The platform for 3D AI agents — render, embody, own, embed.
Drop a 3D model, give it an LLM brain and a voice, mint its identity on-chain, and embed it anywhere — in the browser. No plugins. No installs.
We build the runtime & identity layer for embodied AI agents.
Today's AI agents are text in a chat box — no face, no body, no persistent identity, no place in the 3D/spatial web. As AI moves from chat to digital humans, there is no open platform to create, run, and distribute them.
three.ws is that platform.
| Primitive | What it does | |
|---|---|---|
| Render | WebGL 2.0 viewer (three.js r184) loads & validates glTF/GLB with Draco, KTX2, Meshopt — zero server upload | ✅ Live |
| Embody | LLM brain runs a tool-loop: listen → reason → act (animate, express, skills, memory) → speak with real-time lip-sync | ✅ Live |
| Own | On-chain identity — ERC-8004 on 15+ EVM chains, Metaplex Core on Solana, signed action history & reputation | ✅ Live |
| Embed | <agent-3d> web component + 5 widgets, OpenGraph/oEmbed, versioned CDN — agent into any site | ✅ Live |
Plus a shipped social layer: multiplayer 3D worlds (/play, /club, /city), presence & DMs, voice/pose/mocap studios, and a paid agent-to-agent economy (x402).
A live agent that listens, reasons, animates, and speaks in real time. three.ws/agents
<agent-3d src="…"> — WYSIWYG editor at three.ws/embed.
3 photos → GPU reconstruction → riggable GLB. How it works
Shared 3D space per token — peer avatars, chat, voxel building. three.ws/play
Reviewers approve faster with a live product + public GitHub + demo link. We lead with all three.
Event-bus architecture decouples body, brain, voice, memory & skills — every action flows through typed events with a 200-action replay buffer.
Per-turn tool-loop: STT/text in → system prompt assembled from manifest + recalled memory + installed skills → LLM called with full tool set → up to 8 tool iterations → spoken response with ARKit-52 (52-blendshape) lip-sync.
Skills are composable bundles loadable from IPFS/HTTP, with payment-gated trust modes via x402.
play_clip setExpression lookAt speak remember skill
@three-ws/avatar · /avatar-schema · /avatar-cli · /solana-agent · @three-ws/mcp-server
Claude primary; open models (Llama 3.3 70B, Mistral, Qwen) via OpenRouter/Groq & Livepeer decentralized GPU. Keys never client-side.
Selfie/text → textured, rigged avatar. TRELLIS, Hunyuan3D, UniRig — diffusion+transformer, ~30–90s on A100/H100.
WebGL 2.0 on every client GPU; WebGPU on the roadmap. PBR, skinned mesh, morph targets, BVH raycasting.
Every avatar created and every open-model inference burns GPU cycles — demand scales linearly with users.
three.ws is digital humans + embodied agents — exactly the workload NVIDIA's AI stack targets.
| three.ws component | NVIDIA technology |
|---|---|
| Avatar lip-sync (ARKit-52), facial animation | NVIDIA ACE — Audio2Face |
| Speech-to-text + text-to-speech | NVIDIA Riva |
| LLM agent brains served at scale | NVIDIA NIM + TensorRT-LLM |
| Selfie/text → 3D (TRELLIS, Hunyuan3D, UniRig) | CUDA inference · NIM for 3D / Edify 3D |
| 3D asset pipeline (glTF/GLB, retarget) | Omniverse / OpenUSD interop |
| Decentralized inference network (roadmap) | NVIDIA-accelerated node operators |
The plan is concrete: Inception GPU credits run avatar-gen + open-model inference, and Inception support carries our digital-human stack onto ACE + NIM.
Vanilla JS + Vite · three.js r184 WebGL 2.0 · standalone Svelte chat app
Vercel serverless (Node 24) · Neon Postgres · Cloudflare R2 · Upstash Redis · Colyseus multiplayer (Fly.io)
Replicate · GCP Cloud Run (NVIDIA GPUs) · HF Spaces — pluggable, env-selected generative-3D inference
ERC-8004 (15+ EVM chains) · Metaplex Core (Solana) · x402 payments · OAuth 2.1 · MCP over HTTP · OpenAPI 3.1
Who it's for: creators & brands embedding talking 3D agents; AI builders who need a face/voice/body; web3 & consumer-social communities.
The convergence: capable LLMs + fast image/text→3D + on-chain identity — none existed together 18 months ago. AI is moving from chat to embodied digital humans, and the tooling is fragmented and closed.
Our wedge: the only open, browser-native, on-chain agent platform — vs. closed digital-human SaaS and text-only agent frameworks.
Reviewers expect one credible figure, not a guess.
<agent-3d> + 5 widgets on versioned CDNRoadmap: ACE/NIM migration → personalization & reputation markets → open decentralized inference network.
Selfie/text→avatar generation — GPU cost + margin. Primary variable cost is GPU.
Platform fee on agent-to-agent paid skill calls, asset downloads, and royalties.
Premium widgets, custom voices, private skills, brand analytics.
Name service (.threews.sol), launchpad, token & community tooling.
With Inception: GPU credits + ACE/NIM efficiency gains directly improve margin and let generation scale with demand.
Founder — product, engineering, and the agent platform. Builds three.ws in public at github.com/nirholas.
Developer / Engineer — platform and product engineering.
Engineering-heavy, shipping daily across real-time 3D, agents, and infrastructure.
Evidence of execution: a production platform spanning real-time 3D, LLM agent runtimes, GPU inference orchestration, smart contracts on 15+ chains, and multiplayer infra — live at three.ws.
three.ws — embodied AI, owned by its creators.