rig-worker

An agent describes a character as simple shapes and gets back one closed, vertex-coloured mesh. It then places bones, binds them with automatic (bone-heat) weights, keys an animation, and checks every step with rendered pictures and numbers. Everything runs as plain JavaScript inside a Worker: no Blender, no GPU, nothing to install.

What it does

On a test creature mesh, the dominant bone of 97.6% of vertices matches Blender 4.5's Automatic Weights with the same skeleton. In one test an agent (Claude Opus 5.5) modelled a 44-shape toy character, rigged it with 19 bones and made it walk. It took 14 minutes and cost $2.59 in API calls plus about a cent of Worker CPU.

Related MCP server: modelgen

Use the public demo

The demo needs no token. It allows 30 requests a minute, of which 6 can be model/bind calls. Files are kept per client and deleted after a day.

claude mcp add --transport http rig-worker https://rig-worker-demo.epinnock.workers.dev/mcp

Other MCP clients use {"type": "http", "url": "https://rig-worker-demo.epinnock.workers.dev/mcp"}. Then ask your agent something like "Use rig-worker to model a small toy robot, rig it and make it wave. Show me the pictures."

To run the whole example pipeline from a terminal (model → skeleton → bind → paint → walk → qa → export):

git clone https://github.com/epinnock/rig-worker && cd rig-worker && npm install
node examples/pip/run.mjs        # pictures and pip_rig.glb land in out/pip

Self-host

You need the Workers Paid plan, because a bind on a large mesh can use up to a minute of CPU. Deploying creates the R2 bucket. Set RIG_TOKEN to a long random string. Clients send it as Authorization: Bearer <RIG_TOKEN>, and it also signs download links.

npm install
npx wrangler deploy
openssl rand -hex 32 | npx wrangler secret put RIG_TOKEN
claude mcp add --transport http rig-worker https://rig-worker.<you>.workers.dev/mcp \
  --header "Authorization: Bearer <RIG_TOKEN>"

Besides POST /mcp there is a plain HTTP API: PUT/GET/DELETE /files/<name>, GET /files, GET /dl/<name>?exp&sig and POST /bind (multipart). The public demo is the same code deployed with npx wrangler deploy --env demo; see wrangler.jsonc.

Develop

cp .dev.vars.example .dev.vars   # put a token in it
npm run dev                      # wrangler dev on :8799 with local R2
npm test                         # model -> bind -> paint -> animate -> qa in Node, ~15 s, no network

Conventions. Coordinates are glTF metres: +Y up, +Z front, +X the character's left. Rotations are Euler degrees applied X, then Y, then Z, about world axes through the joint for bones bind made (rigs from other tools turn about their own bone axes). The docs have every shape, field and tool.

Known gaps. There is no smoothing operation for weights and no undo for paints. qa does not separate stretched edges from squashed ones. Rebuilding a model drops its rig, so you bind again. model makes vertex colours only, no UVs or textures (meshes you bring keep theirs). Skinning is linear blend with 4 influences.

Blender

Every GLB rig-worker writes opens in Blender with the default glTF import settings and needs no clean-up (tested with Blender 4.5):

• The rest shape, vertex colours (wired into the material), weights (one vertex group per bone) and every clip (one action each) match rig-worker's own numbers. The deformed mesh is within 0.05 mm at eight sample times per clip, in Blender 4.5.

• Bones point along the limbs, head to tail, as in a rig made in Blender.

• Meshes and rigs you bring from Blender keep their UVs, textures, bone rolls and actions.

• Blender's re-export reads back into rig-worker the same.

The test that checks all of this is in test/blender:

BLENDER=/path/to/blender npm run test:blender   # ~30 s; reports in work/blender

License

MIT © 2026 Ejiro Pinnock