Chisel

An MCP server that lets AI agents build, edit, render, and export real 3D models — from geometric primitives and boolean CSG. No diffusion, no GPU, no browser. The agent calls modeling tools, gets back a multi-view render to see its work, iterates, and exports clean .glb/.obj.

Every model is real, editable geometry — boxes, spheres, cylinders, cones combined with union / subtract / intersect and mirrored for symmetry. The render above (a chair built from 9 tool calls) is produced by a pure-CPU software rasterizer — the same thing the agent sees each step.

Why

Image/mesh diffusion models paint surfaces — you get a blob you can't cleanly edit or boolean. Chisel takes the opposite approach: give an agent a discrete, inspectable modeling API and a see-and-revise loop, and it builds real solids you can open in Blender, slice for printing, or drop into a game engine.

The hard part of 3D-by-agent isn't geometry, it's occlusion — any single view hides half the model. So Chisel renders four canonical views every step (front / side / top / iso) and hands them back as an image, so the agent reasons about the whole form, not one face.

Quick start

Run the server (stdio):

# zero-install, straight from GitHub (builds on first run)
npx -y github:EYamanS/chisel

# …or from a clone
git clone https://github.com/EYamanS/chisel && cd chisel
npm install
npm run mcp

Wire it into any MCP client (claude_desktop_config.json, .mcp.json, etc.):

{
  "mcpServers": {
    "chisel": {
      "command": "npx",
      "args": ["-y", "github:EYamanS/chisel"],
      "env": { "CSG_OUTPUT_DIR": "/absolute/path/to/exports" }
    }
  }
}

Or point at a local clone for speed:

{
  "mcpServers": {
    "chisel": {
      "command": "node",
      "args": ["/absolute/path/to/chisel/dist/server.js"],
      "env": { "CSG_OUTPUT_DIR": "/absolute/path/to/exports" }
    }
  }
}

Then just ask your agent: "model a coffee mug and export it as a glb."

Tools

Tool	What it does
add_box add_sphere add_cylinder add_cone	Add a primitive (size/radius/height, position, rotation, color).
transform	Move / rotate / scale an object (absolute or relative).
union subtract intersect	Boolean CSG — fuse, carve a hole, or clip to an overlap.
mirror	Reflect an object across X/Y/Z. Composes with booleans — build one side, mirror it.
set_color delete select	Recolor, remove, highlight.
get_scene	Return the scene graph as text (every object, shape, transform).
render	Return a 2×2 multi-view PNG (front/side/top/iso) so the agent can see the model.
export_model	Write .glb or .obj to disk; returns the path. inline:true for base64.
reset	Clear the session to an empty scene.

Models persist per session id (default main) for the life of the process, so an agent builds incrementally across calls. Rendering is GPU-free (a software rasterizer over the evaluated CSG triangles), so it runs anywhere Node runs — laptops, CI, containers.

A typical agent loop

add_cylinder { radius: 0.6, height: 1.2, position: [0, 0.6, 0], color: "steel" }   -> obj1
add_cylinder { radius: 0.46, height: 1.1, position: [0, 0.74, 0] }                 -> obj2
subtract     { a: "obj1", b: "obj2" }   # hollow it out                            -> obj3
render       # look at all four views, notice it needs a handle
add_box      { size: [0.16, 0.62, 0.16], position: [0.66, 0.62, 0], color: "steel" }
union        { a: "obj3", b: "obj4", name: "mug" }
render       # looks right
export_model { format: "glb" }          # -> ./exports/main-<ts>.glb

Web playground (optional)

The repo also ships an interactive Next.js playground where an OpenAI model drives the same engine in your browser, with a live 3D viewport and glTF/OBJ export:

cp .env.local.example .env.local   # add OPENAI_API_KEY
npm run dev                        # http://localhost:3000

The Demo buttons (Mug / Table / Rocket) exercise the full engine with no API key.

How it works

src/lib/scene/    types.ts        scene graph: flat list of CSG expression trees
                  operations.ts   deterministic reducer — applies one tool call
                  tools.ts        modeling tool schemas (shared by MCP + web agent)
src/lib/three/    build.ts        scene graph -> Three.js meshes (evaluates CSG)
src/lib/render/   raster.ts       headless software renderer (no GPU) -> 2x2 PNG
                  export.ts       glTF/OBJ export | png.ts encoder | font.ts labels
src/mcp/          server.ts       MCP server (stdio) | engine.ts session store
src/lib/agent/    loop.ts         web playground's render->see->revise agent loop
src/components/   Viewport.tsx    browser viewport (WebGL) + capture + export
src/app/          page.tsx        playground UI | api/agent/route.ts OpenAI proxy

The MCP server and the web playground are two front-ends over one engine — same scene graph, same CSG, same tool schemas. The MCP path is fully headless (software-rendered); the web path uses WebGL for the interactive viewport.

Development

npm run mcp         # run the MCP server from source (tsx)
npm run build:mcp   # bundle the self-contained binary -> dist/server.js
npm run dev         # web playground
npm run build       # production build of the web app

License

MIT © Emir Yaman Sivrikaya