Skip to content

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.

fix(research): replace banned dashes in the landscape notes and reword a dry-run usage comment so the push checks pass

nirholas2c56267e27e7
diff --git a/docs/research/oss-landscape-2026-10-notes/agent-commerce.md b/docs/research/oss-landscape-2026-10-notes/agent-commerce.md
index 067649eaa..0b181a4b6 100644
--- a/docs/research/oss-landscape-2026-10-notes/agent-commerce.md
+++ b/docs/research/oss-landscape-2026-10-notes/agent-commerce.md
@@ -96,29 +96,29 @@ Sources for the table rows: GitHub API reads on 2026-10-09 (stars, license, push
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/)
+- 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.
@@ -137,25 +137,25 @@ three.ws already has the delegated-budget core (signed Intent Mandate with per-c
### 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
+- 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/)
+- 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).
@@ -175,21 +175,21 @@ three.ws already has the delegated-budget core (signed Intent Mandate with per-c
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`
+- 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.
@@ -209,16 +209,16 @@ On Solana the only standard with institutional pull is the ERC-8004 port: the So
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))
+- 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).
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.md
index e7c81b7aa..350b5112a 100644
--- a/docs/research/oss-landscape-2026-10-notes/chat-memory-evals.md
+++ b/docs/research/oss-landscape-2026-10-notes/chat-memory-evals.md
@@ -8,15 +8,15 @@ Method note: star counts and license lines were read from each repo's GitHub pag
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.
+- **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
@@ -24,15 +24,15 @@ Open WebUI (154k) dominates on scale and enterprise features but its license now
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).
+- **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
@@ -40,14 +40,14 @@ Every serious entrant publishes >88% on LoCoMo/LongMemEval, but all numbers are
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).
+- **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
@@ -55,12 +55,12 @@ Langfuse, Opik, Helicone, DeepEval, Promptfoo, OpenLLMetry and Laminar are permi
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.
+- **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.
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.md
index ef70873f5..ebaee4e95 100644
--- a/docs/research/oss-landscape-2026-10-notes/irl-home-robotics.md
+++ b/docs/research/oss-landscape-2026-10-notes/irl-home-robotics.md
@@ -8,25 +8,25 @@ All star counts, licenses and push dates were read from the GitHub REST API (aut
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)
+- 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.
@@ -44,12 +44,12 @@ Do not rebuild any of the smart-home stack. Home Assistant (Apache-2.0, ~91k sta
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)
+- 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.
@@ -64,7 +64,7 @@ MentraOS (Apache-2.0, 2.4k stars) is the clear open glasses OS and is directly a
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)
+- 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.
@@ -78,13 +78,13 @@ openpilot (MIT, 63.9k stars) is the only large open in-car project, but it is dr
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)
+- 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.
@@ -98,13 +98,13 @@ LeRobot (Apache-2.0, 28k stars) is the hub: it bundles SO-101 teleop/training, R
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)
+- 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.
diff --git a/docs/research/oss-landscape-2026-10-notes/worlds-engines.md b/docs/research/oss-landscape-2026-10-notes/worlds-engines.md
index e6133abb0..f6bc99259 100644
--- a/docs/research/oss-landscape-2026-10-notes/worlds-engines.md
+++ b/docs/research/oss-landscape-2026-10-notes/worlds-engines.md
@@ -11,67 +11,67 @@ Of ~60 repos checked, the permissively licensed ones that three.ws can lift code
**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).
+- **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).
+- **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).
+- **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.
+- **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).
+- **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.
+- **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).
@@ -89,14 +89,14 @@ Of ~60 repos checked, the permissively licensed ones that three.ws can lift code
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).
+- 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.
@@ -113,14 +113,14 @@ By October 2026 three.js itself has shipped most of the renderer-level features
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).
+- 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.
@@ -137,12 +137,12 @@ Three architectures dominate in 2026: room-per-process authoritative servers (Co
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).
+- 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.
@@ -160,15 +160,15 @@ Two lineages define the field: Stanford's generative-agent loop (natural-languag
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/).
+- 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.
diff --git a/scripts/x-mention-make-dry-run.mjs b/scripts/x-mention-make-dry-run.mjs
index e74ee9695..1ebac42cf 100644
--- a/scripts/x-mention-make-dry-run.mjs
+++ b/scripts/x-mention-make-dry-run.mjs
@@ -12,7 +12,7 @@
// Rows land under account_ref `trythreews-evidence`, so they never appear in
// the real account's review console or rate limits.
//
-// USAGE (needs DATABASE_URL, e.g. from .env.local)
+// Usage (needs DATABASE_URL, e.g. from .env.local):
// node --env-file=.env.local scripts/x-mention-make-dry-run.mjs
// node --env-file=.env.local scripts/x-mention-make-dry-run.mjs --base https://three.ws --budget-ms 150000
// node --env-file=.env.local scripts/x-mention-make-dry-run.mjs --follow-up-ms 900000 # how long to keep running the follow-up tick on pending rows
diff --git a/tests/x-mention-make.test.js b/tests/x-mention-make.test.js
index 674d66c33..ebfef4bb0 100644
--- a/tests/x-mention-make.test.js
+++ b/tests/x-mention-make.test.js
@@ -81,9 +81,9 @@ describe('reply copy', () => {
for (const kind of ['success', 'failure']) {
const text = make.composeMakeReply(kind, { prompt: long, link: `${BASE}/forge?prompt=${long}` });
expect(weightedLength(text)).toBeLessThanOrEqual(280);
- expect(text).not.toMatch(/[–—]/);
+ expect(text).not.toMatch(/[\u2013\u2014]/);
}
- expect(make.composeMakeReply('unsafe')).not.toMatch(/[–—]/);
+ expect(make.composeMakeReply('unsafe')).not.toMatch(/[\u2013\u2014]/);
});
});
@@ -107,6 +107,13 @@ describe('success path', () => {
expect(bot.service_account).toBe(true);
});
+ it('uses the reference image as media when the model is too large to render', async () => {
+ const ev = await mentionRow();
+ const big = { ...doneJob, size_bytes: 12 * 1024 * 1024, preview_image_url: 'https://cdn.three.test/forge/dragon.png' };
+ const res = await make.handleMake(ev, { ...noDirector, startForge: async () => big });
+ expect(res.mediaUrl).toBe('https://cdn.three.test/forge/dragon.png');
+ });
+
it('hands a live row to the adapter and records the reply id once', async () => {
const ev = await mentionRow({ dryRun: false });
const deliver = vi.fn(async () => '2000000000000000001');