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.
Agent-native 3D — create, embody, and distribute, end to end
An agent given a goal ("get yourself a body") generates the 3D assets it needs and uses them — no browser, no mocks, no human in the loop. This is the roadmap campaign's work order 10 demo (retired; readable in git history): the closing proof that three.ws's MCP tools compose into a real autonomous create-and-use loop, and that the result is trivially distributable everywhere.
What it does
Drives the live, free three.ws "Free 3D Studio" MCP server
(/api/mcp-studio — see docs/mcp-studio.md) through
the full chain:
tools/list— confirms the composable tool contracts are live.mesh_forge(prompt)— text → a static 3D mesh (Granite-directed model chain).rig_mesh(glb_url)— auto-rig it into an animation-ready GLB.create_agent_persona(glb_url, name)— save the rigged GLB as a NAMED, persistent agent body.persona_say(persona_id, text)— perform a line through that body (lip-sync + emotion).get_agent_persona(persona_id)— reload the SAME body in a fresh call, proving continuity (not a per-call random result).- Build every distribution snippet — iframe,
<model-viewer>,<agent-3d>, a@three-ws/page-agenttalking guide, and a@three-ws/walkcompanion — using the exact same pure functions (src/forge-embed-snippets.js) the Forge web app's "Embed this model" panel uses, so what this script prints is byte-identical to what a human copies from the UI.
All six MCP tools it calls are free — no x402 payment, no wallet, no API key. Every request/response is a real HTTP round-trip against production (or your local dev server); nothing here is stubbed.
Run it
cd examples/agent-native-3d
node run.mjs
Against a local dev server instead of production:
npm run dev # from the repo root, in another terminal
MCP_STUDIO_URL=http://localhost:3000/api/mcp-studio node run.mjs
Every request and response is written to
prompts/roadmap/_generated/10/agent-native-3d-transcript.json
after each run.
About the rate-limit fallback
mesh_forge / rig_mesh share one platform-wide free-tier generation lane
(the same one /forge web-page drafts use) with a real capacity ceiling. If
that lane is saturated when you run this, the script backs off using the
server's own quoted retry_after (bounded, never a blind sleep loop) and, if
it's still saturated past the budget, falls back to a known-good rigged GLB
(https://three.ws/avatars/default.glb) so steps 4–7 — embodiment, speech,
continuity, and distribution — still run for real. The fallback is logged
loudly in the console and recorded in the transcript (result.fallback_used),
never silently substituted; the failed generation calls are preserved verbatim
in transcript.steps for audit.
Why this matters
This is the "own → use everywhere" half of the roadmap's closing loop (prompts 08/09/10): generate a body, embody an agent identity with it, and hand out real distribution in one call each — the same primitives a human uses from the Forge UI, callable end to end by an agent with zero UI.