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.
Grok makes a 3D model through three.ws
Grok generates a real, downloadable 3D model (GLB) by calling the free three.ws
MCP server through xAI's Responses API. xAI connects to the server and runs the
tool calls on its side; this script only talks to api.x.ai. Zero dependencies,
Node 20 or newer.
The server is https://three.ws/api/mcp-grok, the three.ws surface built for
Grok: every free studio tool, no account, no key. See the Grok guide
for Grok Bot and Grok connectors, and three.ws/grok for a live demo.
Run it
cd examples/grok-remote-mcp
# 1. Prove the three.ws server answers and offers the tools (no key needed)
node index.mjs --check
# 2. See the exact request the script sends to xAI (no key, no call)
node index.mjs --dry-run
# 3. The real thing
XAI_API_KEY=xai-... node index.mjs "a lunar lander with gold foil legs"
Get a key at console.x.ai. The run prints Grok's reply and the viewer link, which opens an interactive 3D viewer in any browser.
| Variable | Default | Purpose |
|---|---|---|
XAI_API_KEY |
none | Your xAI API key. Required for a real run. |
XAI_MODEL |
grok-4.7 |
Any xAI model that supports remote MCP tools. |
THREE_WS_MCP_URL |
https://three.ws/api/mcp-grok |
Point at a local server (http://localhost:8080/api/mcp-grok) while developing. |
How it works
The request carries one remote MCP tool:
{
"type": "mcp",
"server_url": "https://three.ws/api/mcp-grok",
"server_label": "three-ws",
"allowed_tools": ["search_catalog", "forge_free", "forge_avatar", "check_job"]
}
allowed_tools keeps Grok's context small and limits it to the tools this task
needs. Remove it to expose the whole catalog, including rigging, refinement and
the persona tools that give Grok a talking 3D body.
Slow renders. A model usually lands in under a minute, but a GPU that has
scaled to zero boots first. The server never holds a call past 40 seconds: it
answers with a pending job and tells Grok to collect it with check_job. Grok
cannot wait between calls inside one response, so the script continues the same
conversation with previous_response_id, every 30 seconds, until Grok replies
with a viewer link. Grok Bot needs none of this; it waits on its own.
Quota. Generation is free and rate-limited per MCP session. The server
issues an Mcp-Session-Id on initialize, and xAI echoes it, so your quota is
yours even though every Grok user reaches three.ws from xAI's cloud.