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.
Awesome 3D Agents 
Everything you need to give an AI agent a body.
A curated list of the models, engines, formats, and research that turn a language model into something you can see, animate, talk to, and put on a web page. Generation, rigging, motion, faces, voice, renderers, and the agent frameworks that drive them.
152 entries across 15 sections. Browsable, searchable, and filterable at three.ws/awesome. Every link is fetched and verified before it ships.
Contents
- Text and image to 3D (16)
- Photos and video to 3D (14)
- Avatars and characters (13)
- Rigging and retargeting (8)
- Motion and animation (10)
- Texturing and materials (5)
- Faces, lipsync, and voice (15)
- Rendering on the web (12)
- Formats and asset pipeline (12)
- AR and XR (6)
- Physics, networking, and shared worlds (8)
- Agent frameworks and protocols (13)
- Datasets and simulators (7)
- Free assets (6)
- Learning (7)
Text and image to 3D
Generate a mesh from a prompt or a single photo. This is the front door of the pipeline: the fastest way to get geometry that did not exist an hour ago.
- TRELLIS - Structured latent generation that emits meshes, radiance fields, and Gaussians from the same model. Currently the strongest open text and image to 3D quality.
- Hunyuan3D 2 - Shape generation plus a separate texture synthesis stage, so the mesh and its material are not fighting for the same capacity. Ships usable PBR maps.
- TripoSR - Single image to 3D in well under a second on one GPU. The reference point for fast feed-forward reconstruction.
- InstantMesh - Multi-view diffusion into a large reconstruction model, tuned for clean topology rather than the blobby output that plagues earlier feed-forward work.
- Stable Fast 3D - Image to UV-unwrapped, textured mesh in roughly half a second, with material parameter prediction built in.
- LGM - Large multi-view Gaussian model. High-resolution 3D from text or a single image, then converted to a mesh.
- Wonder3D - Cross-domain diffusion that generates consistent normal maps and colour views together, which is why its surfaces hold up under relighting.
- CRM - Convolutional reconstruction model. Single image to a textured mesh in about ten seconds without per-object optimisation.
- Zero-1-to-3 - The novel-view-synthesis paper most of this section is built on. Learn what your favourite generator is actually doing by reading this one.
- threestudio - One framework, many text-to-3D methods (DreamFusion, Magic3D, ProlificDreamer, and more) behind a shared config system. The right place to compare approaches honestly.
- Shap-E - Generates implicit functions that render as both meshes and neural radiance fields. Old now, but small, permissive, and easy to run.
- Point-E - Text to point cloud on a single GPU in a minute or two. Still the cheapest way to get rough 3D structure from a prompt.
- three.ws Forge - A free hosted text-to-3D lane with no key and no account. Returns a GLB you can download, embed, or rig.
- Meshy - Commercial text and image to 3D with quad remeshing and PBR texturing. Good API, generous enough free tier to evaluate properly.
- Tripo - The hosted product from the TripoSR authors. Fast previews, refined high-poly output, and a documented API.
- Rodin - Generation aimed at production assets: clean topology options, PBR materials, and part-level control.
Photos and video to 3D
Capture reality instead of inventing it. Photogrammetry, neural fields, and Gaussian splatting, plus the viewers that get the result onto the web.
- 3D Gaussian Splatting - The original INRIA implementation that made radiance fields real-time. Everything in this section is downstream of it.
- gsplat - A CUDA-accelerated splatting library with a clean Python API, faster and far less memory-hungry than the reference code.
- Nerfstudio - A modular pipeline for neural radiance fields and splats, with a real-time web viewer and a sane data-processing story.
- Spark - A Gaussian splatting renderer that drops into an existing three.js scene as ordinary objects, so splats and meshes light and sort together.
- SuperSplat - A browser-based editor for cleaning, cropping, and compressing splat captures before you publish them.
- Brush - A splatting engine in Rust and wgpu that trains and renders on the web, macOS, Linux, Windows, and Android from one codebase.
- OpenSplat - Splat training on CPU as well as CUDA and Metal, which makes it the practical option on machines without an NVIDIA card.
- COLMAP - Structure from motion and multi-view stereo. The camera poses almost every method above expects as input come from here.
- DUSt3R - Dense 3D reconstruction from unconstrained image pairs with no camera calibration at all. It quietly deleted a whole preprocessing stage.
- VGGT - A single feed-forward transformer that predicts camera poses, depth, and point maps together in seconds.
- Meshroom - A node-graph photogrammetry application on top of AliceVision. The friendliest free path from a folder of photos to a textured mesh.
- splat - A WebGL splat viewer in a few hundred readable lines. Read it once and you will understand the whole rendering technique.
- Luma AI - Phone capture to a shareable splat, with an embed and a web SDK.
- Polycam - LiDAR and photo capture on phones with mesh, splat, and glTF export. The fastest route from a real object to a file you can use.
Avatars and characters
Humanoid bodies that arrive already rigged, already skinned, and ready to be driven. Building one from scratch is rarely the right first move.
- Avaturn - Photorealistic avatars from a phone photo, exported as GLB or FBX with a standard humanoid skeleton.
- VRoid Studio - A free desktop editor for anime-style VRM characters, with hair, clothing, and expression authoring built in.
- three-vrm - The VRM runtime for three.js: spring bones, look-at, expressions, and material handling that a plain glTF loader will not give you.
- VRM specification - The glTF extension that standardises humanoid bones, expressions, look-at, and licence metadata. Read it before you invent your own avatar format.
- MakeHuman - An open parametric human generator with a full skeleton and topology you can actually edit. Old-fashioned, and still unmatched for free anatomically-driven bodies.
- SMPL-X - A parametric body model with articulated hands and an expressive face. The shared coordinate system that most human motion research speaks.
- Character Creator - Commercial character authoring with production-grade topology, morphs, and direct export to game engines.
- three.ws Avatar Studio - Generate, rig, dress, and pose an avatar in the browser, then embed it anywhere with one script tag.
- MetaHuman - Epic's photorealistic human creator, free to use, with a browser editor and export into Unreal or Maya. The quality ceiling for a digital human.
- VRM Add-on for Blender - Import, edit, and export VRM avatars in Blender, including spring bones and expressions. The missing link between an authoring tool and the avatar format.
- PIFuHD - Reconstructs a clothed 3D human from one photo at a resolution that captures fabric folds. The paper that made single-image humans plausible.
- ECON - Clothed human reconstruction that stays robust on loose garments and unusual poses, where implicit-only methods fall apart.
- Meshcapade - Commercial bodies and motion built on the SMPL family, so what you buy speaks the same parameter space as the research above.
Rigging and retargeting
A mesh without a skeleton is a statue. This is the layer that decides whether an avatar can be animated at all.
- Mixamo - Free automatic humanoid rigging plus a large motion-capture library. Its bone naming has become an informal industry convention.
- UniRig - An autoregressive model that rigs humans, animals, and objects from one checkpoint, predicting the skeleton and the skin weights together.
- RigNet - The neural rigging paper this field grew out of: joint prediction, skeleton connectivity, and skinning learned end to end.
- AccuRIG - A free desktop auto-rigger for static humanoid meshes, with clean weight painting and direct export to standard skeletons.
- Rigify - Blender's bundled meta-rig system. Free, scriptable, and the practical fallback when an automatic rigger gets a character wrong.
- SkeletonUtils - Retargeting helpers in three.js that map a clip authored for one skeleton onto another at runtime. Small, unglamorous, and load-bearing.
- Ossos - A TypeScript skeletal animation library with inverse kinematics and retargeting written to be read, not just imported.
- mixamo_converter - Batch-converts Mixamo clips to in-place root motion in Blender, which is the conversion every game and web project needs and Mixamo does not do.
Motion and animation
Generated motion, captured motion, and the datasets both are trained on. A body that only idles is a prop.
- Motion Diffusion Model - Text to human motion via a lightweight diffusion model. The reference implementation nearly every later method compares against.
- MoMask - Generative masked modelling of motion tokens. Faster and sharper than diffusion baselines, with editing and in-betweening for free.
- MotionGPT - Treats motion as a language, so one model generates, captions, predicts, and in-betweens with the same interface.
- T2M-GPT - VQ-VAE plus a transformer for text-driven motion. A clean, readable codebase to learn the tokenised-motion approach from.
- OmniControl - Adds spatial control signals to motion diffusion, so you can pin a hand to a location instead of hoping the prompt lands.
- HumanML3D - The text-and-motion dataset behind most of this section. If you are benchmarking, you are benchmarking on this.
- AMASS - Fifteen-plus motion capture datasets unified onto one SMPL body. The reason cross-dataset motion training is possible at all.
- Cascadeur - Keyframe animation software with physics-aware assistance and AutoPosing, usable free for individuals under a revenue threshold.
- Rokoko - Motion capture suits and a free video-to-motion tool, plus a Blender plugin that handles retargeting.
- Kalidokit - Turns MediaPipe face, hand, and pose landmarks into VRM rotations in the browser, so a webcam drives an avatar with no capture suit.
Texturing and materials
Geometry is half the asset. These are the tools that decide whether it looks like a surface or like grey clay.
- TEXTure - Text-guided texturing of an existing mesh through iterative depth-conditioned diffusion. The paper that started this category.
- Text2Tex - Generates high-resolution textures for a mesh from a prompt, with a view-selection pass that reduces the seams these methods usually leave.
- Paint3D - Produces lighting-free texture maps, so the result can be relit properly instead of baking one scene's shadows into the albedo.
- xatlas - Mesh parameterisation and UV atlas generation. A generated mesh with no UVs cannot be textured, and this is what fixes that.
- Material Maker - A free node-based procedural material authoring tool built on Godot, in the shape of Substance Designer.
Faces, lipsync, and voice
The difference between a model standing there and an agent talking to you. Speech in, speech out, and a mouth that agrees with the audio.
- TalkingHead - A three.js class that turns a glTF humanoid avatar into a talking head with visemes, gaze, and mood. The most complete open reference for this exact problem.
- Rhubarb Lip Sync - Offline audio to mouth-shape timings as a plain command line tool. No model to host, no API to pay for.
- MediaPipe - Face landmarks and blendshape coefficients in the browser at video rate, which is enough to drive an avatar's face from a webcam.
- ARKit blendshapes - The 52-coefficient facial vocabulary that avatar formats, capture tools, and generators have all standardised on.
- Whisper - Robust multilingual speech recognition that made self-hosted transcription the default rather than the exception.
- faster-whisper - A CTranslate2 reimplementation of Whisper that is several times faster at the same accuracy. Use this in production, not the reference code.
- whisper.cpp - Whisper in plain C and C++ with no dependencies, running on phones, browsers via WebAssembly, and Raspberry Pis.
- Kokoro - An 82M-parameter text to speech model with quality that embarrasses far larger ones, small enough to run in a browser tab.
- Piper - Fast local neural text to speech built for low-power devices, with a large set of pretrained voices.
- F5-TTS - Flow-matching text to speech with convincing zero-shot voice cloning from a few seconds of reference audio.
- Coqui TTS - A deep-learning toolkit covering dozens of text to speech and voice-conversion models, including XTTS multilingual cloning.
- Silero VAD - Voice activity detection in under a millisecond per chunk. The unglamorous component that decides whether a voice agent feels responsive.
- Pipecat - A Python framework for real-time voice and multimodal agents: pipelines, interruption handling, and transport out of the box.
- LiveKit Agents - Build voice agents that join a WebRTC room, with turn detection and telephony handled for you.
- Web Speech API - Recognition and synthesis already in the browser, with zero bytes shipped. Always check this before adding a model.
Rendering on the web
The runtime that actually puts the body on a page. Everything above is upstream of a choice made here.
- three.js - The default WebGL and WebGPU library for the web, with the deepest ecosystem of loaders, controls, and examples anywhere.
- React Three Fiber - A React renderer for three.js. Scene graphs become components, and the reconciler handles the mutation you would otherwise write by hand.
- drei - The helper library that makes React Three Fiber practical: controls, loaders, staging, text, and effects you would otherwise rebuild every project.
- Babylon.js - A batteries-included web engine with a built-in physics layer, node material editor, and an inspector that is genuinely better than the alternatives.
- PlayCanvas - A lightweight engine tuned hard for load time and mobile, backed by a collaborative browser editor.
- model-viewer - One custom element that renders a glTF with sensible defaults, accessibility, and AR on both phone platforms. The lowest-effort correct answer.
- A-Frame - Declarative HTML for WebXR scenes, with an entity-component system underneath. Still the shortest path from markup to headset.
- Threlte - Svelte components for three.js, with a component library, physics bindings, and a preprocessor that keeps bundles small.
- TresJS - Vue components for three.js, with reactivity mapped onto scene-graph updates.
- OGL - A minimal WebGL library that stays close to the API instead of hiding it. The right tool when three.js is more engine than you need.
- Needle Engine - Author in Unity or Blender, ship a small web build. A pragmatic bridge for teams whose content pipeline already exists.
- agent-3d - A web component that drops a rigged, animated, lip-syncing agent avatar onto any page in one tag, with a JS API for moods and gestures.
Formats and asset pipeline
Getting the asset small, valid, and loadable. Most 3D-on-the-web failures are pipeline failures, not rendering failures.
- glTF - The transmission format for 3D on the web, plus its extension registry. If you are choosing a format, this is the answer.
- glTF-Transform - Read, edit, optimise, and validate glTF from Node or the CLI. Deduplicate, resize textures, weld, and compress in a scripted pipeline.
- meshoptimizer - Vertex cache optimisation, simplification, and the meshopt compression that glTF files use. Its gltfpack CLI is the single highest-leverage step in most pipelines.
- Draco - Geometry compression that shrinks meshes and point clouds hard, with decoders for the browser.
- KTX-Software - Basis Universal texture transcoding. GPU-compressed textures that stay compressed in memory, which is usually the real budget problem.
- glTF-Validator - Tells you whether a file is actually valid before a user's browser does. Wire it into CI and stop guessing.
- glTF Sample Assets - The official test models, including deliberately awkward ones. Your loader is not done until these render correctly.
- Blender - Free, scriptable in Python, and headless-capable, which makes it the workhorse of automated 3D pipelines as much as an authoring tool.
- trimesh - The Python library for loading, repairing, measuring, and booleaning meshes. Nearly every 3D backend service has this in it somewhere.
- Assimp - Imports roughly fifty 3D formats into one data structure. The escape hatch when someone sends you a file nothing else opens.
- OpenUSD - Pixar's scene description for large, layered, collaboratively edited worlds. Heavier than glTF, and the right answer above a certain scale.
- three-mesh-bvh - Bounding volume hierarchies for three.js geometry, turning raycasts against dense meshes from unusable into instant.
AR and XR
Getting the body off the page and into the room. The web is the only platform where this needs no install.
- WebXR Device API - The specification behind every headset and phone AR session on the open web. Read the spec, not just a framework's wrapper.
- WebXR Samples - The working group's own runnable examples, from a bare session to hit testing and layers. Read these before any framework tutorial.
- Immersive Web Emulator - A browser extension that emulates a headset and controllers, so WebXR can be developed and debugged without putting a device on.
- AR Quick Look - Apple's system AR viewer. A USDZ behind a link opens full-screen AR on any iPhone with no app and no permission prompt.
- Scene Viewer - The Android equivalent, launched from an intent URL with a glTF. Together with AR Quick Look it is why model-viewer can offer AR everywhere.
- 8th Wall - Markerless AR that runs in the mobile browser rather than an app, with world tracking and face effects.
Physics, networking, and shared worlds
One avatar in a viewer is a demo. Several avatars in a place, with gravity and collisions, is a product.
- Rapier - A Rust physics engine with deterministic cross-platform simulation and first-class WebAssembly bindings.
- Jolt Physics - The multi-core physics engine used in shipped AAA titles, with a WebAssembly build that runs in the browser.
- cannon-es - A maintained fork of cannon.js in modern JavaScript. Small and readable, and enough for character controllers and props.
- Colyseus - Authoritative multiplayer rooms with automatic state synchronisation and delta encoding, in Node.
- Yjs - CRDTs for shared editable state, with providers for WebRTC, WebSocket, and IndexedDB persistence.
- LiveKit - An open WebRTC server for real-time audio and video, with SDKs everywhere and a self-hostable path.
- mediasoup - A selective forwarding unit as a Node library rather than a service, for teams that want to own the media path.
- geckos.io - Unreliable UDP-like messaging in the browser over WebRTC data channels, which is what fast-moving avatar positions actually want.
Agent frameworks and protocols
The mind that drives the body. Listed here because embodiment is a rendering problem bolted onto an agent problem, and most teams underestimate the second one.
- Model Context Protocol - The open protocol for giving a model tools, resources, and prompts from an external server. The cleanest way to let an agent act on a 3D pipeline.
- Claude Agent SDK - Build agents on the same harness Claude Code runs on: tool loops, subagents, permissions, and session state.
- OpenAI Agents SDK - A small framework for multi-agent handoffs, guardrails, and tracing, with a deliberately thin abstraction layer.
- LangGraph - Agents as explicit stateful graphs, with checkpointing and human-in-the-loop pauses. The right shape when a flow must be resumable.
- Vercel AI SDK - One TypeScript interface across model providers, with streaming, structured output, and tool calls that work in the browser.
- smolagents - Agents that write Python to call their tools instead of emitting JSON, in about a thousand lines of library code.
- CrewAI - Role-based multi-agent orchestration with an emphasis on readable task delegation.
- AutoGen - A research-grade framework for conversational multi-agent systems, with an event-driven core and a visual builder.
- Mastra - A TypeScript agent framework with workflows, memory, and evals, aimed at shipping rather than research.
- Letta - Agents with persistent memory as a first-class server concept, out of the MemGPT research line.
- A2A - An open protocol for agents built by different vendors to discover each other and collaborate over a common envelope.
- Generative Agents - The Stanford simulation where twenty-five agents with memory, reflection, and planning produced emergent social behaviour. Still the clearest demonstration of why memory architecture matters.
- Voyager - An embodied agent that plays Minecraft by writing and saving its own skills as code, then composing them. The best existing argument for embodiment as a learning signal.
Datasets and simulators
What the models in this list were trained and evaluated on, and where embodied agents go to practise.
- Objaverse - Around 800,000 annotated 3D objects, and the XL release pushes past ten million. The dataset that made open 3D generation viable.
- Objaverse-XL - The download and processing tooling for the ten-million-object release, including deduplication and rendering scripts.
- ShapeNet - The categorised CAD-model dataset that anchored 3D deep learning for a decade, mirrored on Hugging Face while its original site is unreliable.
- Habitat - A fast photorealistic simulator for training embodied agents to navigate and manipulate in indoor scenes.
- AI2-THOR - Interactive 3D rooms where objects have state: doors open, eggs break, appliances turn on. Built for agents that change the world rather than tour it.
- BEHAVIOR - A benchmark of a thousand everyday household activities defined in a logic language, with simulated physics and states.
- ScanNet - Richly annotated RGB-D scans of real indoor spaces, and the benchmark most scene-understanding work still reports on.
Free assets
Models, textures, and lighting you can ship without a lawyer. Every project in this list looks better with a real HDRI behind it.
- Poly Haven - HDRIs, PBR textures, and models, all CC0, all high quality, with no account required. Start here.
- ambientCG - Thousands of CC0 PBR material sets at multiple resolutions, downloadable in bulk.
- Kenney - Tens of thousands of CC0 game assets, including clean low-poly 3D kits that prototype beautifully.
- Quaternius - Stylised CC0 model packs with consistent proportions across sets, so a scene assembled from several packs still looks intentional.
- Sketchfab - The largest 3D model library on the web, with a downloadable Creative Commons subset and glTF export.
- OpenGameArt - A long-running community archive of openly licensed art. Uneven, and it holds things that exist nowhere else.
Learning
How to get good at this. Read in roughly this order.
- three.js manual - The official guide, rewritten around real problems rather than API surface. The fastest correct start.
- Discover three.js - A free book that teaches the concepts underneath the library, including colour management and performance work most tutorials skip.
- Three.js Journey - Bruno Simon's paid course, and the most thorough path from nothing to shaders and production scenes.
- The Book of Shaders - An interactive introduction to fragment shaders. Work through it once and materials stop being magic.
- WebGL Fundamentals - What every abstraction in this list is hiding, explained from first principles with runnable examples.
- WebGPU Fundamentals - The same treatment for WebGPU, which is where the web's rendering stack is heading.
- Real-Time Rendering resources - The companion site to the standard reference text, with a maintained portal into graphics literature.
Contributing
Additions are welcome. Read CONTRIBUTING.md first: entries live in data/awesome.json, not in this file, which is generated. The list is published under Apache-2.0, and each linked project carries its own license.