hallo3/EchoMimic/LatentSync stale or lower priority.++### Gaps+- Weight licenses for LatentSync, InfiniteTalk, Hallo3, EchoMimic V3 not verified. SoulX-FlashTalk stars/fps not verified.++## 3. Real-time voice-agent frameworks++### Takeaway+Pipecat (BSD-2) and LiveKit Agents (Apache-2.0) are the permissive standards, both pushed within the day; three.ws can adopt them for transport/turn-taking. Moshi/Unmute give full-duplex speech-to-speech with CC-BY weights. TEN's license has extra restrictions.++### Cited Findings+- Pipecat https://github.com/pipecat-ai/pipecat : 16,306 stars (2026-10-09), BSD-2-Clause, pushed 2026-10-09 : [API](https://api.github.com/repos/pipecat-ai/pipecat). Daily-maintained; has avatar/video vendor integrations (not verified individually).+- LiveKit Agents https://github.com/livekit/agents : 14,663 stars, Apache-2.0, pushed 2026-10-09 : [API](https://api.github.com/repos/livekit/agents).+- TEN Framework https://github.com/TEN-framework/ten-framework : 11,155 stars, NOASSERTION ("Apache 2.0 with additional restrictions"; packages are plain Apache-2.0), pushed 2026-10-09 : [API](https://api.github.com/repos/TEN-framework/ten-framework), [README](https://github.com/TEN-framework/ten-framework). Ships TEN VAD, TEN Turn Detection (full-duplex), SIP call agent, ESP32 hardware agent, TMAN visual designer, lip-sync avatar example. Root license restrictions text not read: treat as not adoptable until read.+- Kyutai Moshi https://github.com/kyutai-labs/moshi : 11,191 stars, Apache-2.0 per API (README: Python MIT, Rust Apache, client MIT), weights CC-BY 4.0, theoretical 160 ms / ~200 ms practical on L4, last push 2026-09-09 : [API](https://api.github.com/repos/kyutai-labs/moshi), [README](https://github.com/kyutai-labs/moshi).+- Kyutai Unmute https://github.com/kyutai-labs/unmute : 1,523 stars, MIT, pushed 2026-09-09 : [API](https://api.github.com/repos/kyutai-labs/unmute). Modular STT+LLM+TTS voice wrapper.+- Sesame CSM https://github.com/SesameAILabs/csm : 14,727 stars, Apache-2.0, last push 2025-05-27 (stale) : [API](https://api.github.com/repos/SesameAILabs/csm).++### Inferences+- Verdicts: Pipecat, LiveKit Agents = already covered in spirit (three.ws has own voice lane) but adopt as optional transports; compare turn-taking/interruption; Moshi = adopt design (full-duplex), weights CC-BY ok with attribution; TEN = adopt design only (read license); CSM = stale.++### Gaps+- Whether three.ws voice lane has barge-in/semantic turn detection not checked in docs.++## 4. TTS / STT++### Takeaway+Permissive TTS options exist (Chatterbox MIT, Kokoro Apache, Orpheus Apache, Dia Apache, CosyVoice Apache); F5-TTS weights and Fish Speech are non-commercial/research licensed. VibeVoice says research-only.++### Cited Findings+- Chatterbox (Resemble AI) https://github.com/resemble-ai/chatterbox : 26,820 stars, MIT, pushed 2026-07-21 : [API](https://api.github.com/repos/resemble-ai/chatterbox). README lists MIT for repo, no separate weights terms shown; Perth neural watermark on all output; Turbo 350M English with [laugh] paralinguistic tags (one-step decoder), Nano 110M CPU ~3x realtime, Multilingual V3 500M 23+ languages : [README](https://github.com/resemble-ai/chatterbox). Verdict: adopt code (self-host TTS lane).+- Kokoro (hexgrad) https://github.com/hexgrad/kokoro : 9,227 stars, Apache-2.0, last push 2025-08-06 : [API](https://api.github.com/repos/hexgrad/kokoro).+- Fish Speech https://github.com/fishaudio/fish-speech : 32,978 stars, NOASSERTION; code AND weights under Fish Audio Research License (not commercial-free; commercial needs Fish license); S2-Pro 4B, 10M+ hours, 80+ languages, inline [whisper]/[excited] tags, multi-speaker, streaming 100 ms TTFA on H200 (RTF 0.195); pushed 2026-10-05 : [README](https://github.com/fishaudio/fish-speech), [API](https://api.github.com/repos/fishaudio/fish-speech). Not adoptable commercially without a license; adopt design (inline emotion tags).+- Orpheus-TTS (canopyai) : 6,346 stars, Apache-2.0, pushed 2025-12-05 : [API](https://api.github.com/repos/canopyai/Orpheus-TTS).+- Dia (nari-labs) : 19,400 stars, Apache-2.0, pushed 2025-11-19 : [API](https://api.github.com/repos/nari-labs/dia).+- VibeVoice (Microsoft) https://github.com/microsoft/VibeVoice : 54,709 stars, MIT, pushed 2026-10-08; models: ASR-7B (60 min single pass), ASR-Streaming, ASR-BitNet (CPU), TTS-1.5B (90 min), Realtime-0.5B; README says TTS code removed 2025-09-05 after misuse and models are research/development only, not for commercial use without further testing : [README](https://github.com/microsoft/VibeVoice), [API](https://api.github.com/repos/microsoft/VibeVoice). Do not ship in production on this README alone.+- F5-TTS https://github.com/SWivid/F5-TTS : 15,361 stars, code MIT, pretrained weights CC-BY-NC (Emilia data); pushed 2026-09-21 : [README](https://github.com/SWivid/F5-TTS), [API](https://api.github.com/repos/SWivid/F5-TTS). Weights not commercially usable.+- CosyVoice (FunAudioLLM) : 23,902 stars, Apache-2.0 code, pushed 2026-05-25; latest Fun-CosyVoice3-0.5B-2512 (Dec 2025), bi-streaming latency as low as 150 ms; weights license not stated on README : [API](https://api.github.com/repos/FunAudioLLM/CosyVoice), [README](https://github.com/FunAudioLLM/CosyVoice).+- faster-whisper (SYSTRAN) : 25,776 stars, MIT, pushed 2026-10-06 : [API](https://api.github.com/repos/SYSTRAN/faster-whisper). Whisper weights MIT (background knowledge, unverified here).+- Parakeet (NVIDIA NeMo): not fetched; weights believed CC-BY-4.0 (UNVERIFIED).++### Inferences+- Verdicts: Chatterbox, Kokoro, Orpheus, Dia, CosyVoice = adopt code as optional TTS lanes on GPU fleet after weight check; F5-TTS, Fish Speech, VibeVoice = do not adopt (weights/licence), adopt design; faster-whisper = adopt code.++### Gaps+- Chatterbox/Kokoro/Orpheus/Dia/CosyVoice weight licenses, Parakeet, other Whisper variants not verified via primary page.++## 5. Viseme / blendshape lip-sync libraries++### Takeaway+Rhubarb is the standard offline mouth-cue generator; wawa-lipsync is a tiny MIT browser analyzer; NVIDIA Audio2Face-3D is the only open audio-to-full-face-blendshape model but its weights are NVIDIA Open Model license. three.ws already emits ARKit-52 rigs, so Audio2Face-3D is the real upgrade path over viseme-only lip sync.++### Cited Findings+- Rhubarb Lip Sync https://github.com/DanielSWolf/rhubarb-lip-sync : 2,645 stars, NOASSERTION (README does not show license; believed MIT with bundled PocketSphinx, UNVERIFIED), pushed 2026-06-16; 6 base shapes A-F + G, H, X; outputs TSV/XML/JSON/DAT : [README](https://github.com/DanielSWolf/rhubarb-lip-sync), [API](https://api.github.com/repos/DanielSWolf/rhubarb-lip-sync). Offline, English-centric (background knowledge).+- wawa-lipsync (wass08) https://github.com/wass08/wawa-lipsync : 211 stars, MIT, pushed 2025-11-07 : [API](https://api.github.com/repos/wass08/wawa-lipsync). Real-time browser audio-analysis visemes.+- NVIDIA Audio2Face-3D https://github.com/NVIDIA/Audio2Face-3D : 467 stars (2026-10-09), repo itself no license; SDK + Maya/UE5 plugins MIT; training framework Apache; NIM under NVIDIA software license; models "Nvidia Open Model" (regression v2.3, diffusion v3.0); Audio2Emotion models custom (use with Audio2Face only); sample dataset evaluation-only; outputs mesh deformation, joints or blendshape weights; C++ SDK : [README](https://github.com/NVIDIA/Audio2Face-3D), [API](https://api.github.com/repos/NVIDIA/Audio2Face-3D). three.ws is an NVIDIA Inception member (surfaces list), a favorable relationship.++### Inferences+- Verdicts: Audio2Face-3D = adopt code (SDK MIT; read Open Model license for hosted use) onto ARKit-52 rigs; wawa-lipsync = already covered if three.ws has analyser-based visemes; Rhubarb = adopt design (offline cue baking for pre-recorded lines).++### Gaps+- Open Model license text not read.++## 6. Closed products (features only)++### Takeaway+Closed vendors compete on a rendering+turn-taking+perception stack: Tavus ships purpose-built models (turn-taking, perception, generative video face), Simli sells sub-300 ms speech-to-video, ElevenLabs sells breadth (languages, voices, integrations) with no avatar. All are video/voice-first, none are embeddable 3D rigs with gestures, so three.ws's 3D body is the differentiator; the gaps are perception and turn-taking models.++### Cited Findings+- Tavus: Phoenix-4.5 real-time face rendering (stated 134 ms, new identity from one image), Sparrow-2 turn-taking model (claims #1 on TurnBench), Raven-1 multimodal perception (voice, video, screen, emotion cues, <300 ms context), Griffin preview video-to-video full-duplex model, knowledge-base grounding, CVI developer API with "PALs" : [Tavus](https://www.tavus.io/) (vendor claims).+- Simli: real-time video avatars, speech-to-video under 300 ms, Gaussian "Emotive Faces", face cloning, free plan $10 at signup plus 50 minutes monthly top-up; STT/LLM/TTS latency estimates 100-500/250-450/250-1200 ms : [Simli](https://www.simli.com/) (vendor claims).+- ElevenLabs Agents: 90+ languages with mid-call switching, 16-17k voices, integrations (Zapier, Salesforce, Stripe, HubSpot, Shopify, Zendesk, Genesys, Twilio, Amazon Connect), MCP/API/webhooks, BYO LLM, free tier 15 min with commercial license; no avatars mentioned : [ElevenLabs](https://elevenlabs.io/agents).+- HeyGen interactive avatar: page now 301-redirects to https://www.liveavatar.com/ (not fetched) : [HeyGen redirect](https://www.heygen.com/interactive-avatar).+- Hume EVI: page returned 404; no verified data. Character.AI: not fetched.++### Inferences+- Missing in three.ws relative to closed products: a dedicated turn-taking model (Sparrow-like; open analogues are TEN Turn Detection and Pipecat smart-turn), multimodal perception of the user (camera/emotion), 90+ language voice breadth, CRM-style tool integrations.++### Gaps+- Hume EVI, Character.AI, HeyGen live-avatar details not retrieved.diff --git a/docs/research/oss-landscape-2026-10-notes/chat-memory-evals.md b/docs/research/oss-landscape-2026-10-notes/chat-memory-evals.mdnew file mode 100644index 000000000..e7c81b7aa--- /dev/null+++ b/docs/research/oss-landscape-2026-10-notes/chat-memory-evals.md@@ -0,0 +1,76 @@+# OSS repos that beat three.ws chat, memory, observability and companion surfaces (state as of 2026-10-09)++Method note: star counts and license lines were read from each repo's GitHub page on 2026-10-09 (WebFetch summaries of the repo front page; license text itself not re-read except where a license URL is given). Verdict key: ADOPT-CODE = copy/depend on it, license compatible with Apache-2.0; ADOPT-DESIGN = rebuild the idea; COVERED = three.ws already ships equivalent. three.ws baseline (from /home/user/projects/three.ws): `@three-ws/agent-memory` is a tiered (working/recall/archival) embeddings store with free-first NVIDIA NIM embeddings, deterministic entity graph, no published LoCoMo/LongMemEval score ([README](/home/user/projects/three.ws/packages/agent-memory/README.md)); docs exist for memory, companion, concierge, monitor, persona-hub under /home/user/projects/three.ws/docs.++## Chat UIs++### Takeaway+Open WebUI (154k) dominates on scale and enterprise features but its license now has a branding clause, so reuse is design-only. MIT/Apache chat UIs (LibreChat, AnythingLLM, Jan, assistant-ui, CopilotKit, AG-UI) are safe to depend on or copy. None ships a 3D-embodied agent in chat; that stays three.ws's moat. The gaps to close are generative UI / AG-UI protocol, resumable streams, artifacts, scheduled agent tasks, and dynamic model routing.++### Cited Findings+- **Open WebUI** https://github.com/open-webui/open-webui : ~154.1k stars (2026-10-09). License: "Open WebUI License", modified BSD-3 with branding clause; page points to LICENSE and LICENSE_HISTORY. Clause effective v0.6.6 (2025-04-19): may not alter/remove "Open WebUI" branding unless <=50 users in 30 days, substantive-contributor written permission, or enterprise license; code up to v0.6.5 stays BSD-3-Clause; CLA for later contributions; Nixpkgs treats it as non-free — [license docs](https://docs.openwebui.com/license/), [analysis](https://wz-it.com/en/blog/open-webui-license-branding-business/), [repo](https://github.com/open-webui/open-webui). Features: RBAC/groups, Filters/Actions/Pipes/Tools/Skills plugins, MCP/OpenAPI tool servers, multi-vector-DB RAG, parallel multi-model chat, voice/video calls, persistent memory, notes, channels, calendar + scheduled automations, Open Terminal agent, SCIM/SSO/LDAP, OpenTelemetry. Commercial reuse: an Apache-2.0 platform may not copy post-v0.6.5 code without carrying the branding obligation (and cannot rebrand a 50+ user deployment). Verdict: ADOPT-DESIGN (channels, scheduled automations, parallel multi-model compare, OTel export).+- **LibreChat** https://github.com/danny-avila/LibreChat : ~45.5k stars, MIT. Standout: MCP, reusable Skills and Subagents, sandboxed code interpreter, Artifacts (React/HTML/Mermaid), resumable streams synced across tabs/devices, OTel + Langfuse export, OAuth2/LDAP — [repo](https://github.com/danny-avila/LibreChat). Verdict: ADOPT-CODE for resumable-stream and artifacts patterns (MIT is Apache-compatible, keep notice); three.ws has a Claude.ai artifact embed but not in-chat code artifacts/interpreter.+- **LobeChat (LobeHub)** https://github.com/lobehub/lobe-chat : ~83.1k stars. License ambiguous on page: footer says "LobeHub Community License", badge says Apache 2.0; resolve against the LICENSE file before any reuse. Standout: Agent Builder from a description, Agent Groups with shared context, Operator (hire/schedule/report agents), IM Gateway to messaging apps, editable structured Personal Memory, 10,000+ tools/MCP plugins — [repo](https://github.com/lobehub/lobe-chat). Verdict: ADOPT-DESIGN (agent groups, editable memory view, IM gateway); do not copy code until license confirmed.+- **AnythingLLM** https://github.com/Mintplex-Labs/anything-llm : ~66.9k stars, MIT. Standout: dynamic model routing rules, memories, cron-scheduled agent tasks, skill selection cutting tokens up to 80%, no-code agent builder, embeddable website chat widget (Docker), citations on documents — [repo](https://github.com/Mintplex-Labs/anything-llm). Direct competitor to three.ws Concierge/Assistant widget. Verdict: ADOPT-DESIGN (model routing, source citations, token-saving skill selection); concierge otherwise COVERED, and three.ws adds a 3D body.+- **Jan** https://github.com/janhq/jan : ~44.9k stars, Apache-2.0. Local models from Hugging Face, OpenAI-compatible server at localhost:1337, MCP, custom assistants, fully offline — [repo](https://github.com/janhq/jan). Verdict: COVERED for chat; ADOPT-DESIGN only for offline/local-model lane (desktop companion app).+- **Cherry Studio** https://github.com/CherryHQ/cherry-studio : ~52.5k stars. AGPL-3.0 Community Edition; commercial license with AGPL exemption via [email protected]; Enterprise Edition separate. 300+ preset assistants, multi-model simultaneous chat, WebDAV backup, MCP — [repo](https://github.com/CherryHQ/cherry-studio). AGPL means no code copy into Apache-2.0. Verdict: ADOPT-DESIGN (assistant preset gallery; three.ws already has persona hub).+- **assistant-ui** https://github.com/assistant-ui/assistant-ui : ~12.5k stars, MIT (optional paid Assistant Cloud). Composable Thread/Message/Composer primitives, generative UI from tool calls, inline human approvals, accessibility, LangGraph + AG-UI/A2A adapters — [repo](https://github.com/assistant-ui/assistant-ui). Verdict: ADOPT-CODE as a dependency or reference for tool-call-to-component rendering and approvals.+- **CopilotKit** https://github.com/CopilotKit/CopilotKit : ~37.9k stars, MIT source. Generative UI, shared agent-UI state, human-in-the-loop, `useAgent` hook, React/Angular/Vue/RN/Slack/Teams; paid "Intelligence" platform adds threads, user memories, analytics (not OSS) — [repo](https://github.com/CopilotKit/CopilotKit). Verdict: ADOPT-DESIGN (shared state sync between agent and embodied 3D UI).+- **AG-UI protocol** https://github.com/ag-ui-protocol/ag-ui : ~16.4k stars, MIT. "Open, lightweight, event-based protocol" for agent-to-user-app, complements MCP (tools) and A2A (agent-to-agent). README claims adoption by Google, LangChain, AWS, Microsoft, Mastra, PydanticAI (self-reported) — [repo](https://github.com/ag-ui-protocol/ag-ui). Verdict: ADOPT-CODE (emit/consume AG-UI events from the agent runtime and embodiment layer; three.ws's Embodiment and Spatial MCP could be an AG-UI event extension). Highest-leverage standard to align with.++## Agent-memory systems++### Takeaway+Every serious entrant publishes >88% on LoCoMo/LongMemEval, but all numbers are vendor-reported on differing harnesses and the Mem0-vs-Zep dispute shows they are not comparable. Apache/MIT licensing means Graphiti, Cognee, Mem0, Supermemory and Memobase patterns can be adopted outright. three.ws's gaps: temporal fact invalidation, a published benchmark score, contradiction handling, user-profile synthesis, forget/improve verbs.++### Cited Findings+- **Mem0** https://github.com/mem0ai/mem0 : ~66.9k stars (2026-10-09), Apache-2.0. LoCoMo 92.5 and LongMemEval 94.4 with April 2026 algorithm (old: 71.4 / 67.8); BEAM 64.1 (1M) / 48.6 (10M); under ~7,000 tokens per retrieval. Scores are from the managed platform; OSS SDK "directionally similar but not identical" — [repo](https://github.com/mem0ai/mem0), [research](https://mem0.ai/research). An earlier page cited 91.6/93.4 — [Mem0 vs Zep](https://mem0.ai/blog/zep-vs-mem0-which-ai-memory-layer-should-you-choose). Verdict: ADOPT-DESIGN (single-pass hierarchical extraction, multi-signal retrieval) while keeping packages/agent-memory; ADOPT-CODE is possible (Apache-2.0).+- **Zep / Graphiti** https://github.com/getzep/graphiti : ~31.6k stars, Apache-2.0. Temporal knowledge graph: facts have validity windows and are invalidated not deleted, episode provenance, Pydantic ontologies, hybrid semantic+BM25+graph retrieval, Neo4j/FalkorDB/Neptune backends, MCP server, telemetry on by default (disable `GRAPHITI_TELEMETRY_ENABLED=false`) — [repo](https://github.com/getzep/graphiti). Zep self-reports 94.7% LoCoMo / 90.2% LongMemEval ([Zep](https://getzep.com/research)); Mem0 cites 80.3% / 71.2% for Zep ([Mem0](https://mem0.ai/blog/zep-vs-mem0-which-ai-memory-layer-should-you-choose)); contested. Verdict: ADOPT-DESIGN, top priority: temporal validity + provenance on top of three.ws's existing entity co-occurrence graph (three.ws graph is deterministic co-occurrence only).+- **Letta (MemGPT)** https://github.com/letta-ai/letta : ~25.1k stars, Apache-2.0. Pivoted to stateful agent harness (letta-code), TUI, desktop/web/mobile, Slack/Telegram/Discord, TS Agent SDK; V1 server archived — [repo](https://github.com/letta-ai/letta). 83.2% LoCoMo only in an unverified community table — [comparison](https://cdn.jsdelivr.net/gh/rohitg00/agentmemory@main/benchmark/COMPARISON.md). Verdict: ADOPT-DESIGN (self-editing memory blocks; three.ws has working/recall/archival tiers = COVERED in concept).+- **Cognee** https://github.com/topoteretes/cognee : ~31.9k stars, Apache-2.0. `remember`/`recall`/`improve`/`forget` verbs, local-only mode with GLiNER + local embeddings (no API key), Python/TS/Rust SDKs, MCP, migration tools from Mem0/Letta/Zep/Graphiti, graph viz UI — [repo](https://github.com/topoteretes/cognee). Verdict: ADOPT-DESIGN (improve/forget, import-from-competitor migration, local-no-key mode).+- **Supermemory** https://github.com/supermemoryai/supermemory : ~31.2k stars, MIT. Fact extraction with contradiction/expiry handling, static+recent user profile in one call, hybrid RAG+memory query, connectors (Drive, Gmail, Notion, OneDrive, GitHub), OCR/video, MCP, local mode. Self-reports #1 on LongMemEval, LoCoMo, ConvoMem; LongMemEval 95% Recall@15 at ~720 tokens — [repo](https://github.com/supermemoryai/supermemory). Not independently verified. Verdict: ADOPT-CODE/DESIGN for user-profile synthesis and connectors (three.ws already seeds memory from X, Farcaster, GitHub per docs).+- **Memobase** https://github.com/memodb-io/memobase : ~2.9k stars, Apache-2.0. "Memory for User, not Agent": configurable profile schema + event timeline, buffered batch flush, fixed 3 LLM calls per run, Go/Node/Python SDKs; claims LoCoMo SOTA without numbers on the page — [repo](https://github.com/memodb-io/memobase). Verdict: ADOPT-DESIGN (profile+timeline for the personal companion).+- **MemOS** https://github.com/MemTensor/MemOS : ~11.8k stars, Apache-2.0. Graph memory, multimodal (text/images/tool traces/personas), "memory cubes" with isolation/sharing, async MemScheduler, natural-language memory correction. Self-reported: LoCoMo 88.83, LongMemEval 89.20, PersonaMem v2 40.58, HaluMem 80.91, BEAM-10M 56.75 (OmniMemEval, no independent verification) — [repo](https://github.com/MemTensor/MemOS). Verdict: ADOPT-DESIGN (memory cubes map to per-agent/per-household memory sharing; NL correction UX).+- **2026 entrants (self-reported, single sources):** "Agent Zero Memory" paper claims 95.60% LongMemEval and 93.60% LoCoMo — [arXiv 2608.29606](https://arxiv.org/pdf/2608.29606); agentmemory (rohitg00) 95.2% LongMemEval-S self-measured — [comparison](https://cdn.jsdelivr.net/gh/rohitg00/agentmemory@main/benchmark/COMPARISON.md); Mastra cited at 94.87% LongMemEval in the arXiv paper; memU listed ~13.3k stars by a third-party directory (date unclear) — [SourcePulse](https://www.sourcepulse.org/projects/11539371). Treat as unverified.+- Benchmark caveat: LoCoMo is near-saturated and both benchmarks are contested; numbers from different harnesses/judges are not comparable — [search synthesis](https://mem0.ai/research).++## Observability and evaluation++### Takeaway+Langfuse, Opik, Helicone, DeepEval, Promptfoo, OpenLLMetry and Laminar are permissively licensed (MIT/Apache) and can be integrated or exported to. Phoenix is Elastic License 2.0: usable as a service dependency but not copyable code. three.ws's agent monitor, analytics, session-to-test recorder and feedback loop are product-embedded; competitors beat them on generic LLM tracing, LLM-as-judge metrics, datasets/experiments and prompt versioning.++### Cited Findings+- **Langfuse** https://github.com/langfuse/langfuse : ~35.6k stars; MIT except `ee` folders (not MIT; do not copy those) — [repo](https://github.com/langfuse/langfuse). Tracing, prompt management with caching, LLM-as-judge + manual labeling evals, datasets, playground jump-from-trace. Verdict: ADOPT-CODE as an optional exporter (LibreChat already ships Langfuse); ADOPT-DESIGN for prompt versioning and trace-to-playground in Agent Monitor.+- **Arize Phoenix** https://github.com/Arize-ai/phoenix : ~11.8k stars; **Elastic License 2.0 (ELv2)**, copyright Arize AI, some patent protected — [repo](https://github.com/Arize-ai/phoenix). OTel/OpenInference tracing, versioned datasets+experiments, PXI debugging agent, remote MCP server over traces. Commercial reuse: you may run it, but ELv2 forbids offering it as a hosted/managed service to third parties; cannot copy into an Apache-2.0 distribution. Verdict: ADOPT-DESIGN (MCP server over own traces; agent-that-debugs-traces).+- **Opik (Comet)** https://github.com/comet-ml/opik : ~22.5k stars, Apache-2.0 whole platform. Agent tracing, eval metrics (hallucination, moderation, RAG), Agent Optimizer SDK, guardrails, online eval rules, PyTest integration, MCP server — [repo](https://github.com/comet-ml/opik). Verdict: ADOPT-CODE (guardrail and online-eval rule concepts; Apache-compatible); three.ws recorder is session-to-test, they are CI-eval.+- **Helicone** https://github.com/Helicone/helicone : ~6.2k stars, Apache-2.0, appears actively developed on the page (no maintenance notice found; I did not check the project blog) — [repo](https://github.com/Helicone/helicone). AI gateway (100+ models, fallbacks), cost/latency analytics, PostHog export, public LLM cost API. Verdict: COVERED mostly (platform analytics, upstream resilience toolkit); ADOPT-DESIGN for public cost API.+- **Promptfoo** https://github.com/promptfoo/promptfoo : ~25.8k stars, MIT. README: "Promptfoo is now part of OpenAI" and remains MIT. Local evals, red teaming, side-by-side model comparison, CI code scanning — [repo](https://github.com/promptfoo/promptfoo). Verdict: ADOPT-CODE for CI red-team/eval of agent personas (MIT); watch vendor-neutrality risk after OpenAI acquisition.+- **DeepEval** https://github.com/confident-ai/deepeval : ~18.7k stars, Apache-2.0. Agent metrics (task completion, tool correctness, plan quality), multi-turn conversation and MCP metrics, synthetic datasets, pytest-style runs — [repo](https://github.com/confident-ai/deepeval). Verdict: ADOPT-CODE; strongest fit to extend the session-to-test recorder with multi-turn and tool-correctness scoring.+- **OpenLLMetry (Traceloop)** https://github.com/traceloop/openllmetry : ~7.5k stars, Apache-2.0. OTel instrumentations for LLM providers, vector DBs, frameworks, MCP; ~two dozen backends; telemetry removed in >=0.49.2 — [repo](https://github.com/traceloop/openllmetry). Verdict: ADOPT-CODE (emit standard OTel spans from the agent runtime so any backend works).+- **Laminar** https://github.com/lmnr-ai/lmnr : ~3.4k stars, Apache-2.0, Rust backend. Plain-English "Signals" that watch agent behavior and alert to Slack, SQL over traces via MCP/CLI, dashboards — [repo](https://github.com/lmnr-ai/lmnr). Verdict: ADOPT-DESIGN (natural-language behavior alerts for Agent Monitor fleet board).++## Character and companion platforms++### Takeaway+SillyTavern (AGPL) and RisuAI (GPL-3.0) are copyleft, so only their file formats and ideas can be used. The Character Card V2 spec is the de facto import format; V3 is a draft proposal (MIT repo). AIRI (MIT, ~50k stars) is the closest 3D/Live2D-embodied persistent-companion competitor and is code-adoptable; it still lists its memory ("Memory Alaya") as work in progress.++### Cited Findings+- **SillyTavern** https://github.com/SillyTavern/SillyTavern : ~34.3k stars, AGPL-3.0 (third-party parts MIT/Apache/OFL) — [repo](https://github.com/SillyTavern/SillyTavern). Lorebooks (WorldInfo), many backends, TTS, image gen via A1111/ComfyUI, extensions, visual novel mode. Commercial reuse: copying code forces AGPL on the combined work and network-use source disclosure; no. Verdict: ADOPT-DESIGN (lorebook/world-info keyword-triggered context injection; character card import for persona hub).+- **RisuAI** https://github.com/kwaroran/Risuai : ~1.7k stars, GPL-3.0 (as shown on page) — [repo](https://github.com/kwaroran/Risuai). Emotion images, group chats, plugins, regex scripts, lorebook, HypaMemory V2/V3 and SupaMemory context compression. Verdict: ADOPT-DESIGN (compression-based long-chat memory; emotion-mapped expression, which three.ws does with a 3D body so COVERED). Note: a different fork "RisuAI/RisuAI" has 2 stars; the original is kwaroran.+- **Character Card V2 spec** https://github.com/malfoyslastname/character-card-spec-v2 : fields `spec: "chara_card_v2"`, `data` with name, description, personality, scenario, first_mes, mes_example, creator_notes, system_prompt, post_history_instructions, alternate_greetings, character_book, tags, creator, character_version, extensions; page states no license — [spec](https://github.com/malfoyslastname/character-card-spec-v2). Verdict: ADOPT-CODE (import/export V2 cards in persona hub; format interop, license unstated so implement from the spec).+- **Character Card V3 spec** https://github.com/kwaroran/character-card-spec-v3 : draft "updated spec built on CCv2", MIT license listed, finalization pending ("Living Standard" goal); field/asset details are in SPEC_V3.md which I did not read — [repo](https://github.com/kwaroran/character-card-spec-v3). Verdict: ADOPT-CODE for import once read; supports embedded assets that could carry a GLB/VRM body.+- **AIRI (moeru-ai)** https://github.com/moeru-ai/airi : ~50.2k stars, MIT. VRM and Live2D avatars, auto blink/look-at/idle, browser/Discord audio input, multi-provider TTS (ElevenLabs, Azure, OpenAI-compatible, Kokoro), in-browser DuckDB WASM/pglite, web+macOS+Windows, self-described "self hosted, you-owned Grok Companion"; "Memory Alaya" WIP — [repo](https://github.com/moeru-ai/airi). Verdict: ADOPT-CODE for VRM/Live2D look-at and blink layers and client-side talking detection (check against three.ws procedural animation, partly COVERED); three.ws leads on memory, wallet, marketplace.+- Other 2026 open companions (unverified, from aggregator listings, star counts undated): ClawSoul (AGPL-3.0, Markdown file memory, daily schedule, selfie generation) — [PyPI](https://pypi.org/project/clawsoul-ai/); Digital Life Project 2 / DLP3D (MIT, real-time 3D expressive avatars) and ElizaOS (character framework with pgvector memory) — [listing](https://fast.io/resources/top-openclaw-tools-ai-companion-character-chat-apps/); ZcChat desktop pet; Nekro Agent.++## Inferences+- The ordered adoption list by leverage: (1) AG-UI events in the agent runtime, (2) Graphiti-style temporal fact invalidation + provenance in packages/agent-memory, (3) DeepEval/Promptfoo-style multi-turn and tool-correctness scoring in the session-to-test recorder, (4) OTel export via OpenLLMetry/Langfuse from Agent Monitor, (5) Character Card V2/V3 import and SillyTavern-style lorebooks in persona hub, (6) a published, reproducible LoCoMo/LongMemEval score for agent-memory.+- Licenses that block copying: AGPL (SillyTavern, Cherry Studio CE), GPL-3.0 (RisuAI), ELv2 (Phoenix), Langfuse `ee`, Open WebUI post-v0.6.5 (branding), LobeChat (unresolved).+- No surveyed chat UI or companion repo combines a skinned 3D body that is the chat surface with a wallet-backed agent economy; AIRI is the nearest.++## Gaps+- Exact current versions and release dates were not captured (GitHub pages did not render release data).+- LobeChat's actual LICENSE text (Community License vs Apache-2.0) not resolved.+- Character Card V3 field list (SPEC_V3.md) not read; Memobase exact LoCoMo numbers not found.+- Supermemory, MemOS, Zep, Letta benchmark numbers are all self-reported; no independent leaderboard found. Helicone maintenance status not checked beyond its README.+- Star counts come from WebFetch summarization of the repo page and may be rounded ("k"); no GitHub API cross-check.+- No hands-on audit of three.ws chat UI against each feature; three.ws baseline comes from the agent-memory README and doc filenames only.diff --git a/docs/research/oss-landscape-2026-10-notes/coding-mcp-skills.md b/docs/research/oss-landscape-2026-10-notes/coding-mcp-skills.mdnew file mode 100644index 000000000..8c050c170--- /dev/null+++ b/docs/research/oss-landscape-2026-10-notes/coding-mcp-skills.md@@ -0,0 +1,192 @@+# Open-source coding agents, computer-use/browser agents, MCP ecosystem, agent skills and sandboxes vs three.ws's developer surface (state as of 2026-10-09)++Method note for the writer: every star count below was read from the GitHub REST API (`gh api repos/<owner>/<repo>`, fields `stargazers_count`, `license.spdx_id`, `pushed_at`, `archived`) at 2026-10-09 05:16-05:18 UTC unless a different date is stated. "N/A" SPDX means GitHub could not classify the LICENSE file (custom, multi-license, or missing). Feature claims come from the official docs or repo README fetched the same day. three.ws's own surfaces were read from `docs/mcp.md`, `docs/cli.md`, `docs/agent-skills.md`, `docs/skills.md`, `docs/agent-screen-control.md` and the surfaces list supplied by the coordinator.++Verdict key: **adopt code** = the license permits vendoring/depending on it inside an Apache-2.0 platform; **adopt design, rebuild** = the idea is worth copying but the code is copyleft/source-available/proprietary or architecturally foreign; **already covered** = three.ws ships an equivalent.++## Key question 1: What developer-experience features define the best coding agents in 2026, and which can an agent platform adopt?++### Takeaway+The 2026 bar is set by four extensibility layers that every top agent now ships in some form: lifecycle hooks (Claude Code has 33 named events and five hook types; OpenCode exposes ~30 typed plugin events; Gemini CLI enabled hooks by default in Jan 2026), versioned plugin bundles with a marketplace manifest, subagents with their own tool allowlists and context, and a scriptable headless mode that emits JSON with per-run cost. Checkpoint/rewind, pattern-based permission rules (`Bash(git *)`), and OS-level sandboxes are the second tier. three.ws has none of these as a first-class product surface for its own agents (its in-app skills have a worker sandbox and a trust policy, but no hook events, no plugin manifest, no headless JSON runner), and all of the reference implementations except Claude Code itself are Apache-2.0 or MIT.++### Cited Findings++#### Fact sheet: coding agents (stars and licenses observed 2026-10-09 via GitHub API)++- **anthropics/claude-code**: 149,756 stars; LICENSE.md reads "© Anthropic PBC. All rights reserved. Use is subject to Anthropic's Commercial Terms of Service" (proprietary; the repo is issues, docs, plugins and examples, not the agent source). Source: [raw LICENSE.md](https://raw.githubusercontent.com/anthropics/claude-code/main/LICENSE.md); [repo](https://github.com/anthropics/claude-code). Open companions: **anthropics/claude-plugins-official** 37,563 stars, Apache-2.0 ([repo](https://github.com/anthropics/claude-plugins-official)); **anthropics/claude-agent-sdk-python** 8,229 stars, MIT ([repo](https://github.com/anthropics/claude-agent-sdk-python)); **anthropics/claude-code-action** 9,453 stars, MIT ([repo](https://github.com/anthropics/claude-code-action)); **anthropics/claude-code-security-review** 6,323 stars, MIT, last push 2026-02-11 ([repo](https://github.com/anthropics/claude-code-security-review)); **anthropics/claude-cookbooks** 53,293 stars, MIT ([repo](https://github.com/anthropics/claude-cookbooks)). Verdict: **adopt design, rebuild** for the agent itself; **adopt code** for the MIT/Apache SDK, action and plugin-marketplace format.+- **openai/codex**: 128,265 stars, Apache-2.0, pushed 2026-10-09. Source: [repo](https://github.com/openai/codex). Wikipedia notes the split "CLI: Apache 2.0, Desktop: Proprietary". Source: [Wikipedia](https://en.wikipedia.org/wiki/OpenAI_Codex_(AI_agent)). Verdict: **adopt code** (Rust CLI; sandbox and exec-mode design are the reusable parts).+- **google-gemini/gemini-cli**: 107,263 stars, Apache-2.0, pushed 2026-10-08. Source: [repo](https://github.com/google-gemini/gemini-cli). The official extension docs state "Gemini CLI was replaced by Antigravity CLI on June 18, 2026, for unpaid tier and Google One users". Source: [geminicli.com extensions](https://geminicli.com/docs/extensions/). Verdict: **adopt code** (extension manifest format).+- **anomalyco/opencode** (sst/opencode redirects to it): 212,239 stars, MIT, pushed 2026-10-09, "The open source coding agent." Source: [repo](https://github.com/anomalyco/opencode). The highest-starred coding agent in the set. Verdict: **adopt code** (plugin event API, permissions schema, `/share` session links).+- **cline/cline**: 70,040 stars, Apache-2.0, "Autonomous coding agent as an SDK, IDE extension, or CLI assistant." Source: [repo](https://github.com/cline/cline). Verdict: **adopt code**.+- **RooCodeInc/Roo-Code**: 24,285 stars, Apache-2.0, **archived**, last push 2026-05-15. Source: [repo](https://github.com/RooCodeInc/Roo-Code). Third-party FAQ says Roo's docs announced all products shut down 2026-05-15 and Kilo is the migration target. Source: [baeseokjae Cline alternatives 2026](https://baeseokjae.github.io/posts/cline-alternatives-2026/). Verdict: **already covered / skip** (dead project; mine its Orchestrator-mode design only).+- **Aider-AI/aider**: 49,428 stars, Apache-2.0, last push 2026-05-22 (no commits in ~5 months). Source: [repo](https://github.com/Aider-AI/aider). A 2026 roundup also flags "no commits since May 2026". Source: [pinggy best open source CLI coding agents](https://pinggy.io/blog/best_open_source_cli_coding_agents/). Verdict: **adopt design, rebuild** (git-native "every edit is a reviewable commit" is still the best UX idea here; code is stale).+- **block/goose**: 55,093 stars, Apache-2.0, pushed 2026-10-09; extensibility is MCP-based "extensions"; a 2026 comparison says it "lives under the Linux Foundation and handles more than code". Source: [repo](https://github.com/block/goose); [morphllm OpenCode vs Goose](https://www.morphllm.com/comparisons/opencode-vs-goose). Verdict: **adopt code** (Rust, Apache; MCP-extension model matches three.ws's MCP-first stack).+- **OpenHands/OpenHands** (All-Hands-AI redirects): 90,319 stars, MIT, pushed 2026-10-09. Source: [repo](https://github.com/OpenHands/OpenHands). Headless mode: `openhands --headless -t "task"`, `-f task.txt`, `--json` streams JSONL events, and it runs in `always-approve` mode which "cannot be changed". Source: [OpenHands headless docs](https://docs.openhands.dev/openhands/usage/run-openhands/headless-mode). Verdict: **adopt code** (MIT; the sandboxed headless runner is the closest OSS analogue to an "agent that does work in CI").+- **Kilo-Org/kilocode**: 27,540 stars, MIT, pushed 2026-10-08, "all-in-one agentic engineering platform". Source: [repo](https://github.com/Kilo-Org/kilocode). A 2026 roundup says Kilo "was rebuilt on top of OpenCode in April 2026, and it spans VS Code, JetBrains, and CLI". Source: [pinggy](https://pinggy.io/blog/best_open_source_cli_coding_agents/). Verdict: **adopt code** (MIT).+- **openinterpreter/open-interpreter**: 68,534 stars, Apache-2.0, pushed 2026-10-07, now described as "A coding agent for open models like Kimi K3 and GLM 5.3". Source: [repo](https://github.com/openinterpreter/open-interpreter). Verdict: **adopt code** if an open-model lane is wanted.+- **openclaw/openclaw** (2026 entrant, personal agent rather than coding agent): 391,499 stars, MIT, pushed 2026-10-09, "The AI that really does things. Any OS. Any Platform." Channels: WhatsApp, Telegram, Discord, Slack, Teams, iMessage, Google Chat, Signal "plus 20+ more"; native apps for macOS, iOS, Android, Windows, Linux; model providers and agent harnesses are swappable plugins shared through ClawHub; unknown senders are paired by default and approved with a CLI command; tools run on the host unless sandboxing is configured. Source: [repo](https://github.com/openclaw/openclaw). Verdict: **adopt code** (MIT) for the channel-gateway and pairing model, which overlaps three.ws's Herald/Knock/companion surfaces.++#### Feature-level findings (Claude Code, official docs)++- Plugin = directory with manifest at `.claude-plugin/plugin.json`; components are Skills (`SKILL.md`), Agents (subagent definitions), Hooks, a hooks module ("mod") written as JavaScript that can draw panes, and MCP servers. A marketplace is "a repository or directory with a `.claude-plugin/marketplace.json` file", "a catalog, not a hosted store". Install scopes: user, project (committed `.claude/settings.json`), local. Official marketplace plugins show a "Context cost" estimate before install; `claude plugin disable` exists; three marketplace tiers (official, community, third-party) with names reserved for `github.com/anthropics/` sources. Source: [Plugins overview](https://code.claude.com/docs/en/plugins).+- Hook events (33): SessionStart, Setup, UserPromptSubmit, UserPromptExpansion, PreToolUse, PermissionRequest, PermissionDenied, PostToolUse, PostToolUseFailure, PostToolBatch, Notification, MessageDisplay, SubagentStart, SubagentStop, TaskCreated, TaskCompleted, Stop, StopFailure, TeammateIdle, InstructionsLoaded, ConfigChange, CwdChanged, DirectoryAdded, FileChanged, WorktreeCreate, WorktreeRemove, PreCompact, PostCompact, PreModelSwitch, PostModelSwitch, Elicitation, ElicitationResult, SessionEnd. Hook types: `command`, `http`, `mcp_tool`, `prompt`, `agent`. Returns include `permissionDecision: allow|deny|ask|defer`, `updatedInput` (rewrite tool args), `updatedToolOutput`, `additionalContext`, exit code 2 = block. Source: [Hooks reference](https://code.claude.com/docs/en/hooks).+- Headless: `claude -p`, `--output-format text|json|stream-json`, `--json-schema` for structured output, `--continue`, `--resume <session-id or .jsonl path>`, `--allowedTools "Bash(git diff *)"` prefix rules, `--permission-mode auto|dontAsk|acceptEdits`, `--permission-prompts none`, `--max-budget-usd`, `--bare` (skips hooks/skills/plugins/MCP/CLAUDE.md discovery for reproducible CI), `--mcp-config`, `--plugin-dir`; JSON output "includes `total_cost_usd` and a per-model cost breakdown"; `system/init` event lists `plugins`, `plugin_errors`, `mcp_servers`, `mcp_server_errors` so CI can fail when a plugin or server did not load; subagent messages carry `parent_tool_use_id`. Source: [Run Claude Code programmatically](https://code.claude.com/docs/en/headless).+- Checkpoints: every prompt that starts a turn creates a checkpoint; snapshots kept for the 100 most recent checkpoints and saved with the conversation so `/rewind` works after resume; `/rewind` or Esc Esc offers "Restore code and conversation", "Restore conversation", "Restore code", "Summarize from here", "Summarize up to here"; Bash-made changes, most subagent edits, external edits, and symlinked/hard-linked paths are not restored. Source: [Checkpointing](https://code.claude.com/docs/en/checkpointing).+- Subagents run in their own context, can run in a git worktree, return only a final message, and respect the parent's settings.json permission tree; hooks can be attached to subagents and skills through frontmatter and a Stop hook becomes SubagentStop. Source: [alexop.dev Claude Code explained 2026](https://alexop.dev/posts/understanding-claude-code-full-stack/); [Hooks reference](https://code.claude.com/docs/en/hooks).++#### Feature-level findings (other agents)++- OpenCode plugins: JS/TS files in `.opencode/plugins/` or npm packages listed in the `plugin` array of `opencode.json` (Bun installs them at startup); a plugin exports a function receiving `project`, `directory`, `worktree`, `client`, `$` and returns hooks; events include `tool.execute.before/after`, `shell.env`, `experimental.session.compacting`, `permission.asked`, `permission.replied`, `session.idle`, `session.compacted`, `file.edited`, `lsp.client.diagnostics`, `tui.toast.show`; a `tool` helper with Zod args defines custom tools. Source: [OpenCode plugins](https://opencode.ai/docs/plugins/).+- OpenCode permissions: keys `read`, `edit`, `glob`, `grep`, `bash`, `task` (subagent launches), `skill`, `lsp`, `question`, `webfetch`, `websearch`, `external_directory`, `doom_loop` (same call 3x); values `allow|ask|deny`; per-command patterns `"bash": {"*":"ask","git *":"allow","rm *":"deny"}`; per-agent overrides in agent Markdown frontmatter; `*.env` denied by default; `opencode run --auto`. Source: [OpenCode permissions](https://opencode.ai/docs/permissions/). Also `/undo`, `/redo`, `/share` (public conversation link), `/init` writes `AGENTS.md`, Plan/Build modes on Tab, desktop app and IDE extension, Docker image `ghcr.io/anomalyco/opencode`. Source: [OpenCode docs](https://opencode.ai/docs/).+- Codex CLI: `codex`, `/init` (AGENTS.md), `/permissions` (shows active sandbox and writable roots), `/review` (prioritised findings without modifying the tree), `codex resume`, `codex --image`, `codex --search`, `codex exec` for pipelines, `codex mcp` (adds local or remote MCP servers and authenticates), `codex cloud` (submit work to a configured environment and apply results locally), skills, plugins, hooks (link only) and subagents. Source: [Codex CLI docs](https://learn.chatgpt.com/docs/codex/cli).+- Gemini CLI extensions: `gemini-extension.json` fields `name`, `version`, `description`, `migratedTo`, `mcpServers`, `contextFileName` (defaults to `GEMINI.md`), `excludeTools` (supports `run_shell_command(rm -rf)`), `plan.directory`, `settings` (sensitive values go to the keychain), `themes`; directory conventions `hooks/hooks.json`, `skills/<name>/SKILL.md`, `agents/*.md` (preview), `commands/*.toml` -> `/gcs:sync`, `policies/*.toml`; install from GitHub URL or local path, `--ref` pins a branch/tag/commit, `gemini extensions update --all`, `--auto-update`, gallery at geminicli.com/extensions. Source: [Gemini CLI extension reference](https://geminicli.com/docs/extensions/reference/). Hooks "officially enabled by default" in v0.26.0 (2026-01-28). Source: [gemini-cli discussion #17812](https://github.com/google-gemini/gemini-cli/discussions/17812).+- Benchmarks quoted by third parties: "Codex CLI 83.4% on Terminal-Bench, OpenCode Free at 172K Stars" (page title, June-era snapshot). Source: [morphllm best AI CLI tools 2026](https://www.morphllm.com/best-ai-cli-tools-2026).++#### How three.ws compares (from its own docs)++- three.ws's external-facing skills are plain SKILL.md folders in `.agents/skills/`, shipped as a Claude Code plugin marketplace (`/plugin marketplace add nirholas/three.ws`, `/plugin install three-ws-core@three-ws`) and symlinked into `.claude/skills/` on `npm install`. Source: `docs/agent-skills.md`.+- three.ws's in-app skills are a 4-file bundle (`manifest.json`, `SKILL.md`, `tools.json`, `handlers.js`) run inside a worker sandbox with `sandboxPolicy: "trusted-main-thread"` only honoured for same-origin bundles, an `author` field enforced under `owned-only` trust, and semver `version` in the manifest. Source: `docs/skills.md`.+- The `three-ws` CLI writes hosted MCP servers into Claude Code, Claude Desktop, Cursor, Windsurf, VS Code, Codex, Gemini CLI and Hermes, with `--oauth`, `--device`, `--key`, `--financial` sign-in flags and a live `tools/list` verification. Source: `docs/cli.md`.++### Inferences+- Every feature the owner listed (hooks, plugins, subagents, checkpoints, cost tracking, headless CI) has at least one permissively licensed reference implementation: OpenCode (MIT) for plugins/permissions, OpenHands (MIT) for headless JSONL, Codex (Apache) for sandbox + `exec`, Gemini CLI (Apache) for the extension manifest. Nothing here requires copying Claude Code's proprietary source.+- The highest-leverage adoption for three.ws is a hook/event bus on the agent runtime (`@three-ws/agent-runtime`) with OpenCode's event naming, plus a `three-ws run -p ... --output-format json` headless mode that reports `total_cost_usd`-style spend in $THREE/USDC, since three.ws already meters x402 spend per call.+- three.ws's in-app skill bundle is richer than the SKILL.md spec (it has typed tools and handlers) but lacks the one thing every 2026 host has: a marketplace manifest with versioned install and a "context cost" preview.++### Gaps+- Codex's sandboxing and hooks pages (`/codex/sandboxing`, `/codex/hooks`) returned 404 on 2026-10-09, so the exact sandbox mode names and hook events for Codex are unverified here.+- Cline's hook/subagent surface was not fetched from primary docs; only the repo description (SDK, IDE extension, CLI) is confirmed.+- No neutral 2026 benchmark comparing these agents on the same task set was found; the Terminal-Bench figure comes from a vendor-adjacent blog.++## Key question 2: How do the MCP registries present servers, and what would make three.ws's MCP servers rank and install well?++### Takeaway+Registries now key on four signals: verified namespace ownership (official registry: GitHub OAuth/OIDC for `io.github.*`, DNS or HTTP challenge for reverse-DNS names), live usage counts (Smithery shows "uses", Anthropic's directory shows installs, runs per tool and error rates), letter-graded metadata (Glama grades license, quality and maintenance and lets users filter 40,169 remote-capable servers), and remote Streamable HTTP with OAuth and dynamic client registration (Anthropic's directory accepts a remote server by URL only and no longer accepts MCPB desktop extensions). three.ws already has Streamable HTTP, OAuth 2.1 with RFC 7591 DCR, and `/.well-known/mcp.json`, so the missing work is publishing to the official registry under a DNS-verified `ws.three.*` namespace and submitting the hosted servers as connectors to Anthropic's directory.++### Cited Findings++#### Fact sheet: MCP ecosystem (stars and licenses observed 2026-10-09)++- **modelcontextprotocol/servers**: 91,085 stars; LICENSE file states "The MCP project is undergoing a licensing transition from the MIT License to the Apache License, Version 2.0 ... All new code and specification contributions ... are licensed under Apache-2.0. Documentation contributions (excluding specifications) are licensed under CC-BY-4.0." Source: [raw LICENSE](https://raw.githubusercontent.com/modelcontextprotocol/servers/main/LICENSE); [repo](https://github.com/modelcontextprotocol/servers). Verdict: **adopt code**.+- **modelcontextprotocol/registry**: 7,331 stars; same MIT-to-Apache-2.0 transition LICENSE. Launched in preview 2025-09-08; "breaking changes or data resets may occur" until GA; v0.1 API froze 2025-10-24. Publishing via `mcp-publisher` CLI; `io.github.<user>/...` names require GitHub login or a GitHub Action on that user's repos; reverse-DNS names like `me.<domain>/...` require a DNS or HTTP challenge; auth methods are GitHub OAuth, GitHub OIDC, DNS verification, HTTP verification; API docs at registry.modelcontextprotocol.io/docs; images at ghcr.io/modelcontextprotocol/registry; working group lead Radoslav Dimitrov (Stacklok) with Tadas Antanavicius (PulseMCP), Bob Dickinson, Preeti Dewani. Source: [registry README](https://github.com/modelcontextprotocol/registry); [raw LICENSE](https://raw.githubusercontent.com/modelcontextprotocol/registry/main/LICENSE). A 2026 comparison says it is "now under the Linux Foundation's Agentic AI Foundation (Dec 9, 2025), but remains in preview". Source: [designrevision best MCP marketplaces](https://designrevision.com/blog/best-mcp-marketplaces-and-registries). Verdict: **adopt code** (self-host a subregistry) and **publish to it**.+- **modelcontextprotocol/modelcontextprotocol** (spec): 9,411 stars, same transitional license. Source: [repo](https://github.com/modelcontextprotocol/modelcontextprotocol). **modelcontextprotocol/inspector**: 11,044 stars. Source: [repo](https://github.com/modelcontextprotocol/inspector).+- **Smithery**: "Browse 25,717+ MCPs"; "now a part of Arcade.dev!"; listings show a usage count ("43.39k uses" for Agent News, 22.49k OneSignal, 9.18k Exa Search, 5.56k Context7) and every featured listing has a verified badge; remote Streamable HTTP listings; "handles OAuth flows, credential injection, and retries automatically"; `npx smithery auth login`; built on open-source agent.pw. Source: [smithery.ai](https://smithery.ai/). **smithery-ai/cli**: 840 stars, **AGPL-3.0**, last push 2026-05-31. Source: [repo](https://github.com/smithery-ai/cli). Verdict: **adopt design, rebuild** (AGPL CLI; list there, do not vendor).+- **Glama**: directory lists 97,619 servers; filters: remote-capable 40,169, local-only 35,284, hybrid 17,336, Claimed 6,820; each listing shows three letter grades labelled license, quality, maintenance (e.g. A, A, B) and an "official" badge; sorts by relevance, recent usage, date added/updated, weekly npm/PyPI downloads, GitHub stars; license shown per listing. Source: [glama.ai/mcp/servers](https://glama.ai/mcp/servers). Verdict: **list there** (no code to adopt; glama-ai/glama repo was not found via API).+- **PulseMCP**: "directory plus a newsletter run by two people, one of whom sits on the MCP Steering Committee and maintains the official MCP Registry"; "tracks 16,820+ servers". Source: [designrevision](https://designrevision.com/blog/best-mcp-marketplaces-and-registries). Verdict: **list there**.+- **mcp.run (Dylibso)**: `https://www.mcp.run/` now 301-redirects to `https://turbomcp.ai/`, a self-hosted enterprise MCP gateway page published 2026-01-08 that does not mention mcp.run, servlets or WebAssembly. Source: [turbomcp.ai](https://turbomcp.ai/). The dylibso/mcp.run repo returned no JSON from the API. Verdict: **skip** (product appears to have pivoted; status unconfirmed).+- **jlowin/fastmcp** (Python): 28,018 stars, Apache-2.0. Source: [repo](https://github.com/jlowin/fastmcp). **punkpeye/fastmcp** (TypeScript): 3,274 stars, MIT. Source: [repo](https://github.com/punkpeye/fastmcp). Verdict: **adopt code** for any new Python worker MCP; three.ws's 42 npm servers already use the TS SDK.+- **mcp-use/mcp-use**: 10,733 stars, MIT, "The fullstack MCP framework to develop MCP Apps for ChatGPT / Claude & MCP Servers". Source: [repo](https://github.com/mcp-use/mcp-use). Verdict: **adopt code** (MCP Apps / UI-in-chat is adjacent to three.ws's Spatial MCP and Claude.ai artifact embed).+- **Anthropic directory (claude.ai/directory)**: anyone on Pro/Max/Team/Enterprise can submit; two listing kinds, Plugin bundle (public GitHub repo) and MCP connector (remote server URL, "no repository needed"); "If you have a remote MCP server, always submit it as an MCP connector"; connectors are scanned automatically and "listed as a Community connector" by default, with escalation to "Verified review" where a reviewer tests each tool; MCPB desktop extensions are deprecated and "the directory no longer accepts local MCP servers packaged as desktop extensions"; a published plugin's Usage tab "shows installs, versions, how often each skill and MCP server runs, and error rates"; connector dashboard shows "server health and usage by tool"; updates flow from merges to the tracked branch. Source: [Publish to the directory](https://claude.com/docs/directory/publish).+- Gateways and scanners: **agentgateway/agentgateway** 5,247 stars, Apache-2.0 ([repo](https://github.com/agentgateway/agentgateway)); **IBM/mcp-context-forge** 4,588 stars, Apache-2.0, "AI Gateway, registry, and proxy ... in front of any MCP, A2A, or REST/gRPC APIs" ([repo](https://github.com/IBM/mcp-context-forge)); **docker/mcp-gateway** 1,590 stars, MIT ([repo](https://github.com/docker/mcp-gateway)), "one container per server with resource limits" per a comparison ([ailearningguides](https://ailearningguides.com/mcp-scan-vs-docker-mcp-gateway-2026/)); **microsoft/mcp-gateway** 867 stars, MIT ([repo](https://github.com/microsoft/mcp-gateway)); **lasso-security/mcp-gateway** 395 stars, MIT ([repo](https://github.com/lasso-security/mcp-gateway)); **invariantlabs-ai/mcp-scan** 3,128 stars, Apache-2.0, now published under the `snyk` org as `uvx snyk-agent-scan@latest`, scans MCP servers AND skills (`SKILL.md` single-file scans, `--no-skills` to skip), risks: prompt injection, tool poisoning, tool shadowing, toxic flows, suspicious downloads, credential handling; requires `SNYK_TOKEN`; "does not accept external contributions at this time" ([repo](https://github.com/invariantlabs-ai/mcp-scan)). An October 2026 test found "Two working malicious MCP servers ... got clean results from each of five scanners, including Cisco mcp-scanner, Snyk agent-scan and NVIDIA SkillSpector". Source: [adversa.ai MCP security October 2026](https://adversa.ai/blog/top-mcp-security-resources-october-2026/). Verdicts: agentgateway / context-forge **adopt code** (Apache) if three.ws wants a self-hosted gateway in front of its 7 hosted servers; mcp-scan **adopt design, rebuild** (Apache but Snyk-token-gated and closed to contributions); three.ws's `@three-ws/mcp-policy` and `docs/mcp-safety.md` are **already covered** for policy, not for scanning.++#### How three.ws compares (from its own docs)++- three.ws serves MCP over Streamable HTTP at `https://three.ws/api/mcp`, protocol `2025-06-18`, auth "OAuth 2.1 (end-user) or API key", with DCR at `POST /oauth/register`, PKCE S256, and `GET /.well-known/oauth-authorization-server`; seven hosted remote servers plus 42 `@three-ws/*` npm stdio servers; a `/.well-known/mcp.json` directory consumed by the CLI; MCP client analytics ("which AI clients use our MCP servers") exists as a surface. Source: `docs/mcp.md`, `docs/cli.md`, surfaces list.++### Inferences+- three.ws already meets the two hard technical requirements for a Verified Anthropic connector (remote Streamable HTTP + OAuth with DCR). What ranks well on Smithery/Glama is usage and claimed/official badges, so the lever is distribution: publish each hosted server to the official registry under a DNS-verified `ws.three.*` namespace (the registry supports an HTTP/DNS challenge, which three.ws controls), claim the Glama listings, submit the 7 hosted servers as connectors to the Anthropic directory and bundle the `.agents/skills` pack as the paired plugin.+- Glama grades license, quality and maintenance per listing; three.ws's 42 npm servers each need a visible license, README and recent publish to score A/A/A.+- A self-hosted Apache gateway (agentgateway or IBM context-forge) could front the hosted servers and give the MCP client analytics surface per-tool auth, rate limits and audit without new code.++### Gaps+- The official registry's web UI showed "Loading servers..." and no counts; the registry's own server count and whether it exposes install/usage metrics were not confirmed.+- Smithery's ranking formula and the time window of its "uses" counter are not published on the page.+- The Glama scoring rubric behind the letter grades is not documented on the listing page.+- Whether three.ws has already published to the official registry or claimed Glama/Smithery listings was not checked in the repo.++## Key question 3: What does the Agent Skills ecosystem look like in 2026, and how does three.ws's skill format compare?++### Takeaway+The Agent Skills spec (agentskills.io, Apache-2.0 reference repo) is deliberately minimal: `name` and `description` required, optional `license`, `compatibility`, `metadata` (string map, where `version` lives), and experimental `allowed-tools`; there is no signing, no registry and no version resolution in the spec. Distribution has consolidated around Vercel's `npx skills` CLI and skills.sh leaderboard (1.25M lifetime installs tracked, top skill `find-skills` at 3.8M installs, ~79 target agents), Anthropic's plugin marketplaces, and ClawHub for OpenClaw (which does claim "signed manifests" and "moderated releases"). three.ws's external pack is spec-compliant and plugin-packaged; its in-app bundle is strictly richer (typed `tools.json`, sandboxed `handlers.js`, semver, trust policy) but is a private format no registry indexes.++### Cited Findings++#### The spec++- Required frontmatter: `name` (1-64 chars, lowercase alphanumerics and hyphens, must match the directory name) and `description` (1-1024 chars). Optional: `license`, `compatibility` (max 500 chars), `metadata` (string-to-string map; example `metadata: {author: example-org, version: "1.0"}`), `allowed-tools` (space-separated, "Experimental"). Progressive disclosure: metadata ~100 tokens at startup, body <5000 tokens recommended, resources on demand; keep SKILL.md under 500 lines; validate with `skills-ref validate ./my-skill`. Source: [agentskills.io specification](https://agentskills.io/specification).+- **agentskills/agentskills** (spec repo): 25,981 stars, Apache-2.0, last push 2026-08-09. Source: [repo](https://github.com/agentskills/agentskills).+- **anthropics/skills**: 180,029 stars; GitHub cannot classify a single license; README says "Many skills are Apache 2.0 open source" while `skills/docx`, `skills/pdf`, `skills/pptx`, `skills/xlsx` "are source-available, not open source"; each skill folder carries its own `LICENSE.txt` (confirmed for `skills/pdf`); repo holds `skills/`, `spec/`, `template/`, `.claude-plugin/`; install with `/plugin marketplace add anthropics/skills` then `/plugin install document-skills@anthropic-agent-skills`. Source: [repo](https://github.com/anthropics/skills); [API contents listing](https://api.github.com/repos/anthropics/skills/contents/skills/pdf). Verdict: **adopt code** for Apache-2.0 skills only; the four document skills are **adopt design, rebuild**.++#### Registries and install counts++- **skills.sh (Vercel)**: leaderboard columns rank, skill, source repo, "8W Activity", installs; All Time total 1,246,524; top 10: find-skills (vercel-labs/skills) 3.8M, grill-me (mattpocock/skills) 1.3M, grill-with-docs 1.1M, improve-codebase-architecture 1.1M, agent-browser (vercel-labs/agent-browser) 1.1M, tdd 1.0M, setup-matt-pocock-skills 967.8K, frontend-design (anthropics/skills) 965.8K, handoff 947.0K, triage 911.8K; 20 agents listed on the page; sections `/audits` ("Security audits") and `/official`; no verification or signing process described. Source: [skills.sh](https://skills.sh/). (Note the per-skill numbers exceed the stated "All Time" total; the page does not reconcile the two.)+- **vercel-labs/skills**: 33,475 stars, MIT; commands `npx skills add|find|update|remove|use|list|init`; sources: GitHub shorthand, subpaths, GitLab, Azure Repos, any git URL, local paths, direct `SKILL.md`/archive URLs; supports "OpenCode, Claude Code, Codex, Cursor, and 75 more"; install counts come from CLI telemetry ("anonymous usage data", repo identifiers sent only for public GitHub repos, opt out with `DISABLE_TELEMETRY=1` or `DO_NOT_TRACK=1`); no lockfile format documented. Source: [repo](https://github.com/vercel-labs/skills). Verdict: **adopt code** (MIT) and **publish to it**: three.ws's pack is one `npx skills add nirholas/three.ws` away from appearing on the leaderboard.+- **mattpocock/skills**: 281,398 stars, MIT, "Skills for Real Engineers". Source: [repo](https://github.com/mattpocock/skills). **vercel-labs/agent-browser**: 43,700 stars, Apache-2.0, "Browser automation CLI for AI agents". Source: [repo](https://github.com/vercel-labs/agent-browser).+- **ComposioHQ/awesome-claude-skills**: 76,721 stars, no license detected, last push 2026-09-18. Source: [repo](https://github.com/ComposioHQ/awesome-claude-skills). **heilcheng/awesome-agent-skills**: 6,265 stars, MIT. Source: [repo](https://github.com/heilcheng/awesome-agent-skills). **skillmatic-ai/awesome-agent-skills**: 682 stars, CC0-1.0. Source: [repo](https://github.com/skillmatic-ai/awesome-agent-skills). Verdict: **submit PRs** listing three.ws's 3d/creative skills.+- **ClawHub (openclaw/clawhub)**: 9,498 stars, MIT, "Skill + Plugin Registry for OpenClaw"; site shows 30 skills and 12 plugins with download counts (Memory LanceDB 15.6k, DuckDuckGo 13.2k, Lobster 11k); publishing via `npm i -g clawhub`, `clawhub login`, `clawhub skill publish ./my-skill --slug my-skill --version 1.0.0`; the page lists "signed manifests", "moderated releases", publisher handles, org trust and an Audits section. Source: [clawhub.ai](https://clawhub.ai/); [repo](https://github.com/openclaw/clawhub). A third-party guide says OpenClaw "ships 53 bundled skills out of the box" and recommends checking "the VirusTotal scan on each skill's ClawHub page". Source: [claw.mobile OpenClaw skills list](https://claw.mobile/blog/openclaw-skills-list); [firecrawl OpenClaw skills](https://www.firecrawl.dev/blog/openclaw-skills). Verdict: **adopt code** (MIT registry with signed manifests and versioned publish is the closest OSS match to three.ws's community-skills registry).+- Security: Snyk's agent-scan (mcp-scan lineage) scans `SKILL.md` for prompt injection, suspicious downloads, malicious code, credential handling and secrets. Source: [mcp-scan repo](https://github.com/invariantlabs-ai/mcp-scan).++#### Tool catalogs++- **ComposioHQ/composio**: 30,476 stars, MIT, "1000+ toolkits, tool search, context management, authentication, and a sandboxed workbench"; site claims "1,500+ apps", managed OAuth, tool search that "loads only what your task needs", isolated sandbox per execution, free Hobby plan "100,000 tool calls per month". Source: [repo](https://github.com/ComposioHQ/composio); [composio.dev](https://composio.dev/). Verdict: **adopt code** (MIT SDK) if three.ws wants third-party SaaS tools inside agents; **already covered** for the marketplace concept (three.ws sells agent skills per call over x402, which Composio does not).+- **ArcadeAI/arcade-mcp** (arcade-ai redirects): 1,047 stars, MIT, "MCP Server Framework and Tool Development library". Source: [repo](https://github.com/ArcadeAI/arcade-mcp). Arcade's own blog claims "8,000+ agent-optimized MCP tools" and per-user OAuth scoping tied to enterprise IdPs; Smithery is now part of Arcade. Source: [arcade.dev blog](https://www.arcade.dev/blog/composio-alternatives/); [smithery.ai](https://smithery.ai/). Verdict: **adopt design, rebuild** (per-user OAuth scoping for MCP tools is the feature; three.ws's `--financial` scope split in the CLI is a start).++#### How three.ws compares (from its own docs)++- External pack: SKILL.md folders under `.agents/skills/` with generated `SKILLS.md` and `skills-pack.json` carrying `category`, `origin`, `crossPlatformSafe`; entry-point `https://three.ws/skill.md` and a Grok variant; shipped via the repo's `.claude-plugin/marketplace.json`. Source: `docs/agent-skills.md`.+- In-app bundle: `manifest.json` (semver `version`, `author`, `sandboxPolicy`), `SKILL.md`, `tools.json` (JSON Schema tools), `handlers.js` (worker-sandboxed ES module); dependencies by URI with semver ranges (`ipfs://...`, `https://...`); `owned-only` trust mode. Source: `docs/skills.md`. Surfaces list also names "Skill bundles (seller builder + evidence pricing)", "On-chain skill licenses", "`@three-ws/skill-license`", "Standalone skill repos", "Claude Code plugin marketplace".++### Inferences+- The spec leaves versioning to `metadata.version` and signing to nobody; ClawHub is the only registry in this set advertising signed manifests, and three.ws's on-chain skill licenses plus `@three-ws/skill-license` are a stronger provenance story than any registry here. Exposing that (a signed, versioned manifest with an on-chain license pointer) in the public pack would be a differentiator rather than a gap.+- three.ws's pack is invisible on the two leaderboards that matter (skills.sh counts CLI telemetry; Anthropic's directory counts installs per version) until it is published to both. `npx skills add` requires nothing beyond a public GitHub path, so this is a distribution task, not an engineering one.+- The in-app bundle format (typed tools + handlers) is what Claude Code plugins achieve by bundling an MCP server with a skill. Exporting an in-app skill as a Claude Code plugin (SKILL.md + `.mcp.json` pointing at the hosted server) would let one authored skill reach both ecosystems.++### Gaps+- No primary source for how skills.sh computes "8W Activity" or why per-skill install counts exceed the all-time total shown.+- ClawHub's VirusTotal integration is reported by third-party blogs only; the clawhub.ai page did not mention it.+- Whether Anthropic's `/official` section on skills.sh or the directory's "Verified" label is attainable without an Anthropic partnership was not stated on either page.++## Key question 4: Which open computer-use stacks reach the best benchmark scores, what architecture do they use, and how does that compare with three.ws's agent-screen-control?++### Takeaway+Desktop computer use has crossed the human baseline: OSWorld's paper puts humans at 72.36%, Agent S3 (Apache-2.0, screenshot + grounding model + reflection + code agent + Behavior Best-of-N) reports 72.6% on OSWorld at 100 steps, UI-TARS-2 reports 47.5% in its own paper, and third-party aggregators list Qwen3.8 Max at 86.1% and open-weight Qwen3.8-27B at 84.3% on OSWorld-Verified as of late September 2026. Browser agents split by perception: Browser Use (MIT, 117k stars) and Agent TARS use DOM/hybrid, Skyvern (AGPL) is vision-first on Playwright, Stagehand v4 (MIT) uses hybrid accessibility-tree trimming with `act/extract/observe` primitives, and Cua (MIT core) runs full macOS/Linux VMs. three.ws's agent-screen-control is a human-in-the-loop remote-control channel over a throwaway Chromium (lease tokens, SSRF guard, owner-only), not an autonomous perception stack, so the comparison is architectural: it is already covered on safety, and behind on autonomy.++### Cited Findings++#### Benchmarks++- OSWorld paper: "humans can accomplish over 72.36% of the tasks" and "the best model achieves only 12.24% success" at publication; OSWorld-Verified introduced 2025-07-28; OSWorld 2.0 announced 2026-06-26; the live leaderboard table did not render (showed "Loading verified benchmark data..."). Source: [osworld-v1.xlang.ai](http://osworld-v1.xlang.ai/). **xlang-ai/OSWorld** repo: 3,186 stars, Apache-2.0. Source: [repo](https://github.com/xlang-ai/OSWorld).+- Agent S3: 66% on OSWorld at 100 steps alone, 72.6% with Behavior Best-of-N; previous SOTA listed as 63.4% (GTA1 w/ GPT-5); human "~72%"; WindowsAgentArena 50.2% to 56.6% with 3 rollouts; AndroidWorld 68.1% to 71.6%; Sai reports 73% on OSWorld 2.0 (108 tasks); Simular API went live 2026-10-01. Source: [Agent-S README](https://github.com/simular-ai/Agent-S).+- UI-TARS-2: 47.5% OSWorld, 50.6% WindowsAgentArena, 73.3% AndroidWorld. Source: [digitalapplied open-source browser/computer-use agents 2026](https://www.digitalapplied.com/blog/open-source-browser-computer-use-agents-2026).+- Third-party OSWorld-Verified aggregators (vendor-reported, not the official board): "As of September 29, 2026, Qwen3.8 Max leads the OSWorld-Verified leaderboard with 86.1%, followed by Claude Fable 5 (85%) and Claude Mythos 5 (85%)". Source: [benchlm OSWorld-Verified](https://benchlm.ai/benchmarks/osworld-verified). LLM Stats lists Qwen3.8-27B as best open-source at 0.843 (#3). Source: [llm-stats OSWorld-Verified](https://llm-stats.com/benchmarks/osworld-verified); contradicted on the open-weights ranking by [datalearner](https://www.datalearner.com/en/benchmarks/osworld-verified) which lists Kimi K3 at 84.8. Steel's leaderboard notes some entries lack stated step budgets or run counts. Source: [leaderboard.steel.dev OSWorld](https://leaderboard.steel.dev/leaderboards/osworld/).+- Browser benchmarks are not comparable across vendors: "browser-use, Skyvern and Stagehand each publish a headline score from a different suite, graded by a different judge, against a different set of live sites". Browser Use claims 89.1% WebVoyager (and 97% Online-Mind2Web for its hosted cloud agent, 63.3% on BU Bench V1 for the pip library); Skyvern 85.85% WebVoyager and 64.4% on WebBench (a benchmark Skyvern co-authored). Source: [digitalapplied](https://www.digitalapplied.com/blog/open-source-browser-computer-use-agents-2026); [firecrawl best browser agents](https://www.firecrawl.dev/blog/best-browser-agents); [Skyvern README](https://github.com/Skyvern-AI/skyvern).++#### Fact sheet: computer-use and browser agents (stars and licenses observed 2026-10-09)++- **browser-use/browser-use**: 117,335 stars, MIT, pushed 2026-10-07; Python 3.11+ library with a TypeScript version; connects to local, cloud or existing remote browsers over CDP; Claude toolset integration "31 browser actions plus Bash"; Chrome-profile reuse and cloud cookie sync; cloud at $0.02 per browser-hour; BU Bench v2 60-task subset scored by rubric and cost. Source: [repo](https://github.com/browser-use/browser-use). Verdict: **adopt code** (MIT; the natural autonomous layer on top of the caster pool's Chromium).+- **browserbase/stagehand**: 25,584 stars, MIT, now v4 (Go SDK v4.0.0); `act`, `extract`, `observe` in TypeScript, Python and Go plus Playwright-style `goto/click/locator/screenshot`; "hybrid accessibility-tree trimming"; "runs as an extension next to the browser"; server-side cache (`cache: true`) replays identical calls without model tokens; hosted Browserbase MCP server exposing `navigate/act/observe/extract` with `claude mcp add` setup. Source: [repo](https://github.com/browserbase/stagehand). v3 (Feb 2026) was "a complete rewrite ... communicates directly with the browser via Chrome DevTools Protocol ... 44% faster". Source: [dexto browser automation tools](https://www.dexto.ai/blog/best-browser-automation-tools-for-ai-agents/). Verdict: **adopt code** (MIT).+- **browserbase/open-operator**: 1,952 stars, **archived**, last push 2026-05-04. Source: [repo](https://github.com/browserbase/open-operator). **steel-dev/surf.new**: 510 stars, MIT, **archived**. Source: [repo](https://github.com/steel-dev/surf.new). Verdict: **skip** both.+- **steel-dev/steel-browser**: 7,757 stars, Apache-2.0; session persistence across requests (`/sessions`), Puppeteer/Playwright/Selenium over CDP, live session viewer with request logging, proxy chain rotation, stealth plugins and fingerprint management, custom Chrome extensions, `/screenshot` and `/pdf` plus markdown/readability endpoints, Docker image `ghcr.io/steel-dev/steel-browser` on ports 3000/9223. Source: [repo](https://github.com/steel-dev/steel-browser). Verdict: **adopt code** (Apache; this is almost exactly three.ws's "on-demand caster pool" as a packaged product, with the live viewer already built).+- **Skyvern-AI/skyvern**: 23,160 stars, **AGPL-3.0** ("managed cloud's anti-bot measures are excluded from this open-source license"); vision-LLM over Playwright, "swarm of agents"; workflows with blocks (browser task, extraction, validation, loops, file parsing, email, HTTP, custom code); 2FA via TOTP/email/SMS; 1Password/Bitwarden/LastPass; `run_task` SDK with extraction schema; MCP support; WebBench 64.4%. Source: [repo](https://github.com/Skyvern-AI/skyvern). Verdict: **adopt design, rebuild** (AGPL; workflow blocks and 2FA handling are the ideas).+- **magnitudedev/magnitude**: 6,506 stars, Apache-2.0, but the repo is now described as "Open source inference engine for agents that optimizes itself for your exact hardware", i.e. no longer a browser agent. Source: [repo](https://github.com/magnitudedev/magnitude). Verdict: **skip** (pivoted).+- **anthropics/anthropic-quickstarts** (computer-use demo): 17,832 stars, MIT. Source: [repo](https://github.com/anthropics/anthropic-quickstarts). Verdict: **adopt code** (MIT reference for the screenshot-action loop).+- **simular-ai/Agent-S**: 12,566 stars, Apache-2.0; grounding model required (UI-TARS-1.5-7B recommended, grounding dims must match its resolution), `OSWorldACI` translates actions to executable Python, reflection agent (`enable_reflection` default True), `max_trajectory_length` default 8 image turns, optional `call_code_agent` for Python/Bash, Behavior Best-of-N. Source: [repo](https://github.com/simular-ai/Agent-S). Verdict: **adopt code** (Apache; the strongest open desktop agent architecture).+- **bytedance/UI-TARS**: 11,568 stars, Apache-2.0, last push 2026-01-27 (model/paper repo). Source: [repo](https://github.com/bytedance/UI-TARS). **bytedance/UI-TARS-desktop**: 39,219 stars, Apache-2.0; UI-TARS Desktop = native GUI agent with screenshot recognition, local and remote computer/browser operators; Agent TARS = multimodal agent CLI/Web UI whose "kernel is built on MCP" and controls browsers "through a GUI Agent (visual grounding), DOM, or a hybrid of both"; models UI-TARS-1.5 7B and Seed-1.5-VL/1.6; latest listed release Agent TARS CLI v0.3.0 on 2025-11-05. Source: [repo](https://github.com/bytedance/UI-TARS-desktop). Verdict: **adopt code** (Apache; open-weight 7B grounding model is the only self-hostable VLM grounding option here, fits the GPU fleet).+- **xlang-ai/OpenCUA**: 851 stars, MIT, last push 2026-05-25. Source: [repo](https://github.com/xlang-ai/OpenCUA). Verdict: **adopt code** (MIT open foundations; low activity).+- **trycua/cua**: 29,078 stars; "Everything outside Cua Spaces is MIT-licensed", Cua Spaces is **FSL-1.1-MIT** (source-available, converts to MIT two years after each release), `cua-som` is AGPL-3.0-or-later; Lume runs macOS/Linux VMs on Apple Silicon via Virtualization.Framework, Lumier is a Docker-compatible interface, gVisor containers, `cua-spacesd` in-sandbox daemon on port 3211 (processes, files, screenshots, input), SDKs in Python/TypeScript/Swift/Kotlin, CUA-S1 small models, Cua Bench, Spaces runnable in AWS/GCP/Modal. Source: [repo](https://github.com/trycua/cua). Verdict: **adopt code** for the MIT driver/SDK; **adopt design, rebuild** for Spaces (FSL).+- **microsoft/magentic-ui** (MagenticLite): 10,093 stars, MIT, "experimental agent that works across the browser and local file system". Source: [repo](https://github.com/microsoft/magentic-ui). **nanobrowser/nanobrowser**: 14,016 stars, Apache-2.0, Chrome extension multi-agent web automation. Source: [repo](https://github.com/nanobrowser/nanobrowser). **e2b-dev/open-computer-use**: 2,343 stars, Apache-2.0, "AI computer use powered by open source LLMs and E2B Desktop Sandbox". Source: [repo](https://github.com/e2b-dev/open-computer-use). Verdicts: **adopt code** (all permissive); nanobrowser overlaps three.ws's Walk Avatar / Checkout companion browser extensions.++#### Fact sheet: sandboxes and runtimes (stars and licenses observed 2026-10-09)++- **e2b-dev/E2B**: 14,243 stars, Apache-2.0; SDKs `e2b` (JS/Python), `@e2b/code-interpreter`, `@e2b/desktop` / `e2b-desktop`; self-host via Terraform to AWS, GCP, Azure, or a Linux machine ("Embed"). Source: [repo](https://github.com/e2b-dev/E2B). **e2b-dev/infra**: 1,688 stars, Apache-2.0, "The runtime behind every E2B stack: Cloud, Enterprise, and your own machine." Source: [repo](https://github.com/e2b-dev/infra). Third-party 2026 comparisons describe E2B as "Apache-2.0, self-hostable" and the clearest open option. Source: [bex.co self-hosted sandbox comparison](https://bex.co/blog/2026/09/14/ai-agent-sandbox-market-e2b-modal-daytona). Verdict: **adopt code** (Apache; deployable on the pre-approved GCP credits as the sandbox behind headless agent runs and skill handlers that need a real shell).+- **daytonaio/daytona**: 71,593 stars, **archived 2026-10-03**, last push 2026-07-24; LICENSE at v0.190.0 is **AGPL-3.0**; README: "As of June 2026, Daytona's core development has moved to a private codebase ... remains public and free to use, fork, and build on under the LICENSE". Source: [repo](https://github.com/daytonaio/daytona); [raw LICENSE v0.190.0](https://raw.githubusercontent.com/daytonaio/daytona/v0.190.0/LICENSE). Verdict: **skip** (AGPL and unmaintained).+- **modal-labs/modal-client**: 522 stars, Apache-2.0 (SDK only; platform is closed and "no self-host"). Source: [repo](https://github.com/modal-labs/modal-client); [bex.co](https://bex.co/blog/2026/07/07/ai-agent-sandbox-wars). Verdict: **already covered** (three.ws runs its own GPU fleet on GCP; Modal would be a new external paid API, which CLAUDE.md gates).+- **firecracker-microvm/firecracker**: 37,249 stars, Apache-2.0. Source: [repo](https://github.com/firecracker-microvm/firecracker). Verdict: **adopt code** if building a microVM runner in-house instead of E2B.+- **stackblitz/webcontainer-core**: 4,646 stars, MIT, last push 2025-04-22, but that repo is the issue tracker; the runtime is closed and StackBlitz's FAQ says "Organizations wanting to use the API for production usage in a for-profit setting are required to purchase a WebContainer API License". Source: [repo](https://github.com/stackblitz/webcontainer-core); [StackBlitz FAQ](https://developer.stackblitz.com/guides/user-guide/general-faqs). Verdict: **adopt design, rebuild** (three.ws's Live Docs runnable samples would need a paid license to use WebContainers commercially).++#### How three.ws compares (from its own docs)++- agent-screen-control: owner-only "Take control" of an agent's live cast browser with real mouse/keyboard/scroll/URL navigation; the cast browser "holds no wallet and no keys" (sandboxed throwaway Chromium); a 15-second lease token refreshed per input batch and every 8 s, never riding the login cookie; coordinates clamped, keys whitelisted, navigation SSRF-guarded (no localhost, private ranges or metadata endpoints); one driver at a time (`409 in_use`); API `POST /api/agent-screen-control` with `acquire` then bearer-token input; the doc calls this "the first slice ... the same control channel is what a full desktop will ride on next". Source: `docs/agent-screen-control.md`. Related surfaces: "On-demand caster pool (the live browser behind a watched agent)", `workers/agent-screen-worker/`, `services/agent-screen-caster/`.++### Inferences+- Architecturally the field has settled on three perception inputs: screenshots with a VLM grounding model for desktop (Agent S3, UI-TARS, Cua), DOM/accessibility-tree with CDP for browser speed and cost (Browser Use, Stagehand v4, Agent TARS hybrid), and vision-first for anti-fragility against site changes (Skyvern). three.ws's screen-control channel is input/output plumbing with no perception layer; adding Browser Use or Stagehand (both MIT) on the agent side of the existing caster Chromium would give the "agent drives" half of the feature the doc already promises, with the owner's take-the-wheel lease as the interrupt.+- Steel Browser (Apache) is the closest OSS equivalent to the caster pool and ships the pieces three.ws lists as future (session persistence, proxy rotation, live viewer, PDF/screenshot APIs); vendoring or depending on it would be cheaper than extending the caster.+- For a self-hosted desktop (the doc's "full desktop" next step), the open-weight path is UI-TARS-1.5-7B grounding + Agent S3 loop on E2B Desktop or a Firecracker microVM, all Apache-2.0/MIT, and all runnable on the pre-approved GCP GPU credits.+- three.ws's safety model (no keys in the browser, capability leases, SSRF guard) is stricter than any OSS stack reviewed, none of which separate wallet custody from the driven browser; that is a design worth keeping and documenting as a differentiator.++### Gaps+- The official OSWorld-Verified leaderboard table did not render, so the current top entries, step budgets and run counts could not be read from the primary source; the aggregator figures (86.1% Qwen3.8 Max, 85% Claude) are vendor-reported.+- Agent S3's OSWorld-Verified (as opposed to OSWorld) score is not given in its README.+- No OSWorld or WebArena score was found for Stagehand, Cua or Steel.+- Browser Use's current version number and whether it ships an MCP server were not stated on the README.+- WebArena 2026 results were not retrieved; only OSWorld, WebVoyager, WebBench and Online-Mind2Web figures appear above.diff --git a/docs/research/oss-landscape-2026-10-notes/irl-home-robotics.md b/docs/research/oss-landscape-2026-10-notes/irl-home-robotics.mdnew file mode 100644index 000000000..ef70873f5--- /dev/null+++ b/docs/research/oss-landscape-2026-10-notes/irl-home-robotics.md@@ -0,0 +1,135 @@+# Open-source competitors to three.ws physical-world surfaces (smart home, voice hardware, glasses, car, robotics, sign language), October 2026++All star counts, licenses and push dates were read from the GitHub REST API (authenticated `gh api`) on 2026-10-09 unless noted. A repo's license URL is `https://github.com/<owner>/<repo>/blob/<default-branch>/LICENSE`, and the SPDX value comes from `https://api.github.com/repos/<owner>/<repo>`. Context: three.ws already ships `home-assistant-integration/` (HACS custom component, LAN house dials out to the cloud relay), `services/home-relay`, `@three-ws/home-mcp`, `@three-ws/home-bridge`, `/smart-home`, `/voice/home`, `/drive`, `@three-ws/irl`, `docs/home-satellite.md`, `docs/glasses.md`, and `@three-ws/sign-language` (local docs: docs/smart-home.md, home-voice.md, home-satellite.md, glasses.md, irl.md, sign-language.md).++## 1. Smart home and voice stack: which repos beat three.ws, and integrate or rebuild?++### Takeaway+Do not rebuild any of the smart-home stack. Home Assistant (Apache-2.0, ~91k stars) is the device and voice layer; three.ws should stay an integration on top of its open APIs (Assist/conversation, MCP server, WebSocket/REST, Wyoming) and keep its differentiator, the 3D embodied agent. Verdicts: Home Assistant = "already covered" via integration, deepen it; ha-mcp = "adopt design, rebuild" (87-tool surface is a benchmark for `@three-ws/home-mcp`); Wyoming/openWakeWord/linux-voice-assistant = "adopt code" (all permissive) for the satellite; Frigate = integrate via its HA integration, MIT; Matter SDK = do not rebuild, go through HA; openHAB = EPL-2.0, integrate only if demanded.++### Cited Findings+- Home Assistant core: 91,335 stars (2026-10-09), Apache-2.0, pushed 2026-10-09. Same license as three.ws, so code can be copied with attribution. — [repo](https://github.com/home-assistant/core), [license](https://github.com/home-assistant/core/blob/dev/LICENSE.md)+- Core ships first-party LLM/agent integrations as components: `anthropic`, `openai_conversation`, `ollama`, `conversation`, `assist_pipeline`, `assist_satellite`, `mcp` (HA as MCP client) and `mcp_server` (HA as MCP server, depends on `conversation`, mcp==1.28.1, quality scale silver, local_push). — [components dir](https://github.com/home-assistant/core/tree/dev/homeassistant/components), [mcp_server manifest](https://github.com/home-assistant/core/blob/dev/homeassistant/components/mcp_server/manifest.json), [docs](https://www.home-assistant.io/integrations/mcp_server)+- The official MCP server only reaches Assist-exposed entities, has ~10 intent-based tools, and cannot create or edit automations, scripts or dashboards (third-party comparison, not the official docs; verify). — [raspberry.tips guide](https://raspberry.tips/en/smart-home/home-assistant-mcp-server-ai-setup)+- Home Assistant 2026.2 renamed "Add-ons" to "Apps" in the UI (secondary source). — [raspberry.tips](https://raspberry.tips/en/smart-home/home-assistant-mcp-server-ai-setup)+- ha-mcp (homeassistant-ai/ha-mcp): 4,997 stars (2026-10-09), MIT, pushed 2026-10-09, README badge advertises 87 tools, installs via HACS custom component with in-process server, an admin HA-MCP sidebar panel, backups/restore of helper edits, opt-in file and YAML editing flags. Far broader than the official server (automation/script/dashboard/helper authoring). — [repo](https://github.com/homeassistant-ai/ha-mcp), [license](https://github.com/homeassistant-ai/ha-mcp/blob/main/LICENSE.md)+- Other HA MCP server: voska/hass-mcp 343 stars, MIT. — [repo](https://github.com/voska/hass-mcp)+- Wyoming protocol now lives at OHF-Voice/wyoming: 401 stars, MIT (rhasspy/wyoming URL redirects here). — [repo](https://github.com/OHF-Voice/wyoming)+- rhasspy/wyoming-satellite: 1,241 stars, MIT, last push 2026-01-24 (maintenance mode; successor is below). — [repo](https://github.com/rhasspy/wyoming-satellite)+- OHF-Voice/linux-voice-assistant: 647 stars, Apache-2.0, pushed 2026-10-01. Linux ESPHome-protocol voice satellite (successor path for DIY satellites). — [repo](https://github.com/OHF-Voice/linux-voice-assistant)+- OHF-Voice/intents (Assist sentence/intents library, many languages): 628 stars, CC-BY-4.0 (data, attribution required). — [repo](https://github.com/OHF-Voice/intents)+- openWakeWord (dscripka): 2,828 stars, Apache-2.0, last push 2025-12-30; wyoming-openwakeword 205 stars, Apache-2.0. Pretrained models carry their own terms (check per model; the author notes CC BY-NC-SA on some training data derived models). — [repo](https://github.com/dscripka/openWakeWord), [wyoming wrapper](https://github.com/rhasspy/wyoming-openwakeword)+- microWakeWord (kahrendt): 24 stars, Apache-2.0; powers on-device wake words on Voice PE ("Okay Nabu", "Hey Jarvis", "Hey Mycroft"). — [repo](https://github.com/kahrendt/microWakeWord), [HA blog](https://home-assistant.io/blog/2024/12/19/voice-preview-edition-the-era-of-open-voice)+- Piper TTS: rhasspy/piper 11,297 stars, MIT, archived-era last push 2025-08-26; successor OHF-Voice/piper1-gpl 5,807 stars, GPL-3.0 (copyleft: do not link into an Apache-2.0 product; run as separate process/service only). — [piper](https://github.com/rhasspy/piper), [piper1-gpl](https://github.com/OHF-Voice/piper1-gpl)+- Home Assistant Voice Preview Edition (esphome/home-assistant-voice-pe): 766 stars, license reported as "Other"/NOASSERTION (mixed firmware and hardware files; read LICENSE before copying), ESP32-S3 + XMOS XU316 (echo cancel, noise suppression), on-device wake word, fully open software/firmware/hardware per vendor. — [repo](https://github.com/esphome/home-assistant-voice-pe), [LICENSE](https://github.com/esphome/home-assistant-voice-pe/blob/dev/LICENSE), [HA blog](https://home-assistant.io/blog/2024/12/19/voice-preview-edition-the-era-of-open-voice)+- ESPHome: 11,799 stars, SPDX NOASSERTION (mixed: GPLv3 for C++ runtime/firmware, MIT for Python tooling per its README; confirm before vendoring). — [repo](https://github.com/esphome/esphome), [license](https://github.com/esphome/esphome/blob/dev/LICENSE)+- Home Assistant Android companion app: 3,956 stars, Apache-2.0. — [repo](https://github.com/home-assistant/android)+- Frigate NVR: 36,447 stars, MIT, pushed 2026-10-09; local object detection, HA integration, GenAI event descriptions. — [repo](https://github.com/blakeblackshear/frigate), [license](https://github.com/blakeblackshear/frigate/blob/dev/LICENSE)+- Matter SDK (project-chip/connectedhomeip): 8,961 stars, Apache-2.0. — [repo](https://github.com/project-chip/connectedhomeip), [license](https://github.com/project-chip/connectedhomeip/blob/master/LICENSE)+- openHAB: openhab-core 1,148 stars and openhab-addons 2,063 stars, both EPL-2.0 (weak copyleft, file-level; fine to call over APIs, avoid copying files into Apache-2.0 code without keeping EPL on those files); openhab-android 652 stars EPL-2.0. — [core](https://github.com/openhab/openhab-core), [addons](https://github.com/openhab/openhab-addons)++### Inferences+- three.ws's unique edge versus all of these is the 3D embodied agent plus the dial-out relay (no port forwarding); no surveyed repo has a 3D avatar front end. Adopting the wider surface from HA is cheaper than rebuilding it.+- Concrete adoption list: (a) widen `@three-ws/home-mcp` toward ha-mcp's categories (automation/script/helper authoring with backup-before-edit) but only via HA's WebSocket/REST API, keeping MIT attribution if any code is copied; (b) speak Wyoming/ESPHome-voice to satellites so a Voice PE or linux-voice-assistant node can front the 3D agent (`docs/home-satellite.md` already exists, so audit it against Wyoming events); (c) expose Frigate events as agent cues through HA's Frigate integration rather than touching Frigate directly; (d) use HA `intents` data (CC-BY-4.0) for local intent matching with attribution.+- Piper (GPL) and ESPHome (mixed GPL) are process-boundary only for an Apache-2.0 platform.++### Gaps+- Could not read the official Home Assistant 2026.x release notes (search returned only third-party pages), so exact 2026 Assist/LLM features (e.g. Assist API changes, conversation agent tools) are unverified.+- Did not verify the Voice PE hardware-design license text or ESPHome's per-directory licenses; search showed no explicit hardware license.+- Did not find OHF-Voice/wyoming-satellite (the old rhasspy repo is the one that exists).++## 2. Wearables and smart glasses: which repos beat three.ws?++### Takeaway+MentraOS (Apache-2.0, 2.4k stars) is the clear open glasses OS and is directly adoptable: three.ws's "nearest-agent cue" should ship as a MentraOS MiniApp to reach Mentra Live, Even Realities G2 and others without hardware work. Omi (MIT, 13.7k stars) is the biggest open wearable-AI stack; adopt its design and possibly its protocol/SDK. Open Interpreter's 01 hardware repo is AGPL and stale; skip.++### Cited Findings+- MentraOS (Mentra-Community/MentraOS): 2,382 stars (2026-10-09), Apache-2.0, pushed 2026-10-09. README: "Write Once, Run on Any Smart Glasses", handles pairing/connection/streaming/hardware access, a MiniApp Store, developer console, supports Mentra Live and Even Realities G2 among others; "every component is open source under Apache 2.0". — [repo](https://github.com/Mentra-Community/MentraOS), [README](https://github.com/Mentra-Community/MentraOS/blob/main/README.md)+- Omi (BasedHardware/omi): 13,673 stars (2026-10-09), MIT, pushed 2026-10-09; open wearable AI (device firmware, mobile app, backend, app/plugin store). — [repo](https://github.com/BasedHardware/omi), [license](https://github.com/BasedHardware/omi/blob/main/LICENSE)+- Brilliant Labs Frame: frame-codebase 529 stars, SPDX NOASSERTION (check the file; likely mixed), last push 2025-10-05; frame-sdk-python 11 stars, MIT; frame_realtime_gemini_voicevision 81 stars, BSD-3-Clause (2026-02-09). — [codebase](https://github.com/brilliantlabsAR/frame-codebase), [sdk-python](https://github.com/brilliantlabsAR/frame-sdk-python), [gemini demo](https://github.com/brilliantlabsAR/frame_realtime_gemini_voicevision)+- xg-glass-sdk (hkust-spark): 46 stars, Apache-2.0, pushed 2026-10-05; cross-glasses SDK, a small rival to MentraOS. — [repo](https://github.com/hkust-spark/xg-glass-sdk)+- Open Interpreter 01 (OpenInterpreter/01): 5,159 stars, AGPL-3.0, last push 2024-11-01 (stale). 01-app 342 stars AGPL-3.0. The org's open-interpreter repo is 68,534 stars Apache-2.0 (code-executing agent, not hardware). — [01](https://github.com/OpenInterpreter/01), [open-interpreter](https://github.com/openinterpreter/openinterpreter)+- openglasses-fingerspelling-ctc on Hugging Face (Core ML fingerspelling model for glasses, Apache-2.0 weights) shows glasses + sign language is an emerging intersection. — [HF](https://huggingface.co/Skunk0/openglasses-fingerspelling-ctc)++### Inferences+- Verdict: MentraOS = "adopt code" (build a three.ws MiniApp; Apache-2.0 compatible; also lets glasses show the nearest-agent cue and AR avatar). Omi = "adopt design, rebuild" (always-on memory/conversation capture plus app store); AGPL 01 = do not touch.++### Gaps+- Did not read MentraOS SDK API details or Omi's backend license split (the repo root says MIT; subdirectories not checked). Brilliant Labs frame-codebase license not read.+- No 2026 Meta/Android XR open stack found; none searched.++## 3. In-car++### Takeaway+openpilot (MIT, 63.9k stars) is the only large open in-car project, but it is driving-assistance on comma hardware, not a conversational agent; it is a design reference at most. No open Android Auto assistant of note was found. /drive can be called "already covered" for the conversational side.++### Cited Findings+- openpilot: 63,851 stars (2026-10-09), MIT, pushed 2026-10-09; "operating system for robotics", upgrades driver assistance on 300+ supported cars. — [repo](https://github.com/commaai/openpilot), [license](https://github.com/commaai/openpilot/blob/master/LICENSE)++### Inferences+- openpilot's value to three.ws is limited to driver-monitoring and vehicle-signal ideas; it must not be presented as a safety system, and repo licensing says nothing about road legality.++### Gaps+- GitHub search for open Android Auto / CarPlay LLM assistants returned nothing relevant; absence is not proof none exist.++## 4. Robotics and embodiment++### Takeaway+LeRobot (Apache-2.0, 28k stars) is the hub: it bundles SO-101 teleop/training, Reachy 2, Unitree G1 and a model zoo (Pi0, Pi0.5, SmolVLA, GR00T N1.7, XVLA, etc.). three.ws should integrate (render a LeRobot robot's URDF/state as a 3D agent body, stream joint states) rather than rebuild; watch weight licenses, which vary.++### Cited Findings+- LeRobot (huggingface/lerobot): 28,035 stars (2026-10-09), Apache-2.0, pushed 2026-10-09. Supported hardware: SO100, LeKiwi, Koch, HopeJR, OMX, EarthRover, Reachy2, Unitree G1 and more; policies: Pi0, Pi0Fast, Pi0.5, GR00T N1.7, SmolVLA, XVLA, EO-1, MolmoAct2, WALL-OSS and others. — [repo](https://github.com/huggingface/lerobot), [license](https://github.com/huggingface/lerobot/blob/main/LICENSE)+- SO-ARM100/101 (TheRobotStudio): 7,706 stars, Apache-2.0 (hardware CAD and firmware repo). — [repo](https://github.com/TheRobotStudio/SO-ARM100)+- Reachy Mini (pollen-robotics): 1,540 stars, Apache-2.0; wireless (Raspberry Pi), Lite (USB) and MuJoCo simulation variants. Hugging Face launched a Reachy Mini app store (~200 apps) in May 2026, reported ~10,000 units sold; $299 Lite, $449 wireless at launch; no creator monetization yet. reachy_mini_conversation_app: 320 stars, Apache-2.0. — [repo](https://github.com/pollen-robotics/reachy_mini), [conversation app](https://github.com/pollen-robotics/reachy_mini_conversation_app), [Axios](https://axios.com/2026/05/06/hugging-face-consumer-robot-app-store), [Robot Report](https://www.therobotreport.com/hugging-face-launches-agentic-toolkit-for-reachy-mini/)+- openpi (pi0/pi0.5): 14,151 stars, code Apache-2.0, last push 2026-08-24; checkpoints (pi0.5 base, LIBERO, DROID) served from `gs://openpi-assets`. Weight license not stated in the README lines I read; pi0 derives from PaliGemma so Gemma terms may apply (unverified). — [repo](https://github.com/Physical-Intelligence/openpi), [LICENSE](https://github.com/Physical-Intelligence/openpi/blob/main/LICENSE)+- Isaac-GR00T: 8,179 stars, code Apache-2.0, model weights under the NVIDIA Open Model License (permits commercial use with conditions; not Apache). — [repo](https://github.com/NVIDIA/Isaac-GR00T), [NVIDIA OML](https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/)+- OpenVLA: 7,127 stars, MIT code, last push 2025-03-23; weights derive from Llama 2 and are subject to the Llama Community License (not OSI open). openvla-oft 1,415 stars MIT. — [repo](https://github.com/openvla/openvla), [Llama license](https://ai.meta.com/llama/license/)+- ROS 2 core: ros2/ros2 6,141 stars; rclpy 484 stars Apache-2.0. GitHub search found no standout "ROS 2 LLM bridge" repo with real traction; chrismatthieu/openclaw-robotics (4 stars, MIT) is tiny. — [ros2](https://github.com/ros2/ros2), [rclpy](https://github.com/ros2/rclpy)++### Inferences+- three.ws's differentiator is the avatar/3D rig and browser rendering; a LeRobot bridge (joint states to a three.ws rig) and a Reachy Mini app (the agent as a Reachy app via the conversation-app pattern) are cheap integrations. Verdict: LeRobot "adopt code" as a dependency/bridge; Reachy Mini "adopt design" (app-store distribution model); VLAs "already out of scope", consume only via LeRobot with per-weight license checks.++### Gaps+- SmolVLA and pi0 weight license texts not read. No ROS 2 LLM bridge evaluated in depth.++## 5. Sign-language AI++### Takeaway+No open repo clearly beats three.ws's webcam ASL recognition on a published, comparable accuracy. The best leads are the Skunk0 Apache-2.0 fingerspelling CTC model (20.8% mean CER on a 300-sequence held-out set) and the OpenFS paper (93.7 letter accuracy on FSboard, 8.81M params; code/weights availability unverified). sign.mt is NC-licensed, so only design is adoptable.++### Cited Findings+- MediaPipe: 37,202 stars, Apache-2.0, pushed 2026-10-09; Holistic/Hand landmarks are the standard input for ASL models. — [repo](https://github.com/google-ai-edge/mediapipe), [license](https://github.com/google-ai-edge/mediapipe/blob/master/LICENSE)+- sign/translate (sign.mt): 794 stars, license CC BY-NC-SA 4.0 (non-commercial, share-alike; cannot be reused in a commercial Apache-2.0 platform), last push 2026-08-11. Two-way spoken/signed translation, offline, photorealistic avatars. — [repo](https://github.com/sign/translate), [LICENSE](https://github.com/sign/translate/blob/master/LICENSE.md), [EMNLP 2024 paper](https://aclanthology.org/2024.emnlp-demo.19)+- sign-language-processing org: `pose` 114 stars MIT (2026-09-19); `spoken-to-signed-translation` 103 stars MIT (2026-10-02); `sign-language-processing.github.io` 147 stars CC-BY-4.0 (survey); `recognition` 1 star MIT (stale 2024). — [pose](https://github.com/sign-language-processing/pose), [spoken-to-signed](https://github.com/sign-language-processing/spoken-to-signed-translation)+- Skunk0/openglasses-fingerspelling-ctc (Hugging Face): Core ML CTC over MediaPipe Holistic landmarks, Apache-2.0 weights, 20.8% mean CER (median 10.8%) on 300-seq held-out, trained on Google ASL Fingerspelling (CC-BY 4.0). — [HF](https://huggingface.co/Skunk0/openglasses-fingerspelling-ctc)+- ColdSlim/ASL-TFLite-Edge: Apache-2.0 TFLite, 59 ASL classes from hand landmarks, no accuracy reported. — [HF](https://huggingface.co/ColdSlim/ASL-TFLite-Edge/blob/main/README.md)+- OpenFS (arXiv 2602.22949, Feb 2026): 93.7 letter accuracy, 59.4 top-1 on FSboard, 8.81M params vs 300M ByT5-based baseline. — [paper](https://arxiv.org/abs/2602.22949)+- Older: WLASL dataset repo 1,281 stars, no license on repo (dataset terms are restrictive, treat as research-only); AI4Bharat OpenHands 135 stars Apache-2.0, stale since 2023; sign-language-translator 376 stars Apache-2.0, stale 2024. — [WLASL](https://github.com/dxli94/WLASL), [OpenHands](https://github.com/AI4Bharat/OpenHands)++### Inferences+- three.ws's `@three-ws/pose` plus the camera-graded Sign Mirror are ahead of GitHub's hobby ASL repos (single-digit stars). Worth benchmarking its recognizer against the Skunk0 CTC model on the same held-out set. The Google ASL Fingerspelling corpus (CC-BY 4.0) is the best commercially usable training data; sign.mt avatar/translation ideas must be rebuilt, not copied.++### Gaps+- No verified 2026 open-weights ASL word-level (isolated or continuous) model with published accuracy found. OpenFS weights not confirmed. three.ws's own accuracy number was not compared (not measured here).++## Summary verdict table++| Repo | Stars (2026-10-09) | License | Verdict |+|---|---|---|---|+| home-assistant/core | 91,335 | Apache-2.0 | Integrate via open APIs (already covered, deepen) |+| homeassistant-ai/ha-mcp | 4,997 | MIT | Adopt design, rebuild; or recommend alongside |+| OHF-Voice/wyoming, linux-voice-assistant, openWakeWord | 401 / 647 / 2,828 | MIT / Apache-2.0 / Apache-2.0 | Adopt code |+| piper1-gpl | 5,807 | GPL-3.0 | Process boundary only |+| Frigate | 36,447 | MIT | Integrate via HA, do not rebuild |+| Matter SDK | 8,961 | Apache-2.0 | Via HA only |+| ESPHome / Voice PE | 11,799 / 766 | NOASSERTION / Other | Adopt design; check files |+| openHAB core/addons | 1,148 / 2,063 | EPL-2.0 | API integration only |+| MentraOS | 2,382 | Apache-2.0 | Adopt code (MiniApp) |+| Omi | 13,673 | MIT | Adopt design, rebuild |+| Open Interpreter 01 | 5,159 | AGPL-3.0 | Skip (stale) |+| openpilot | 63,851 | MIT | Design reference only |+| LeRobot / SO-ARM100 | 28,035 / 7,706 | Apache-2.0 | Adopt code (bridge) |+| Reachy Mini | 1,540 | Apache-2.0 | Adopt design (app distribution) |+| openpi / GR00T / OpenVLA | 14,151 / 8,179 / 7,127 | Apache / Apache+NVIDIA OML / MIT+Llama | Via LeRobot, check weight license |+| MediaPipe | 37,202 | Apache-2.0 | Already covered (in use) |+| sign.mt | 794 | CC BY-NC-SA 4.0 | Non-commercial; design only |diff --git a/docs/research/oss-landscape-2026-10-notes/worlds-engines.md b/docs/research/oss-landscape-2026-10-notes/worlds-engines.mdnew file mode 100644index 000000000..e6133abb0--- /dev/null+++ b/docs/research/oss-landscape-2026-10-notes/worlds-engines.md@@ -0,0 +1,181 @@+# OSS repos that beat three.ws worlds: web 3D engines/editors, multiplayer frameworks, AI-NPC simulations, world generation, Gaussian splats, WebXR (state as of 2026-10-09)++Method note for the writer: every star count below was read from the GitHub REST API (`gh api repos/<owner>/<repo>`) on **2026-10-09**; "last push" is the API `pushed_at` date. License SPDX ids come from the API `license` field, and for every repo the API reported as `NOASSERTION` the LICENSE file text was read directly (links given). three.ws's own stack was read from its `package.json` and docs: `three ^0.184.0` (178 `WebGLRenderer` references vs 9 `WebGPURenderer`), `@sparkjsdev/spark ^2.2.0` for splats, `@dimforge/rapier3d-compat ^0.19.3` for physics, `colyseus.js 0.16.22` for the `/play` client, `three-pathfinding` + recast for NPC navigation, `postprocessing ^6.39.5`, `@gltf-transform/*`, `meshoptimizer`, `draco3dgltf`, `three-gpu-pathtracer`, `three-mesh-bvh`, and a vendored three.js editor as Scene Studio.++## Fact sheet: the leading repos (stars dated 2026-10-09, SPDX license, what an Apache-2.0 platform may reuse, standout features, verdict)++### Takeaway+Of ~60 repos checked, the permissively licensed ones that three.ws can lift code from outright are the PlayCanvas engine/editor/SuperSplat trio (MIT), Spark (MIT), Babylon.js 9 (Apache-2.0), Colyseus 0.18 (MIT), AI Town (MIT), Stanford Generative Agents (Apache-2.0), the pmndrs ecosystem (MIT/ISC), Matrix-Game and ABot-World (MIT/Apache-2.0), and model-viewer (Apache-2.0). The ones to copy only the design from are Hyperfy (GPL-3.0), SpacetimeDB (BSL 1.1), iR Engine (CPAL-1.0), Hytopia's SDK (platform-locked limited-use license), Needle and Wonderland (proprietary engines), the Inria 3DGS trainer (non-commercial), lingbot-world-v2 (CC BY-NC-SA 4.0), and Tencent HY-World (territory-restricted community license).++### Cited Findings++**Web 3D engines and editors**++- **three.js** (mrdoob/three.js): 116,161 stars on 2026-10-09, MIT, last push 2026-10-09; latest release r186 published 2026-09-24 per the GitHub releases API — [Source](https://github.com/mrdoob/three.js). Three.ws is on `three ^0.184.0`, two releases behind. Verdict: **already covered** as the base; adopt the r185/r186 WebGPU features listed under Key Question 1.+- **Babylon.js** (BabylonJS/Babylon.js): 26,144 stars, Apache-2.0, last push 2026-10-09; latest release tag 9.30.0 on 2026-10-08 — [Source](https://github.com/BabylonJS/Babylon.js). Babylon.js 9.0 was announced 2026-03-26 with Clustered Lighting for hundreds of lights, a Frame Graph for custom render pipelines, Node Particle Editor with Flow Maps, OpenPBR (alpha), SDF 3D text, 3MF export, and Gaussian-splat shadows, SOG and Triangle Splatting support — [Source](https://blogs.windows.com/windowsdeveloper/2026/03/26/announcing-babylon-js-9-0); [Source](https://radiancefields.com/babylon.js-v9.0-3dgs-gets-shadows-sogs-and-triangle-splatting-support-announced). Babylon 8.0 (March 2025) rewrote all core shaders in WGSL so WebGPU needs no translation layer and the WebGPU bundle is roughly half the size, and shipped Node Render Graph (alpha), GPU mesh picking, Smart Filters, area lights, and SPZ/compressed-PLY splats at 32 bytes/splat — [Source](https://blogs.windows.com/windowsdeveloper/2025/03/27/announcing-babylon-js-8-0/). Apache-2.0 is directly compatible with three.ws's license. Verdict: **adopt design, rebuild** (engine mismatch; copy the Node Material Editor / Frame Graph / clustered-lighting / splat-shadow ideas; the SOG loader and OpenPBR mapping are small enough to port as code).+- **PlayCanvas engine** (playcanvas/engine): 16,996 stars, MIT, last push 2026-10-09; latest release v2.23.1 on 2026-10-07 — [Source](https://github.com/playcanvas/engine). 2026 engine work is splat-centric: v2.16.0 (2026-02-03) added 2D Gaussian Splatting and WGSL subgroups; v2.17.0 (March 2026) added half-precision SH on WebGPU, a GPU-driven sort/cull pipeline and a global splat budget distributed by distance and screen coverage; v2.19.0 (June 2026) moved sorting from CPU workers to GPU compute shaders with automatic WebGL2 fallback, with a single Japanese report of up to 5.7x frame-rate gains on M4 Max; v2.22.0 introduced error-driven LOD and v2.23.0 reverted to distance-only LOD, added target/limit splat budgets and sort-free stochastic rendering on WebGPU — [Source](https://app.cinevva.com/news/2026-03-06-playcanvas-v217-webgpu-splatting); [Source](https://radiancefields.com/playcanvas-engine-v2.23.0-returns-gaussian-splat-lod-to-distance-only-selection); [Source](https://cgworld.jp/flashnews/01-202606-SuperSplat.html). Verdict: **adopt code** for the splat LOD budget and GPU sort (MIT), **adopt design** for the rest.+- **PlayCanvas Editor** (playcanvas/editor): 1,335 stars, MIT, last push 2026-10-07; latest release v2.34.0 on 2026-10-07 — [Source](https://github.com/playcanvas/editor). The README states the Editor "includes a built-in Model Context Protocol (MCP) integration, allowing AI assistants to inspect and modify projects directly through an open Editor session" (repo playcanvas/editor-mcp-server) and that the open frontend is loaded against playcanvas.com with `?use_local_frontend`, i.e. the backend is the hosted service — [Source](https://github.com/playcanvas/editor). The frontend was open-sourced under MIT and the hosted editor/backend remains a commercial product — [Source](https://blog.playcanvas.com/playcanvas-editor-frontend-is-now-open-source/); [Source](https://app.cinevva.com/blog/2026-06-09-web-game-engines-2026-comparison). Editor 2.32 (2026-09-22) added an inspector-configured Joint component (hinges, sliders, ball joints, springs), parallax occlusion, per-cascade shadow control, text justification and reworked splat LOD controls — [Source](https://blog.playcanvas.com/). Real-time collaboration is Google-Docs style with a presence bar, Scripts 2.0 brought real-time collaborative code editing, and the platform has version control with checkpoints and branches; pricing is a free tier, Personal about $15/month, Organization about $50/seat/month — [Source](https://developer.playcanvas.com/user-manual/editor/realtime-collaboration); [Source](https://blog.playcanvas.com/playcanvas-scripts-2-0/); [Source](https://app.cinevva.com/guides/playcanvas-vs-threejs). Verdict: **adopt design, rebuild** (the frontend is MIT but useless without their proprietary realtime backend; the editor-MCP-server pattern, presence bar, and branch/checkpoint model are the things to copy into Scene Studio).+- **SuperSplat** (playcanvas/supersplat): 10,313 stars, MIT, last push 2026-10-05; latest release v3.5.2 on 2026-10-05 — [Source](https://github.com/playcanvas/supersplat). 2026 posts: Editor 3.0 rebuilt on WebGPU (2026-09-09) for larger scenes with less memory; a Publishing API (2026-08-17) with Varjo Teleport, XGRIDS and LichtFeld Studio as launch partners; "vibe code splat apps", 360° video and SPZ export (2026-07-09); WebGPU renderer plus streamed automatic LOD (2026-06-03); collision generation so splats can be walked on (2026-05-14); a tutorial adding physics colliders, light probes, a Recast navmesh, NPCs and an FPS loop to a splat scene (2026-04-22) — [Source](https://blog.playcanvas.com/). Verdict: **adopt code** (MIT) for SOG/LOD generation, collision generation and walk mode in three.ws's Splat Viewer.+- **Needle Engine** (needle-tools/needle-engine-support): 609 stars, no license file (docs/samples repo; engine is closed), last push 2026-10-08 — [Source](https://github.com/needle-tools/needle-engine-support). Features: scenes export from Unity and from a Blender add-on with 100+ components, networking built into the runtime (networked components, variable sync, voice chat, XR), Rapier physics, automatic glTF/texture compression, VR (Quest, Vision Pro), Android WebXR AR, and iOS AR via App Clips/USDZ "Everywhere Actions" — [Source](https://engine.needle.tools/docs/explanation/core-concepts/features-overview). Commercial use requires a Pro license, flat fee with no per-view charges — [Source](https://engine.needle.tools/docs/reference/faq.html). Verdict: **adopt design, rebuild** (proprietary; copy the DCC-export and USDZ-action pipeline ideas).+- **Wonderland Engine** (WonderlandEngine/api): 9 stars, MIT for the JS API only, last push 2024-10-14; the editor is closed — [Source](https://github.com/WonderlandEngine/api). Pricing: free up to $120k annual revenue, 10% royalty above, Enterprise seated licensing — [Source](https://wonderlandengine.com/pricing/). Verdict: **adopt design, rebuild** (proprietary editor).+- **Godot** (godotengine/godot): 118,104 stars, MIT, last push 2026-10-08; latest release 4.7.2-stable on 2026-08-18 — [Source](https://github.com/godotengine/godot). Web export remains WebGL 2.0 via the Compatibility renderer only; Forward+/Mobile are unsupported in browsers and "Godot currently does not support WebGPU, which is a prerequisite" — [Source](https://docs.godotengine.org/en/4.7/tutorials/export/exporting_for_web.html). A secondary report says 4.7 added wasm64 exports lifting the 4 GB ceiling, and corrected itself on 2026-07-14 after wrongly claiming WebGPU export behind a flag — [Source](https://app.cinevva.com/news/2026-06-19-godot-4-7-released). C# projects cannot be exported to the web — [Source](https://docs.godotengine.org/en/4.5/tutorials/export/exporting_for_web.html). Verdict: **already covered** (not a fit for a three.js platform; no WebGPU on web).+- **Bevy** (bevyengine/bevy): 48,736 stars, Apache-2.0 (dual MIT in practice; API reports Apache-2.0), last push 2026-10-08; v0.20.0 released 2026-10-08 — [Source](https://github.com/bevyengine/bevy). Bevy 0.20 (2026-10-08, 227 contributors, 817 PRs) ships Solari real-time path tracer now on Metal with DLSS-RR 4.5, BSN scene syntax, WESL shaders, Feathers editor-centric widgets, and states there is still no full editor (a `.bsn` asset format and remote inspector are planned); mesh shaders are "not supported on web platforms" — [Source](https://bevy.org/news/bevy-0-20/). Verdict: **adopt design, rebuild** (Rust/wasm; nothing to lift into JS).+- **React Three Fiber ecosystem**: react-three-fiber 32,804 stars, MIT, v9.8.1 on 2026-09-24 — [Source](https://github.com/pmndrs/react-three-fiber); drei 9,913 stars, MIT — [Source](https://github.com/pmndrs/drei); uikit 3,248 stars, MIT text under Bela Bohlender copyright (API shows NOASSERTION), latest v1.0.46 on 2025-10-16, last push 2026-10-02 — [Source](https://github.com/pmndrs/uikit/blob/main/LICENSE); pmndrs/xr 2,616 stars, same MIT text — [Source](https://github.com/pmndrs/xr/blob/main/LICENSE); react-three-rapier 1,440 stars, MIT, last push 2025-11-03 (11 months stale) — [Source](https://github.com/pmndrs/react-three-rapier); dimforge/rapier 5,831 stars, Apache-2.0, latest tag v017.1; the separate rapier.js repo is archived and migrated into the main repo — [Source](https://github.com/dimforge/rapier); koota ECS 751 stars, ISC — [Source](https://github.com/pmndrs/koota); viverse toolkit 132 stars, MIT text — [Source](https://github.com/pmndrs/viverse); triplex "open source visual workspace for React / Three Fiber" 1,323 stars but **no LICENSE file detected by the API**, last push 2026-06-11 — [Source](https://github.com/pmndrs/triplex). Verdict: **adopt code** for uikit (3D in-world UI toolkit, which three.ws lacks) and pmndrs/xr; **already covered** for rapier (three.ws is on rapier3d-compat 0.19.3); triplex needs a license check before any reuse.+- **A-Frame** (aframevr/aframe): 17,650 stars, MIT, v1.8.0 on 2026-06-24 — [Source](https://github.com/aframevr/aframe). **Cocos engine**: 9,850 stars, MIT text (API NOASSERTION) — [Source](https://github.com/cocos/cocos-engine/blob/v3.8.9/LICENSE.md). **Filament** (google/filament, native PBR reference): 20,577 stars, Apache-2.0 — [Source](https://github.com/google/filament). Verdict: already covered / reference only.+- **Three.js editor forks with traction**: z2586300277/three-editor 594 stars (MulanPSL-2.0), spite/ThreeJSEditorExtension 186 (MIT), zhangbo126/ThreeFlow 91 (AGPL-3.0) — [Source](https://github.com/z2586300277/three-editor). None adds collaboration; a community prototype added realtime collab to the three.js editor in ~150 lines using Ably, and visual scripting is only "planned" in community roadmaps — [Source](https://discourse.threejs.org/t/three-js-editor-collaborative-in-150-lines-of-javascript/91900). Verdict: already covered (three.ws's vendored editor is on par with these).++**Multiplayer and metaverse frameworks**++- **Hyperfy** (hyperfy-xyz/hyperfy): 305 stars, **GPL-3.0**, last push 2026-04-21; latest GitHub release v0.10.0 (2025-04-01) while CHANGELOG.md already lists v0.16.0 — [Source](https://github.com/hyperfy-xyz/hyperfy); license file — [Source](https://github.com/hyperfy-xyz/hyperfy/blob/main/LICENSE). README: standalone persistent worlds on your own domain, realtime in-world content creation, JavaScript app system, portable avatars, PhysX physics, WebXR; "still in alpha" while moving reference-platform code (hyperfy-ref) into self-hostable worlds; v0.16.0 added "support for AI and AI generated apps via ChatGPT, Grok, Gemini", remote databases, XR building basics, `prim` shapes optimized for thousands of unique primitives, `player.push(force)`, and a distance-based `animate` event — [Source](https://github.com/hyperfy-xyz/hyperfy). GPL-3.0 means copying any Hyperfy source into an Apache-2.0 codebase would force the combined work under GPL. Verdict: **adopt design, rebuild** (in-world building UI, sandboxed JS "apps", AI-generated apps).+- **Colyseus** (colyseus/colyseus): 7,353 stars, MIT, last push 2026-10-07; 0.18 released 2026-08-25 — [Source](https://github.com/colyseus/colyseus). 0.18 release notes: "Client-side prediction & lag compensation: server-authoritative simulation with built-in prediction, reconciliation, and rewind. Existing rooms are unaffected"; server `defineInput()`, `setFixedTimestep()`, `allowRewindState()` + `rewind.lastSeenBy()` for hit tests "against what the shooter saw"; client `room.input()`, `Predict` reconciler, `room.clock` with `serverNow()/renderNow()/rtt()/jitter()`; a drift-telemetry and latency/jitter simulator in `@colyseus/sdk/debug`; a Prediction Playground with 12 labs; sponsors include Bloxd, Poki, Pixels — [Source](https://github.com/colyseus/colyseus/releases/tag/0.18). Managed Colyseus Cloud starts at $15/month — [Source](https://tech-insider.org/colyseus-cloud-setup-2026/). Three.ws's `/play` client pins `colyseus.js 0.16.22`. Verdict: **adopt code** (same framework; upgrade to 0.18 for its netcode, see Key Question 2).+- **Hathora**: shut down game-server hosting on 2026-05-05 after the team joined Fireworks AI; customers were pointed to GameFabric by Nitrado; Stormgate lost its online modes — [Source](https://crux.supercraft.host/blog/hathora-shut-down-where-to-go-after-may-2026/); [Source](https://www.techspot.com/news/111969-stormgate-servers-go-dark-following-ai-focused-hosting.html). The `hathora/hathora` GitHub repo returns 404 — observed via API 2026-10-09. Verdict: **not a competitor anymore**.+- **Croquet / Multisynq** (croquet/croquet): 38 stars, Apache-2.0, last push 2026-07-26, 4,123 commits; multisynq/multisynq-client 236 stars, Apache-2.0, **archived**, last push 2025-08-25 — [Source](https://github.com/croquet/croquet); [Source](https://github.com/multisynq/multisynq-client). Model: "the multiplayer code is executed on each client in a synchronized virtual machine, rather than on a server", with a lightweight reflector ordering messages; sessions run over "Multisynq's global DePIN network" with an API key from multisynq.io — [Source](https://github.com/croquet/croquet); Multisynq client 1.0 public release 2025-04-23 — [Source](https://unpkg.com/@multisynq/[email protected]/README.md). Verdict: **adopt design, rebuild** (deterministic replicated simulation is interesting for cheap co-presence rooms, but low traction and a token-gated network).+- **Nakama** (heroiclabs/nakama): 13,481 stars, Apache-2.0, last push 2026-10-08; v3.41.0 on 2026-09-18 — [Source](https://github.com/heroiclabs/nakama). Architecture details under Key Question 2. Verdict: **adopt design, rebuild** (leaderboards/matchmaking/party primitives; three.ws already has its own DB-backed economy and social graph).+- **SpacetimeDB** (clockworklabs/SpacetimeDB): 25,267 stars, **BSL 1.1** (API NOASSERTION), last push 2026-10-09; v2.11.0 on 2026-10-07 — [Source](https://github.com/clockworklabs/SpacetimeDB). License text: "Business Source License 1.1 ... Additional Use Grant: You may make use of the Licensed Work provided your application or service uses the Licensed Work with no more than one SpacetimeDB instance in production and provided that you do not use the Licensed Work for a Database Service" — [Source](https://github.com/clockworklabs/SpacetimeDB/blob/master/LICENSE.txt). Verdict: **adopt design, rebuild** (not OSI open source; the one-instance grant blocks multi-region worlds).+- **Mozilla Hubs successors**: Hubs-Foundation/hubs 2,218 stars, MPL-2.0, last push 2026-08-23, latest tag prod-2026-03-11 — [Source](https://github.com/Hubs-Foundation/hubs); Hubs-Foundation/reticulum 460 stars, MPL-2.0 — [Source](https://github.com/Hubs-Foundation/reticulum); Hubs-Foundation/Spoke 679 stars, MPL-2.0 with `src/editor/` derived from the MIT three.js editor — [Source](https://github.com/Hubs-Foundation/Spoke/blob/master/LICENSE). Mozilla ended support 2024-05-31; the Community Edition continues as a Kubernetes-deployable community project — [Source](https://ryanschultz.com/2024/02/26/mozilla-ceases-support-for-mozilla-hubs-but-the-open-source-hubs-community-edition-continues/). Verdict: **adopt design, rebuild** (MPL file-level copyleft is workable but the A-Frame stack is dated; Spoke proves a three.js-editor-derived world editor can ship).+- **Third Room** (matrix-org/thirdroom): 648 stars, Apache-2.0, **last push 2023-08-30** — [Source](https://github.com/matrix-org/thirdroom). Verdict: dormant; **adopt code** possible under Apache-2.0 for its Matrix-based room/identity design, but no maintenance.+- **iR Engine / Ethereal Engine** (ir-engine/ir-engine): 104 stars, **CPAL-1.0** (Common Public Attribution License, API NOASSERTION), last push 2025-07-17; the old etherealengine repo (704 stars) is archived — [Source](https://github.com/ir-engine/ir-engine/blob/dev/LICENSE). Verdict: **adopt design, rebuild** (CPAL's attribution-display clause and copyleft are incompatible with quietly folding code into an Apache-2.0 product).+- **Webaverse** (webaverse/app): 376 stars, MIT, last push 2022-12-20 — [Source](https://github.com/webaverse/app). Verdict: dead; nothing current to adopt.+- **Hytopia** (hytopiagg/sdk): 98 stars, **HYTOPIA Limited Use License** (API NOASSERTION), last push 2026-03-02; npm `hytopia` 0.15.2; the license permits "Internally incorporating the Licensed Software into your larger applications for deployment through the HYTOPIA platform" and grants Hytopia an unrestricted license to Feedback — [Source](https://github.com/hytopiagg/sdk/blob/main/LICENSE.md). The separate hytopiagg/hytopia-source engine repo is MIT, 63 stars, last push 2026-03-02 — [Source](https://github.com/hytopiagg/hytopia-source/blob/main/LICENSE.md). Verdict: **adopt design, rebuild** (platform-locked SDK; the MIT engine is small and six months idle).+- **Decentraland**: decentraland/unity-explorer 23 stars, Apache-2.0, last push 2026-10-08 (official Decentraland 2.0 desktop client); decentraland/godot-explorer 18 stars, Apache-2.0, last push 2026-10-08; unity-renderer archived — [Source](https://github.com/decentraland/unity-explorer); [Source](https://github.com/decentraland/godot-explorer). Verdict: **adopt design, rebuild** (not browser-native).++**AI-driven NPC and social simulation**++- **AI Town** (a16z-infra/ai-town): 10,604 stars, MIT, last push 2026-08-26 — [Source](https://github.com/a16z-infra/ai-town). Architecture under Key Question 3. Verdict: **adopt design, rebuild** (MIT allows code, but it is 2D PixiJS on Convex; the memory/conversation engine design is what transfers).+- **Stanford Generative Agents** (joonspk-research/generative_agents): 22,199 stars, Apache-2.0, last push 2024-08-05 — [Source](https://github.com/joonspk-research/generative_agents). **genagents** (StanfordHCI/genagents, the 1,000-persona follow-up): 617 stars, MIT, last push 2024-11-18 — [Source](https://github.com/StanfordHCI/genagents). Verdict: **adopt design, rebuild** (Apache-2.0 code is Python/Django; the architecture is the asset).+- **Project Sid / Altera** (altera-al/project-sid): 1,385 stars, **no license file**, last push 2024-11-04; the repo holds only the README, the paper PDF, a visual abstract and a demo video, no source — [Source](https://github.com/altera-al/project-sid). Verdict: **adopt design, rebuild** (paper only).+- **OASIS** (camel-ai/oasis): 5,246 stars, Apache-2.0, last push 2026-10-05, "Open Agent Social Interaction Simulations with One Million Agents" — [Source](https://github.com/camel-ai/oasis). **AgentSociety 2** (tsinghua-fib-lab/AgentSociety): 1,328 stars, Apache-2.0, last push 2026-10-01 — [Source](https://github.com/tsinghua-fib-lab/AgentSociety). **Voyager** (MineDojo/Voyager, skill-library embodied agent): 7,252 stars, MIT, last push 2024-04-03 — [Source](https://github.com/MineDojo/Voyager). Verdict: adopt design for scale-out and skill libraries.+- **2026 "agents in a 3D town" repos**: a GitHub search on 2026-10-09 for "ai town" found only 2D derivatives: zblcving9-gif/xianxia-AI-town-AI- 1,287 stars MIT (LLM-driven NPC decisions, "npc and players on equal footing"), mewamew/my_ai_town 902 stars no license, recursal/ai-town-rwkv-proxy 161, ykhli/cat-town 160 — [Source](https://github.com/zblcving9-gif/xianxia-AI-town-AI-). Searches for "generative agents 3d", "generative agents threejs", "generative agents unity", "llm npc unity" returned no repos. The only Godot hit was elefant-ai/player2-ai-npc-godot, 31 stars, GPL-3.0 — [Source](https://github.com/elefant-ai/player2-ai-npc-godot).+- **Convai / Inworld**: Convai offers Unity, Unreal and a Web SDK that integrates "real-time conversations, actions, and facial expressions"; Inworld's GitHub org lists Unity, Unreal and Node.js SDKs, and 2026 guides say Inworld "repositioned itself as B2B AI infrastructure" — [Source](https://github.com/topics/convai); [Source](https://www.github.com/inworld-ai); [Source](https://app.cinevva.com/guides/ai-npcs-dialogue). The exact repo names `Conv-AI/Convai-Web-SDK` and `inworld-ai/inworld-unity-sdk` returned 404 from the API on 2026-10-09, so star counts could not be verified. Verdict: **already covered** (both are thin clients to proprietary backends; three.ws runs its own agent brain and voice stack).+- **Unity / Unreal AI editor plugins**: CoplayDev/unity-mcp 14,778 stars, MIT, last push 2026-10-08 — [Source](https://github.com/CoplayDev/unity-mcp); chongdashu/unreal-mcp 2,091 stars, **no license file**, last push 2025-04-22 — [Source](https://github.com/chongdashu/unreal-mcp); Unity ML-Agents 19,727 stars, Apache-2.0 — [Source](https://github.com/Unity-Technologies/ml-agents/blob/develop/LICENSE.md). Verdict: **adopt design** (an editor-side MCP for Scene Studio mirroring PlayCanvas's and Unity's).++**World and level generation**++- **Matrix-Game** (SkyworkAI/Matrix-Game): 2,346 stars, MIT, last push 2026-09-29; "Matrix-Game 3.0: Real-Time and Streaming Interactive World Model with Long-Horizon Memory" released 2026-03-27; 2.0 on 2025-08-12, 1.0 on 2025-05-12 — [Source](https://github.com/SkyworkAI/Matrix-Game); license — [Source](https://github.com/SkyworkAI/Matrix-Game/blob/main/LICENSE). Third-party: 2.0 runs 25 fps with minute-long sessions — [Source](https://aireiter.com/blog/genie-3-alternative).+- **ABot-World** (amap-cvlab/ABot-World): 2,513 stars, Apache-2.0, last push 2026-10-09; ABot-World-0 (checkpoint `ABot-World-0-5B-LF`) streams 720p at 16 FPS with 1.2 s latency on a single RTX 5090 using 19 GB; "LongForcing" training for open-ended rollout; model, inference code and Gradio demo released 2026-07-09, tech report 2026-07-22; online playgrounds at abot-world.amap.com and a Hugging Face Space — [Source](https://github.com/amap-cvlab/ABot-World).+- **LingBot-World**: Robbyant/lingbot-world 4,513 stars, Apache-2.0, last push 2026-07-09 — [Source](https://github.com/Robbyant/lingbot-world); Robbyant/lingbot-world-v2 1,850 stars, **CC BY-NC-SA 4.0** (non-commercial), last push 2026-09-10 — [Source](https://github.com/Robbyant/lingbot-world-v2/blob/main/LICENSE.txt); a community fork claims the 1.3B v2 model runs at 16 FPS on an RTX 5090 — [Source](https://github.com/kaarelkaarelson/lingbot-world-v2-realtime).+- **Tencent HY-World**: HY-World-2.0 2,705 stars, Tencent HY-World 2.0 Community License dated 2026-04-15, which "DOES NOT APPLY IN THE EUROPEAN UNION, UNITED KINGDOM AND SOUTH KOREA" and carries an Acceptable Use Policy — [Source](https://github.com/Tencent-Hunyuan/HY-World-2.0/blob/main/License.txt); it outputs native 3D assets — [Source](https://worldsimulator.ai/blog/articles/best-ai-world-models). HY-WorldPlay (HY-World 1.5, real-time interactive) 1,611 stars, same license family — [Source](https://github.com/Tencent-Hunyuan/HY-WorldPlay); HunyuanWorld-1.0 2,943 stars, same territory exclusion — [Source](https://github.com/Tencent-Hunyuan/HunyuanWorld-1.0/blob/main/LICENSE).+- **Other 2026 open world models**: mira-wm/mira "Multiplayer Interactive World Models with Representation Autoencoders" 570 stars, Apache-2.0, last push 2026-09-16 — [Source](https://github.com/mira-wm/mira); shengshu-ai/minWM minimal real-time interactive world-model framework 850 stars, Apache-2.0 — [Source](https://github.com/shengshu-ai/minWM); AlayaLab/Evoke "Endless Interactive World with Bounded State" 849 stars, Apache-2.0 — [Source](https://github.com/AlayaLab/Evoke); etched-ai/open-oasis 2,139 stars, MIT, last push 2024-11-08 — [Source](https://github.com/etched-ai/open-oasis); a curated survey repo — [Source](https://github.com/liujiuming123/Awesome-Interactive-World-Model). Verdict for the class: **adopt code** (MIT/Apache) as a GPU-worker lane, never in-browser.+- **Procedural terrain/city generators**: the largest three.js-specific hits are tiny: jeromeetienne/threex.proceduralcity 146 stars MIT, photonlines/Procedural-City-Generator 47, MHillier98 city generator 48; Unity: Szuszi/CityGenerator-Unity 89, oddmax/unity-wave-function-collapse-3d 232; infinite terrain: mesqme/infinite-terrain 210 MIT, SirNeirda/godot_procedural_infinite_world 73 — [Source](https://github.com/jeromeetienne/threex.proceduralcity); [Source](https://github.com/oddmax/unity-wave-function-collapse-3d); [Source](https://github.com/mesqme/infinite-terrain). Verdict: **already covered** (Diorama); no open repo here has traction that beats a bespoke generator.++**Gaussian splat and NeRF viewers**++- **Inria reference 3DGS** (graphdeco-inria/gaussian-splatting): 24,145 stars, custom non-commercial license: "Inria and the Max Planck Institut for Informatik (MPII) hold all the ownership rights"; the Software "may be used 'non-commercially', i.e., for research and/or evaluation purposes only"; commercial use requires "prior and explicit consent of Licensors" via [email protected]; `utils/loss_utils.py` is MIT — [Source](https://github.com/graphdeco-inria/gaussian-splatting/blob/main/LICENSE.md). The license text does not say whether it covers outputs (trained splats) or the method.+- **Spark** (sparkjsdev/spark): 3,705 stars, MIT, last push 2026-10-07; v2.3.1 on 2026-10-01; "Built by World Labs"; formats PLY (incl. compressed), SPZ, SPLAT, KSPLAT, SOG; multiple splat objects with sorting and multiple viewpoints; per-splat transforms for animation incl. skeletal animation; GPU shader graph; "targeting 98%+ WebGL2 support"; the README does not mention WebGPU, SH or LOD/streaming — [Source](https://github.com/sparkjsdev/spark). Three.ws already depends on `@sparkjsdev/spark ^2.2.0`. Verdict: **already covered** (bump to 2.3.x).+- **gsplat** (nerfstudio-project/gsplat): 5,762 stars, Apache-2.0, v1.5.3 on 2025-07-04 (CUDA trainer, not a viewer) — [Source](https://github.com/nerfstudio-project/gsplat); nerfstudio 12,055 stars, Apache-2.0 — [Source](https://github.com/nerfstudio-project/nerfstudio). Verdict: **adopt code** on the GPU fleet for a commercially clean training lane.+- **luma-web** (`@lumaai/luma-web` on npm): MIT, version 0.2.2, last modified 2026-05-05, "includes `LumaSplatsWebGL`, which is a WebGL-only gaussian splatting implementation" — [Source](https://www.npmjs.com/package/@lumaai/luma-web) (read via `npm view`; the web page returned 403). The GitHub repo `lumalabs/luma-web` returned 404.+- **GaussianSplats3D** (mkkellogg): 2,905 stars, MIT, last push 2025-10-19; a 2026 guide says it "is no longer in active development" — [Source](https://github.com/mkkellogg/GaussianSplats3D); [Source](https://www.utsubo.com/blog/gaussian-splatting-guide). **antimatter15/splat**: 3,071 stars, MIT, last push 2025-11-16 — [Source](https://github.com/antimatter15/splat).++**WebXR / AR**++- **model-viewer** (google/model-viewer): 8,266 stars, Apache-2.0, v4.3.1 on 2026-06-04; wires AR Quick Look (USDZ) on iOS and WebXR/Scene Viewer on Android — [Source](https://github.com/google/model-viewer). Verdict: **already covered** (three.ws has "View in your space" exports) but the AR routing table is worth cross-checking.+- **Apple `<model>` element**: enabled by default in visionOS 26 with a new API, declarative `autoplay`/`loop`, USDZ assets — [Source](https://webkit.org/blog/17118/a-step-into-the-spatial-web-the-html-model-element-in-apple-vision-pro/); a WWDC26 session says the element "comes to iOS, iPadOS, and macOS" and presents it as a native alternative to model-viewer — [Source](https://developer.apple.com/videos/play/wwdc2026/215/); the W3C incubation repo immersive-web/model-element has 86 stars — [Source](https://github.com/immersive-web/model-element). WebXR immersive-vr is on by default in visionOS Safari since visionOS 2 / Safari 18, immersive-ar is listed but non-functional, and Safari on iOS/iPadOS/macOS still has no WebXR — [Source](https://www.testmuai.com/learning-hub/webxr-compatible-browsers/); [Source](https://xrdoctors.pro/blog/webxr-on-ios-what-actually-works).+- **WebXR Layers** (immersive-web/layers): 96 stars, W3C Document License; Quest Browser has shipped Layers and Media Layers by default since version 16.1 (2021); a July 2026 secondary report says the W3C moved Layers to Working Draft on 2026-06-12; Meta notes other browsers need the polyfill — [Source](https://github.com/immersive-web/layers); [Source](https://developers.meta.com/horizon/blog/achieve-better-rendering-and-performance-with-webxr-layers-in-oculus-browser/); [Source](https://vr.org/articles/aframe-1-8-0-webxr-open-source-june-2026).+- **8th Wall alternatives**: AR-js-org/AR.js 5,997 stars, MIT, last push 2026-06-21 — [Source](https://github.com/AR-js-org/AR.js); hiukim/mind-ar-js 2,744 stars, MIT, last push 2024-06-11 (two years idle) — [Source](https://github.com/hiukim/mind-ar-js); Needle positions itself to 8th Wall users — [Source](https://engine.needle.tools/docs/tutorials/for-8thwall-users.html). Verdict: **adopt code** (MIT) if AR Studio needs markerless image tracking without a vendor.++### Inferences+- The star-weighted center of gravity in web 3D is still three.js (116k) with Babylon (26k) and PlayCanvas (17k) as the two engines with first-party editors; three.ws made the right base choice but is now the only one of the three without a WebGPU-first renderer in production (178 vs 9 renderer references).+- The metaverse-framework generation that competed with three.ws's worlds (Webaverse, Third Room, Ethereal/iR, Hubs) is dead or in maintenance; the living competitors are narrower: Hyperfy (small, GPL), Hytopia (platform-locked), and Decentraland (native clients). The field is open for an Apache-2.0 browser world platform.+- The real 2026 competitive pressure is not another world platform but (a) PlayCanvas's splat-era tooling (SuperSplat, LOD budgets, GPU sort, publishing API) and (b) open interactive world models that run on one consumer GPU, both of which three.ws can adopt under MIT/Apache.++### Gaps+- Star counts for Convai and Inworld SDK repos could not be verified (exact repo slugs 404ed).+- The GitHub API returned no LICENSE for pmndrs/triplex, altera-al/project-sid, chongdashu/unreal-mcp and needle-engine-support; treat all four as "all rights reserved" until a license file appears.+- Hytopia's full limited-use license was only sampled by keyword; a lawyer-grade read is still needed before any design borrowing that could be construed as derivative.++## Key Question 1: What do the best web 3D editors and engines offer in 2026 that a vendored three.js editor does not (WebGPU, visual scripting, collaborative editing, asset pipelines, LOD/streaming, instancing, GI, post-processing)?++### Takeaway+By October 2026 three.js itself has shipped most of the renderer-level features (WebGPU default with WebGL2 fallback, TSL node materials on both renderers, a `RenderPipeline` post-processing stack, clustered lighting, cascaded `SunLight` shadows, `BatchedMesh` per-instance opacity), so the gap is not the engine but the editor: PlayCanvas and Babylon offer realtime collaboration with presence, branches/checkpoints, node-based material/particle/render-graph editors, an editor MCP server, and splat LOD budgets, none of which the stock three.js editor has.++### Cited Findings+- three.js r186 (`[email protected]`, 2026-09-08 per the guide; GitHub release 2026-09-24): WebGPU is enabled by default in Chrome, Edge, Safari 26 and Firefox and `WebGPURenderer` falls back to WebGL 2 automatically; r184 brought TSL and node materials to `WebGLRenderer` via a compatibility layer; r183 renamed `PostProcessing` to `RenderPipeline`; r185 added WebXR on `WebGPURenderer` and clustered (Forward+) lighting; r186 added `SunLight` with cascaded shadows for both renderers and `compileComputeAsync()`; r183 added `BatchedMesh` per-instance opacity; `Clock` deprecated for `Timer`; `PCFSoftShadowMap` removed in r186; minified builds removed and CommonJS deprecated — [Source](https://www.utsubo.com/blog/threejs-2026-what-changed); release tag — [Source](https://github.com/mrdoob/three.js/releases/tag/r186). r180 (2025-09-03) added BC4/BC5 compression and manual mipmaps to the WebGPU backend and a procedural-wood TSL material — [Source](https://github.com/mrdoob/three.js/releases/tag/r180).+- Babylon.js 8.0: all core shaders available in both GLSL and WGSL, Node Material Editor outputs WGSL, Node Render Graph (alpha), GPU mesh picking, Smart Filters node editor, area lights, Node Geometry (lattice deformer, subdivision) — [Source](https://blogs.windows.com/windowsdeveloper/2025/03/27/announcing-babylon-js-8-0/). Babylon.js 9.0 (2026-03-26): Clustered Lighting for hundreds of lights, Frame Graph, Node Particle Editor with Flow Maps, OpenPBR alpha, SDF text, splat shadows and Triangle Splatting, new inspector — [Source](https://blogs.windows.com/windowsdeveloper/2026/03/26/announcing-babylon-js-9-0); [Source](https://blogs.windows.com/windowsdeveloper/2026/04/02/part-3-babylon-js-9-0-openpbr-and-additional-engine-updates/).+- PlayCanvas Editor: realtime collaboration with a presence bar, collaborative code editing (Scripts 2.0), version control with checkpoints and branches, hot reload, a built-in MCP server for AI agents, Joint component, parallax occlusion, per-cascade shadow control, splat LOD controls (Editor 2.32, 2026-09-22) — [Source](https://developer.playcanvas.com/user-manual/editor/realtime-collaboration); [Source](https://github.com/playcanvas/editor); [Source](https://blog.playcanvas.com/). The editor's scripting is code-based; no visual scripting was found in 2026 posts — [Source](https://blog.playcanvas.com/).+- PlayCanvas engine LOD/streaming: a global splat budget auto-distributed by distance and screen coverage (v2.17), GPU compute sorting with WebGL2 fallback (v2.19), error-driven then distance-only LOD with target/limit budgets and sort-free stochastic rendering on WebGPU (v2.22 to v2.23) — [Source](https://radiancefields.com/playcanvas-engine-v2.23.0-returns-gaussian-splat-lod-to-distance-only-selection); [Source](https://app.cinevva.com/news/2026-03-06-playcanvas-v217-webgpu-splatting).+- Needle's asset pipeline: automatic glTF and texture compression on export from Unity/Blender, built-in networking and Rapier — [Source](https://engine.needle.tools/docs/explanation/core-concepts/features-overview).+- Bevy 0.20 ships Solari path tracing (DLSS-RR 4.5, Metal) but still no editor; its Feathers widgets are "components of a future scene editor" — [Source](https://bevy.org/news/bevy-0-20/). Godot's web export has no WebGPU or Forward+ — [Source](https://docs.godotengine.org/en/4.7/tutorials/export/exporting_for_web.html).+- The three.js editor has no native collaboration; a community clone added it with Ably in ~150 lines and offers "a subset of features"; node-based visual programming is only a planned community feature — [Source](https://discourse.threejs.org/t/three-js-editor-collaborative-in-150-lines-of-javascript/91900); [Source](https://app.cinevva.com/guides/playcanvas-vs-threejs).+- Hubs' Spoke editor is a three.js-editor derivative under MPL-2.0 that shipped as a full world-composition tool — [Source](https://github.com/Hubs-Foundation/Spoke/blob/master/LICENSE).++### Inferences+- Upgrading three.ws from `three ^0.184` to r186 and flipping Scene Studio, `/play`, and the walkaround world to `WebGPURenderer` would close the engine-level gap with Babylon/PlayCanvas at near-zero license cost; the TSL-on-WebGL compatibility layer (r184) means one material graph can serve both renderers during the transition.+- The editor-level gap ranks, by user-visible impact: (1) realtime co-editing with presence, (2) checkpoints/branches on scenes, (3) an editor MCP server so agents can build scenes (three.ws already has Spatial MCP; pointing it at Scene Studio mirrors PlayCanvas and Unity MCP, 14.8k stars), (4) a node material editor, (5) splat LOD budgets. All are "adopt design, rebuild" because PlayCanvas's backend is proprietary.+- A `Joint` component in the inspector (PlayCanvas 2.32) is a cheap win on top of three.ws's existing Rapier integration.++### Gaps+- No primary three.js source was fetched for the r181 to r186 feature list; the per-release detail comes from a well-regarded third-party guide and should be spot-checked against the r186 migration guide before quoting externally.+- Babylon's WebGPU-by-default status (vs. opt-in) in 9.x was not confirmed.++## Key Question 2: What networking and simulation architectures do leading multiplayer repos use, and how do they scale?++### Takeaway+Three architectures dominate in 2026: room-per-process authoritative servers (Colyseus, which added client prediction, reconciliation and rewind-based lag compensation in 0.18), match-handlers-as-goroutines on one node with Enterprise-only clustering (Nakama, with a third-party 2M CCU test), and database-as-server with a single-threaded actor per database scaled by sharding into child databases (SpacetimeDB, ~300k TPS replicated, BitCraft as proof). Deterministic client-side replication (Croquet) survives only as a niche. Dedicated-server orchestration vendors are consolidating (Hathora shut down 2026-05-05).++### Cited Findings+- Colyseus 0.18 (2026-08-25): "server-authoritative simulation with built-in prediction, reconciliation, and rewind"; `defineInput()` typed per-client buffered input; `setFixedTimestep()` fixed-rate loop that advertises the tick rate; `allowRewindState()` + `rewind.lastSeenBy()` for hit tests against the shooter's view; client `Predict` reconciler, `room.clock.serverNow()/renderNow()/rtt()/jitter()`; drift telemetry and a latency/jitter simulator; "Existing rooms are unaffected" — [Source](https://github.com/colyseus/colyseus/releases/tag/0.18). Colyseus "deliberately stays focused on the real-time state-sync layer", persistence is bring-your-own; Colyseus Cloud from $15/month — [Source](https://app.cinevva.com/guides/multiplayer-browser-game); [Source](https://tech-insider.org/colyseus-cloud-setup-2026/).+- Nakama: authoritative matches run "as lightweight goroutines (think green threads) inside the Nakama server"; match state lives on one node that becomes host; "any single Nakama node capable of running thousands of matches depending on hardware and player count"; matches are isolated and keep ticking when empty; runtimes are Lua, Go and TypeScript; any PostgreSQL-wire database — [Source](https://heroiclabs.com/docs/nakama/concepts/multiplayer/authoritative/); [Source](https://heroiclabs.com/docs/nakama/getting-started/architecture/). Multi-node presence replication and message routing are Enterprise-only ("you can mesh nakama nodes together ... using ENTRPRISE EDITION") — [Source](https://forum.heroiclabs.com/t/zoning-using-authoritive-servers-nakama-scaling-questions-and-architecture/2025); [Source](https://forum.heroiclabs.com/t/how-can-i-scale-nakama-open-source/1832). Code Wizards' AWS load test reached 2M CCU — [Source](https://heroiclabs.com/blog/code-wizards-scale-test-of-nakama-2m-ccu/).+- SpacetimeDB: "Each SpacetimeDB database currently runs as a single-threaded actor"; replicated databases reached roughly 300k TPS on benchmark transactions; all table data lives in memory on the leader so data size is capped by RAM; scale by sharding across child databases with a root database tracking them (BitCraft uses a global root plus region databases); roadmap: async inter-database calls and tiered disk/object storage both dated 2026-10-31 ("Spacetime Continuum"), read replicas, inter-database transactions and intra-database partitioning undated; Standalone is on GitHub, Cloud is proprietary and clustered — [Source](https://spacetimedb.com/blog/how-does-spacetime-scale); [Source](https://runtimewire.com/article/spacetimedb-october-31-scaling-plan-single-thread). BitCraft open-sourced its server modules Phase 1 on 2026-01-21, claiming global plus region modules "allow BitCraft to scale its worldsize indefinitely" — [Source](https://bitcraftonline.com/news/bitcraft-open-sourcing-update). A 2026 comparison calls its scaling "the least tested of the three in public benchmarks" — [Source](https://tech-insider.org/spacetimedb-vs-playfab-vs-nakama-2026/). License: BSL 1.1 with a one-production-instance, no-database-service grant — [Source](https://github.com/clockworklabs/SpacetimeDB/blob/master/LICENSE.txt).+- Croquet: multiplayer code "is executed on each client in a synchronized virtual machine, rather than on a server", a reflector orders messages; the design was presented as "moves all computation to the clients" with a lightweight reflector in 2020 — [Source](https://github.com/croquet/croquet); [Source](https://conf.researchr.org/details/dls-2020/dls-2020-papers/10/Croquet-A-Unique-Collaboration-Architecture-Keynote-). The Multisynq client repo is archived and sessions require an API key on Multisynq's DePIN network — [Source](https://github.com/multisynq/multisynq-client); [Source](https://unpkg.com/@multisynq/[email protected]/README.md).+- Hyperfy: self-hosted Node server per world with PhysX and a JS app system; v0.16.0 added remote databases and `player.push(force)` physics events — [Source](https://github.com/hyperfy-xyz/hyperfy).+- Hathora: announced 2026-03-04, shut down 2026-05-05, team to Fireworks AI, customers routed to GameFabric (Agones SDK) — [Source](https://crux.supercraft.host/blog/hathora-shut-down-where-to-go-after-may-2026/); [Source](https://gameye.com/blog/hathora-migration-gameye-vs-gamefabric/).+- Hytopia's SDK is positioned for "massively multiplayer, block-based games using JavaScript or TypeScript" with a server-authoritative model compiled from the MIT engine source; still "alpha ... Expect breaking changes" — [Source](https://github.com/hytopiagg/sdk/); [Source](https://github.com/hytopiagg/hytopia-source).+- MIRA explores "Multiplayer Interactive World Models", i.e. neural world models with more than one player — [Source](https://github.com/mira-wm/mira).++### Inferences+- three.ws's Colyseus server already implements the authoritative model (docs contain 411 "reconcil" and 267 "authoritative" references); the cheapest, highest-leverage upgrade is Colyseus 0.18's first-party prediction/rewind, which would replace bespoke reconciliation in `/play` combat and vehicles with maintained code (MIT). The client pin `colyseus.js 0.16.22` means two major versions of migration.+- For world-scale growth, SpacetimeDB's root-plus-region sharding is the design to copy (not the code, given BSL): a root "world registry" room plus per-region Colyseus rooms with async handoff mirrors what BitCraft does.+- Nakama's value for three.ws is its feature catalog (leaderboards, parties, matchmaking, tournaments) rather than its scaling story, since its clustering is paywalled.++### Gaps+- No public per-room CCU or tick-rate benchmark for Colyseus 0.18 was found.+- BitCraft player counts and transactions/second are not published in the SpacetimeDB scaling post.++## Key Question 3: What NPC cognition architecture do the leading AI-town repos use, and what would an embodied-agent platform gain from it?++### Takeaway+Two lineages define the field: Stanford's generative-agent loop (natural-language memory stream, retrieval, periodic reflection into higher-level memories, daily planning and reacting, validated with 25 agents) and Altera's PIANO (ten concurrent stateless modules reading and writing a shared agent state at different time scales, with a Cognitive Controller behind an information bottleneck for coherence, validated at 50 to 500 agents in Minecraft with 1000+ attempted). AI Town is the deployable MIT implementation of the first lineage on Convex; no 2026 open repo with traction puts these agents in a 3D town.++### Cited Findings+- Generative Agents (Park et al., 2023): the architecture extends an LLM "to store a complete record of the agent's experiences using natural language, synthesize those memories over time into higher-level reflections, and retrieve them dynamically to plan behavior"; 25 agents in a Sims-inspired town; from one seeded intent ("wants to throw a Valentine's Day party") agents spread invitations, make acquaintances, ask each other out and arrive on time; ablations show "observation, planning, and reflection each contribute critically to the believability" — [Source](https://arxiv.org/abs/2304.03442). Repo: Apache-2.0, `reverie/backend_server` simulation plus a Django frontend, "One game step represents 10 seconds in the game", history files are memory records injected into the memory stream, OpenAI API only, last push 2024-08-05 — [Source](https://github.com/joonspk-research/generative_agents).+- Project Sid / PIANO (Altera, arXiv 2411.00114): modules run concurrently at different time scales as stateless functions reading and writing a shared Agent State; named modules are Memory ("stores and retrieves conversations, actions, and observations across timescales"), Action Awareness, Goal Generation, Social Awareness, Talking, Skill Execution, with the Cognitive Controller "solely responsible for making high-level deliberate decisions"; a bottleneck "reduces the amount of information presented to the Cognitive Controller" and its decision is broadcast to condition the talk modules for speech/action coherence; reflex modules use small non-LLM networks; experiments: 25 agents/30 min and 49 agents/4 h item-acquisition benchmarks (full PIANO agents averaged 17 unique items, top 30 to 40; 49 agents saturated at ~320 items), 50-agent societies, 30-agent villages for specialization, a 29-agent tax society (deposits fell from ~20% to ~9% after amendment to 5 to 10%), 500-agent cultural transmission, runs over 1000 agents "sporadically unresponsive"; social-perception correlations r = 0.606 to 0.901 — [Source](https://arxiv.org/html/2411.00114v1). The GitHub repo has no code and no license — [Source](https://github.com/altera-al/project-sid).+- AI Town: Convex is the "Game engine, database, and vector search" layer that "natively supports shared global state, transactions, and a simulation engine"; memory embeddings use the chosen embedding model with `NUM_MEMORIES_TO_SEARCH` in `constants.ts`; cron jobs in `convex/crons.ts` stop inactive worlds; rendering is 2D PixiJS with Tiled maps; LLMs default to Ollama (`llama3`, `mxbai-embed-large`) with OpenAI, Together and any OpenAI-compatible API supported; deployable via Docker Compose with self-hosted Convex or Fly.io — [Source](https://github.com/a16z-infra/ai-town).+- genagents (Stanford, 1,000 interview-grounded personas): MIT, 617 stars, last push 2024-11-18 — [Source](https://github.com/StanfordHCI/genagents). OASIS scales social simulation to "One Million Agents" (Apache-2.0, 5,246 stars, active 2026-10-05) — [Source](https://github.com/camel-ai/oasis). AgentSociety 2 is "an LLM-native agent simulation platform designed for social science" (Apache-2.0, active 2026-10-01) — [Source](https://github.com/tsinghua-fib-lab/AgentSociety). Voyager's skill library for open-ended embodied agents is MIT — [Source](https://github.com/MineDojo/Voyager).+- AI Metropolis proposes out-of-order execution to scale LLM multi-agent simulation — [Source](https://arxiv.org/pdf/2411.03519). No 2026 open repo combining generative agents with a 3D town surfaced in web or GitHub search; the active AI Town derivatives (xianxia-AI-town 1,287 stars, my_ai_town 902) are 2D — [Source](https://github.com/zblcving9-gif/xianxia-AI-town-AI-).+- Convai's pipeline covers "Speech Recognition, Language Understanding, and Generation, Text-to-Speech, Text-to-Action, and Lipsync"; Inworld has "repositioned itself as B2B AI infrastructure" — [Source](https://assetstore.unity.com/packages/package/9235479); [Source](https://app.cinevva.com/guides/ai-npcs-dialogue).++### Inferences+- three.ws's `/play` NPCs (traffic, vendors, quest-givers, mobs) and `@three-ws/agent-memory` sit on the "scripted plus chat" tier; adding the generative-agent loop (observation log, importance-scored retrieval, nightly reflection, daily plan with reactive re-planning) is the design to copy from Apache-2.0/MIT sources, and it maps naturally onto the existing Colyseus tick and the Convex-style cron pattern AI Town uses.+- PIANO's split between fast non-LLM reflex modules and a slow deliberate controller is the right fit for a real-time 3D world where LLM latency cannot block locomotion; the bottleneck idea also addresses the speech/action incoherence users notice first.+- Because no open 3D agent town has traction, a three.ws world where generative agents with wallets and skills live in a 3D town would be a first-of-kind public surface rather than a catch-up feature.++### Gaps+- The retrieval scoring weights (recency, importance, relevance) and reflection triggers were not read from the Generative Agents paper body, only from the abstract.+- The four unnamed PIANO modules (10 total) are not described in the fetched portion of the paper.+- No cost or latency numbers per agent-step were found for any of the repos.++## Key Question 4: What is the state of commercially usable Gaussian splatting on the web in 2026?++### Takeaway+Rendering splats on the web is commercially clean: Spark (MIT, World Labs), SuperSplat (MIT), Babylon 9 (Apache-2.0), PlayCanvas engine (MIT), luma-web (MIT) and antimatter15/splat (MIT) all permit commercial use. The licensing hazard is on the training side: the Inria/MPII reference implementation is research-and-evaluation only, so a commercial platform should train with gsplat/nerfstudio (Apache-2.0) or a vendor that has cleared rights. The 2026 feature frontier is WebGPU compute sorting, streamed LOD with a per-scene splat budget, SOG compression, splat shadows, collision generation and publishing APIs.++### Cited Findings+- Inria license: non-commercial "for research and/or evaluation purposes only"; commercial use requires prior explicit consent; contact [email protected] — [Source](https://github.com/graphdeco-inria/gaussian-splatting/blob/main/LICENSE.md).+- Spark: MIT, v2.3.1 (2026-10-01), 3,705 stars; PLY/SPZ/SPLAT/KSPLAT/SOG; multi-object sorting, per-splat editing and skeletal animation, GPU shader graph; WebGL2 for 98%+ device coverage; built by World Labs — [Source](https://github.com/sparkjsdev/spark); [Source](https://radiancefields.com/platforms/spark). One listing credits a different author, which conflicts with the World Labs attribution — [Source](https://www.utsubo.com/blog/gaussian-splatting-guide).+- SuperSplat: MIT, v3.5.2 (2026-10-05), 10,313 stars; Editor 3.0 rebuilt on WebGPU, streamed automatic LOD, collision generation, walk mode, 360° video, SPZ export, Publishing API with Varjo/XGRIDS/LichtFeld partners, creator licensing and attribution — [Source](https://github.com/playcanvas/supersplat); [Source](https://blog.playcanvas.com/).+- PlayCanvas engine v2.23.0: distance-only LOD with target/limit budgets, sort-free stochastic rendering on WebGPU, per-axis focal-length fix for stereo; v2.19.0 GPU compute sorting with WebGL2 fallback — [Source](https://radiancefields.com/playcanvas-engine-v2.23.0-returns-gaussian-splat-lod-to-distance-only-selection); [Source](https://cgworld.jp/flashnews/01-202606-SuperSplat.html).+- Babylon.js 9.0: splats cast shadows, SOG support, Triangle Splatting, global sorting across multiple splat assets; 8.0 cut VRAM to 32 bytes/splat with SPZ and compressed PLY — [Source](https://radiancefields.com/babylon.js-v9.0-3dgs-gets-shadows-sogs-and-triangle-splatting-support-announced); [Source](https://blogs.windows.com/windowsdeveloper/2025/03/27/announcing-babylon-js-8-0/).+- gsplat: Apache-2.0 CUDA rasterizer/trainer, v1.5.3 (2025-07-04), 5,762 stars — [Source](https://github.com/nerfstudio-project/gsplat); paper — [Source](https://arxiv.org/html/2409.06765v1). nerfstudio: Apache-2.0, 12,055 stars — [Source](https://github.com/nerfstudio-project/nerfstudio).+- luma-web: MIT, 0.2.2, modified 2026-05-05, WebGL-only `LumaSplatsWebGL` — [Source](https://www.npmjs.com/package/@lumaai/luma-web). GaussianSplats3D: MIT, last push 2025-10-19, reported no longer actively developed — [Source](https://github.com/mkkellogg/GaussianSplats3D); [Source](https://www.utsubo.com/blog/gaussian-splatting-guide). antimatter15/splat: MIT, 3,071 stars — [Source](https://github.com/antimatter15/splat).+- Apple's `<model>` element renders USDZ with stereoscopic depth on visionOS 26 and is expanding to iOS/iPadOS/macOS per WWDC26, but it is USDZ-based, not splat-based — [Source](https://webkit.org/blog/17118/a-step-into-the-spatial-web-the-html-model-element-in-apple-vision-pro/); [Source](https://developer.apple.com/videos/play/wwdc2026/215/).+- PlayCanvas's "Turning a Gaussian Splat Into a Videogame" tutorial adds physics colliders, baked light probes, a Recast navmesh, NPCs and an FPS loop to a splat scene, with the project public — [Source](https://blog.playcanvas.com/).++### Inferences+- three.ws already ships Spark (`^2.2.0`), so the renderer choice is validated; the missing pieces relative to PlayCanvas are streamed LOD with a splat budget (needed for photoreal `/play` districts on phones), SOG compression in the asset pipeline, and collision generation so avatars can walk on splat captures. All three exist under MIT in SuperSplat/engine and can be ported or vendored.+- Spark's README does not mention WebGPU; with three.js r186 defaulting to `WebGPURenderer`, three.ws should verify Spark 2.3.x on the WebGPU path (or its WebGL2 fallback) before flipping the renderer.+- A "selfie to splat avatar" lane trained on gsplat (Apache-2.0) is commercially clean; any lane that shells out to the Inria trainer is not.++### Gaps+- Whether Spark 2.3.x supports WebGPU or spherical harmonics was not confirmed from the README.+- No independent benchmark compares Spark, PlayCanvas and Babylon splat throughput on the same scene; the 5.7x PlayCanvas figure is single-source.+- Whether trained splats (outputs) produced by the Inria code are themselves restricted is not addressed in the license text.diff --git a/docs/research/oss-landscape-2026-10.md b/docs/research/oss-landscape-2026-10.mdnew file mode 100644index 000000000..915f4174b--- /dev/null+++ b/docs/research/oss-landscape-2026-10.md@@ -0,0 +1,74 @@+# Giant open-source repos win stars, not capability++**No open-source repo beats three.ws on its core combination (a skinned 3D body that is the agent's chat surface, wired to a Solana wallet, x402 commerce and a marketplace), but a dozen repos beat it badly on breadth, packaging and standards alignment, and almost all of that breadth is legally adoptable.** The star-count leaders are OpenClaw (**391k stars**, MIT) ([repo](https://github.com/openclaw/openclaw)) and Hermes Agent (**252k**, MIT) ([repo](https://github.com/NousResearch/hermes-agent)), and both are chat-assistant harnesses, not embodied-agent platforms. The real gaps are concrete: channel breadth (three.ws has Telegram and Discord only), a hook/event bus and headless runner, AG-UI and A2A interoperability, x402 v2 and MPP session features, a Solana agent registry, temporal memory, workflow suspend/resume, non-humanoid rigging, and distribution on the MCP, Anthropic and skills.sh registries. Most of the best sources are MIT or Apache-2.0 and can be vendored. The exceptions that need a design-only rebuild are the AGPL, GPL, non-commercial, branding-clause and source-available projects (Open WebUI, Cherry Studio CE, SillyTavern, Skyvern, Dify, n8n, SpacetimeDB, Hyperfy, the Inria trainer and others). The single urgent item is a compliance risk, not a feature: three.ws's live Hunyuan3D lanes sit under a Tencent license that excludes the EU, UK and South Korea and forbids displaying outputs outside its territory, and nobody has checked whether a geo-gate exists. Build order: fix that gate, then ship channels, AG-UI, hooks and the x402/registry upgrades, because each is small to medium and uses permissive code.++## Where three.ws leads and where "beats" only means traction++Read carefully, the leaderboard is mostly a packaging contest. OpenClaw and Hermes are multi-channel personal-assistant harnesses with skill stores, and the big frameworks (Mastra, LangGraph, CrewAI, Agno, Mem0) are libraries, not products with a body, a wallet and a market. In the chat-UI and companion survey, no repo combines a skinned 3D body as the chat surface with a wallet-backed agent economy, and the nearest, AIRI (**~50k stars**, MIT, VRM and Live2D), still lists its memory layer as work in progress ([AIRI](https://github.com/moeru-ai/airi)). In worlds, no 2026 open repo with traction puts generative agents in a 3D town; the active derivatives are 2D ([AI Town](https://github.com/a16z-infra/ai-town)). Hyperfy (**305 stars**, GPL-3.0) and Hytopia (platform-locked SDK) are the living metaverse rivals, and both are small ([Hyperfy](https://github.com/hyperfy-xyz/hyperfy), [Hytopia](https://github.com/hytopiagg/sdk/blob/main/LICENSE.md)). In sign language, GitHub's open ASL repos are hobby-scale, so three.ws's recognizer is ahead of them ([sign.mt](https://github.com/sign/translate), [Skunk0 CTC model](https://huggingface.co/Skunk0/openglasses-fingerspelling-ctc)). On the screen-control safety model, three.ws is stricter than every open computer-use stack reviewed (no wallet or keys in the driven browser, 15-second capability leases, SSRF guard); none of the open stacks separates custody from the driven browser ([Agent S](https://github.com/simular-ai/Agent-S), [Cua](https://github.com/trycua/cua)). Add the integrated Forge-to-rig-to-animation pipeline, universal humanoid retargeting, x402 plus mandates plus escrow on Solana, and on-chain skill licenses (a stronger provenance story than any registry reviewed; ClawHub is the only one advertising signed manifests ([ClawHub](https://clawhub.ai/))), and the lead is real.++Where it genuinely trails, the lead is in infrastructure the repos treat as table stakes. OpenClaw reaches **Discord, iMessage, Slack, Teams, Telegram, WhatsApp, Signal and 20+ more channels** ([README](https://raw.githubusercontent.com/openclaw/openclaw/main/README.md)). OpenCode, Claude Code and Gemini CLI all expose lifecycle hooks, plugin manifests and headless JSON modes with per-run cost ([OpenCode plugins](https://opencode.ai/docs/plugins/), [Claude Code hooks](https://code.claude.com/docs/en/hooks), [headless](https://code.claude.com/docs/en/headless)). Memory systems publish **LoCoMo scores above 90%** (self-reported, on non-comparable harnesses) while `@three-ws/agent-memory` publishes none and keeps a deterministic co-occurrence graph rather than temporal fact invalidation ([Graphiti](https://github.com/getzep/graphiti), [Mem0](https://mem0.ai/research)). Commercial 3D rivals ship quad topology at generation, non-biped rigging and segmentation that three.ws lacks ([Meshy changelog](https://docs.meshy.ai/en/changelog), [Tripo docs](https://platform.tripo3d.ai/docs)). Finally, the engine layer: three.ws is on `three ^0.184` with 178 `WebGLRenderer` references against 9 for WebGPU, while r186 ships WebGPU by default with automatic fallback, and PlayCanvas ships streamed splat LOD budgets and an editor MCP server ([three.js r186](https://github.com/mrdoob/three.js/releases/tag/r186), [PlayCanvas editor](https://github.com/playcanvas/editor)).++## Licenses split the field into adopt, integrate and rebuild++The legal line falls where each source's code license and its weights license disagree, so the triage below tracks both. "Adopt" means a permissive code license and permissive weights (where models are involved) that allow commercial use inside an Apache-2.0 product with notices retained. A line needing a human legal read is flagged in the caveats section.++| Bucket | Projects | One-line reason |+|---|---|---|+| ADOPT | TRELLIS.2, PartCrafter, SkinTokens, MoMask, Direct3D-S2, Stable3DGen, Step1X-3D | MIT or Apache-2.0 code and weights ([TRELLIS.2](https://github.com/microsoft/TRELLIS.2), [PartCrafter](https://github.com/wgsxm/PartCrafter), [SkinTokens](https://github.com/VAST-AI-Research/SkinTokens)) |+| ADOPT | SAM 3D Objects and Body | SAM License is worldwide-commercial but carries attribution and Trade Control clauses ([LICENSE](https://github.com/facebookresearch/sam-3d-objects/blob/HEAD/LICENSE)) |+| ADOPT | Spark, SuperSplat, PlayCanvas engine, gsplat, glTF-Transform, meshoptimizer | MIT or Apache-2.0; gsplat is the commercially clean splat trainer ([Spark](https://github.com/sparkjsdev/spark), [SuperSplat](https://github.com/playcanvas/supersplat), [gsplat](https://github.com/nerfstudio-project/gsplat)) |+| ADOPT | OpenClaw channel adapters, Mastra core, Mem0, Langfuse core, A2A, AG-UI, assistant-ui, LibreChat patterns | MIT or Apache-2.0; exclude `ee/` directories in Mastra and Langfuse ([Mastra](https://raw.githubusercontent.com/mastra-ai/mastra/main/README.md), [AG-UI](https://github.com/ag-ui-protocol/ag-ui), [LibreChat](https://github.com/danny-avila/LibreChat)) |+| ADOPT | OpenCode, OpenHands, Codex CLI, Goose, Gemini CLI manifest, Browser Use, Stagehand, Steel Browser, Agent S, E2B | MIT or Apache-2.0 reference code for hooks, headless runs, browser and desktop agents, sandboxes ([OpenCode](https://github.com/anomalyco/opencode), [Steel](https://github.com/steel-dev/steel-browser), [E2B](https://github.com/e2b-dev/E2B)) |+| ADOPT | x402 v2, MPP spec, Verifiable Intent, Web Bot Auth, MCP Tasks, ERC-8004 and 8004-solana | Apache, MIT or CC0 standards and reference code ([x402](https://github.com/x402-foundation/x402), [MPP](https://github.com/tempoxyz/mpp-specs), [8004-solana](https://github.com/QuantuLabs/8004-solana)) |+| ADOPT | Pipecat, LiveKit Agents, Chatterbox, Kokoro, faster-whisper, wawa-lipsync, MuseTalk | BSD, Apache or MIT; verify each weight license before shipping |+| ADOPT | LeRobot, MentraOS, Wyoming, openWakeWord, Colyseus 0.18 | Apache-2.0 or MIT, direct integration targets ([LeRobot](https://github.com/huggingface/lerobot), [MentraOS](https://github.com/Mentra-Community/MentraOS), [Colyseus 0.18](https://github.com/colyseus/colyseus/releases/tag/0.18)) |+| INTEGRATE VIA API | Home Assistant, Frigate, Matter, openHAB | Home Assistant is Apache-2.0, but three.ws's value is the 3D agent on top, so deepen the existing relay and MCP integration instead of copying ([HA core](https://github.com/home-assistant/core)); openHAB is EPL-2.0 file-level copyleft, API-only ([openHAB](https://github.com/openhab/openhab-core)) |+| INTEGRATE VIA API | Langfuse, Opik, Helicone, Promptfoo, DeepEval, OpenLLMetry | Export traces to them, or run evals in CI; no need to host or fork ([OpenLLMetry](https://github.com/traceloop/openllmetry), [DeepEval](https://github.com/confident-ai/deepeval)) |+| INTEGRATE VIA API | Piper TTS successor, ESPHome firmware | GPL-3.0 or mixed GPL: run as a separate process only ([piper1-gpl](https://github.com/OHF-Voice/piper1-gpl)) |+| REBUILD FROM DESIGN | Hunyuan3D 2 and 2.1, HY-Motion, HY-World 2.0 | Tencent license excludes EU, UK and South Korea, forbids outputs outside the Territory, has a 1M MAU gate ([LICENSE](https://github.com/Tencent-Hunyuan/Hunyuan3D-2/blob/HEAD/LICENSE)); keep running only behind a geo-gate, and stop adding new ones |+| REBUILD FROM DESIGN | Open WebUI (post-v0.6.5), Dify, n8n, AutoGPT platform | Branding clause, multi-tenant ban, internal-use-only and PolyForm Shield competing-use bar ([Open WebUI license](https://docs.openwebui.com/license/), [Dify](https://raw.githubusercontent.com/langgenius/dify/main/LICENSE), [n8n](https://raw.githubusercontent.com/n8n-io/n8n/master/LICENSE.md), [AutoGPT](https://raw.githubusercontent.com/Significant-Gravitas/AutoGPT/master/LICENSE)) |+| REBUILD FROM DESIGN | Cherry Studio CE, SillyTavern, Skyvern, Smithery CLI, Daytona, RisuAI, sign.mt, 01 | AGPL, GPL or CC BY-NC-SA; copying forces copyleft on the combined work ([SillyTavern](https://github.com/SillyTavern/SillyTavern), [Skyvern](https://github.com/Skyvern-AI/skyvern), [sign.mt](https://github.com/sign/translate/blob/master/LICENSE.md)) |+| REBUILD FROM DESIGN | Hyperfy, iR Engine, SpacetimeDB, Hytopia SDK, Needle, Wonderland | GPL-3.0, CPAL-1.0, BSL 1.1 with one-instance grant, platform-locked, or proprietary ([SpacetimeDB license](https://github.com/clockworklabs/SpacetimeDB/blob/master/LICENSE.txt), [iR Engine](https://github.com/ir-engine/ir-engine/blob/dev/LICENSE)) |+| REBUILD FROM DESIGN | Inria 3DGS trainer, LingBot-World v2, Lyra 2.0 weights, Fish Speech, F5-TTS weights, VibeVoice | Non-commercial or research-only terms ([Inria](https://github.com/graphdeco-inria/gaussian-splatting/blob/main/LICENSE.md), [LingBot v2](https://github.com/Robbyant/lingbot-world-v2/blob/main/LICENSE.txt), [Lyra](https://github.com/nv-tlabs/lyra)) |+| REBUILD FROM DESIGN | AITuberKit, Open-LLM-VTuber, Phoenix (ELv2), Stability SF3D and SPAR3D, ERC-8183 SDKs | Non-commercial or custom license, ELv2 forbids managed service, Stability free only below US $1M revenue, Virtuals SDKs carry no license file ([Stability](https://github.com/Stability-AI/stable-fast-3d/blob/HEAD/LICENSE.md), [Phoenix](https://github.com/Arize-ai/phoenix)) |+| REBUILD FROM DESIGN | Hermes learning loop, LangGraph interrupts, CrewAI crews, Agno approvals, Letta memory blocks | Permissive licenses but Python-native and architecturally foreign; take the design into the TypeScript runtime ([Hermes](https://raw.githubusercontent.com/NousResearch/hermes-agent/main/README.md)) |++## Fifteen gaps ranked by leverage++The ranking weighs three things: user-visible impact for the Solana-first, embodied-agent positioning, effort, and whether permissive code exists. Effort sizes are S (days), M (weeks) and L (a quarter). Everything marked ADOPT uses permissively licensed code; REBUILD means the reference is a design source only, or is permissive but foreign to the TypeScript runtime.++| # | Gap and action | Effort | Impact | Inspiring repo | Mode |+|---|---|---|---|---|---|+| 1 | Geo-gate or replace the Hunyuan lanes, including gallery and marketplace display of their outputs | S | Removes a licensing exposure in the EU, UK and South Korea | [Tencent license](https://huggingface.co/tencent/Hunyuan3D-2.1/raw/main/LICENSE) | REBUILD (swap to TRELLIS.2 and Step1X-3D) |+| 2 | Channel breadth: Slack, WhatsApp, Signal, iMessage, Teams, Google Chat adapters, plus unknown-sender pairing, into the existing agent gateway | M | Largest reach gain per week of work | [OpenClaw](https://github.com/openclaw/openclaw) | ADOPT |+| 3 | New image-to-3D lanes: TRELLIS.2 (PBR, MIT) and SAM 3D Objects (real photo to mesh and layout), bake-off against Pixal3D | S to M | Forge quality and licensing safety at once | [TRELLIS.2](https://github.com/microsoft/TRELLIS.2), [SAM 3D](https://github.com/facebookresearch/sam-3d-objects) | ADOPT |+| 4 | Hook/event bus on the agent runtime, plus `three-ws run -p ... --output-format json` reporting spend in USDC and $THREE | M | Lets developers extend and automate agents; three.ws already meters x402 spend per call | [OpenCode](https://opencode.ai/docs/plugins/), [OpenHands headless](https://docs.openhands.dev/openhands/usage/run-openhands/headless-mode) | ADOPT |+| 5 | Emit and consume AG-UI events from the runtime and the embodiment layer | S to M | Interop with CopilotKit, assistant-ui and LangChain front ends; the Embodiment and Spatial MCP could become an AG-UI extension | [AG-UI](https://github.com/ag-ui-protocol/ag-ui) | ADOPT |+| 6 | Solana Agent Registry (register, feedback) alongside the existing Metaplex and memo reputation path | M | Native discoverability; three.ws's own docs name the missing Solana ERC-8004 equivalent | [8004-solana](https://github.com/QuantuLabs/8004-solana) | ADOPT |+| 7 | x402 v2 extensions (SIWx, sessions, hooks, dynamic payTo) and MPP charge, session and refund methods with the Solana SPL rail | M | Session deposits and in-protocol refunds are the missing commerce primitives | [x402](https://github.com/x402-foundation/x402), [MPP](https://github.com/tempoxyz/mpp-specs) | ADOPT |+| 8 | Discovery and identity: A2A signed Agent Card per agent, Web Bot Auth signed requests, publish hosted MCP servers to the official registry under a DNS-verified namespace and to Anthropic's directory, `npx skills add` listing | S | Distribution task, not engineering; three.ws already meets the remote Streamable HTTP plus OAuth DCR requirements | [A2A](https://github.com/a2aproject/A2A), [Web Bot Auth](https://github.com/cloudflare/web-bot-auth), [registry](https://github.com/modelcontextprotocol/registry), [directory docs](https://claude.com/docs/directory/publish) | ADOPT |+| 9 | Evaluator escrow with expiry refund as a Solana Anchor program, extending AgenC task escrow | L | Trust layer for paid agent jobs; no licensed reference exists | ERC-8183 spec (CC0); the Virtuals SDKs carry no license file | REBUILD |+| 10 | Temporal fact validity, provenance and contradiction handling in `@three-ws/agent-memory`, plus a published LoCoMo and LongMemEval score | M | The credibility gap versus every competing memory layer | [Graphiti](https://github.com/getzep/graphiti), [Supermemory](https://github.com/supermemoryai/supermemory) | REBUILD (Apache, but Python and Neo4j) |+| 11 | Workflow suspend/resume and general tool-approval human-in-the-loop on the durable runs engine, beyond the financial guard | M | Long-running agents and approval UX | [Mastra](https://raw.githubusercontent.com/mastra-ai/mastra/main/README.md), [assistant-ui](https://github.com/assistant-ui/assistant-ui) | ADOPT |+| 12 | Autonomous browser agent on the caster pool's Chromium, with the owner's lease as interrupt | M | Delivers the "agent drives" half that `docs/agent-screen-control.md` promises | [Browser Use](https://github.com/browser-use/browser-use), [Steel Browser](https://github.com/steel-dev/steel-browser) | ADOPT |+| 13 | Non-humanoid and any-mesh rigging with SkinTokens, plus part-level generation with PartCrafter | M | Closes the rigging gap against Meshy and Tripo | [SkinTokens](https://github.com/VAST-AI-Research/SkinTokens), [PartCrafter](https://github.com/wgsxm/PartCrafter) | ADOPT |+| 14 | Generative-agent NPC loop for `/play` (observation stream, importance-scored retrieval, reflection, daily plan) with PIANO-style fast reflex and slow deliberation split; upgrade to Colyseus 0.18 prediction and rewind | M | First-of-kind public 3D town of wallet-holding agents; no open 3D rival has traction | [AI Town](https://github.com/a16z-infra/ai-town), [Generative Agents](https://github.com/joonspk-research/generative_agents), [Project Sid paper](https://arxiv.org/html/2411.00114v1) | REBUILD (design), ADOPT (Colyseus) |+| 15 | Skill supply-chain scanner and a self-improving skill loop with full-text conversation search | M | ClawHub's 341 malicious skills show the exposure; a loop is the Hermes differentiator | [ClawHub incident](https://thehackernews.com/2026/02/researchers-find-341-malicious-clawhub.html), [Hermes](https://raw.githubusercontent.com/NousResearch/hermes-agent/main/README.md) | REBUILD |++The next tier is worth naming so it is not lost. Upgrade `three` to r186 and move Scene Studio, `/play` and the walkaround to `WebGPURenderer`, after confirming Spark 2.3.x on that path ([Spark](https://github.com/sparkjsdev/spark)). Port streamed splat LOD budgets, SOG compression and collision generation from the MIT PlayCanvas stack ([PlayCanvas engine](https://github.com/playcanvas/engine)). Point Spatial MCP at Scene Studio, mirroring PlayCanvas's editor MCP. Export OTel spans through OpenLLMetry and add multi-turn scoring from DeepEval to the session-to-test recorder ([DeepEval](https://github.com/confident-ai/deepeval)). Ship a MentraOS MiniApp for the nearest-agent cue on glasses ([MentraOS](https://github.com/Mentra-Community/MentraOS)). Add Character Card V2 import to the persona hub ([spec](https://github.com/malfoyslastname/character-card-spec-v2)). Add a voice turn-taking and full-duplex lane via Pipecat (BSD-2) or LiveKit Agents (Apache). Deepen the Home Assistant tool surface toward the broader ha-mcp category set through HA's own APIs ([ha-mcp](https://github.com/homeassistant-ai/ha-mcp)).++## Flag list: crypto projects among the findings++CLAUDE.md gates any commit that names a crypto project other than $THREE, so the safest pattern in committed docs, code and changelog entries is to cite standards by name (x402, MPP, ERC-8004, ERC-8183, A2A) and never the project or token behind a reference implementation. The projects below appear in the notes and need explicit owner approval before any commit mentions them. Solana itself and the platform's own $THREE work are the chain home and need no flag. The flagged set: Virtuals Protocol and its ACP SDKs (VIRTUAL token) and its game-node framework; elizaOS (token declared dead 2026-08-04, MIT code continues); SendAI and Solana Agent Kit (token reported by a Solana directory; Apache-2.0); Tempo (Stripe and Paradigm L1) and the MPP spec it hosts; Circle, Arc and USDC; Coinbase, Base and Coinbase Agentic Wallets and AgentKit; Dexter and PayAI (Solana facilitators); QuantuLabs and its 8004-solana registry (the one flagged item that is also a top-15 adoption target, so ask before the first commit); 8004scan, 8004market and SATI; Nevermined (token status unchecked, treat as flagged); Skyfire; Multisynq (token-gated DePIN network); MetaMask, Phantom, Privy, Crossmint and Turnkey agent wallets; Lightning and Bitcoin L402; and Decentraland. Hyperfy, Hytopia and Croquet are metaverse projects with no crypto tie in the notes, so they need no flag. The other-coin runtime exceptions in CLAUDE.md (user-supplied mint inputs, platform launch directories) do not apply to any of this reference material.++## Caveats the reader must weigh++Several claims could not be verified, and the conclusions above rest on the verified parts. Star counts come mostly from GitHub's API on 2026-10-09 in the coding, IRL and worlds notes, but the 3D, chat-UI and agent-framework figures are from WebFetch summaries and may be rounded. Benchmarks are the weakest evidence: every memory score (Mem0 92.5 LoCoMo, Zep 94.7, Supermemory, MemOS, Agent Zero) is vendor-reported on different harnesses and contested ([Mem0 vs Zep](https://mem0.ai/blog/zep-vs-mem0-which-ai-memory-layer-should-you-choose)); OSWorld-Verified figures such as Qwen3.8 Max at 86.1% come from aggregators, not the official board, which did not render ([BenchLM](https://benchlm.ai/benchmarks/osworld-verified)); and no independent 3D-generation leaderboard exists. x402 and Solana volume claims (Solana passing Base in daily transactions on 2026-08-25) rest on secondary sources, and several dates for the x402 Foundation conflict. Whether three.ws already has a Hunyuan geo-gate was not checked. Pixal3D's "academic only" blog claims conflict with its MIT LICENSE file, and Hunyuan3D-2.1's `hy3dpaint` carries a non-commercial header.++Items that need a human legal read before any code is copied: Tencent Hunyuan's territorial limits, 1M monthly-active-user gate (measured on each model's release date) and the ban on using outputs to improve other models; TEN Framework's "Apache 2.0 with additional restrictions" text, which the researcher did not read at the root, so treat it as non-adoptable until read; LobeChat, whose page shows both a LobeHub Community License and an Apache badge ([repo](https://github.com/lobehub/lobe-chat)); Mastra's `ee/LICENSE` and the PolyForm Shield clause, neither quoted; Inngest's SSPL conversion terms; weight licenses of Chatterbox, Kokoro, Orpheus, CosyVoice and the NVIDIA Audio2Face-3D "NVIDIA Open Model" text (relevant because three.ws is an NVIDIA Inception member); Gemma-derived weights for pi0 and SmolVLA; the Voice PE and ESPHome per-directory licenses; Hytopia's full limited-use license (only sampled by keyword); and the QuantuLabs program-id discrepancy between its repos. Four repos with no LICENSE file (Triplex, Project Sid, unreal-mcp, Needle's support repo) default to all rights reserved. Lastly, the Roo Code shutdown, the Skyfire shutdown claim, the Virtuals fee split and the Daytona AGPL status each rest on a single or secondary source, and Daytona's archive on 2026-10-03 is from the GitHub API only.++## Conclusion++The sharper lesson is that three.ws competes on a different axis than the repos that outrank it: they win on surface area and standards alignment, and every one of those gaps closes with MIT or Apache code in weeks, while the combination three.ws owns cannot be copied from any single repo. That reframes the roadmap: the first six items are integration work (channels, AG-UI, hooks, registries, x402 v2, Solana registry), not research, and the only L-sized item worth the quarter is evaluator escrow, because it has no licensed reference and is a Solana-native trust primitive nobody else offers.++Two second-order findings change priorities. First, the Hunyuan exposure means the most-used 3D lanes may be the least safely distributable, so item 1 is a prerequisite for any marketplace growth, not a nice-to-have. Second, three.ws's cleanest differentiators (wallet-free driven browsers, on-chain skill licenses, a 3D town of wallet-holding agents) are already ahead of every open rival, so documenting and exposing them (signed skill manifests with license pointers, a public screen-control safety spec) is cheaper than any catch-up feature and turns a hidden lead into a visible one.