lego-mcp

An MCP server where a model can only build physically valid Lego. Instead of letting the model free-form geometry (and hallucinate impossible structures), every action goes through a constraint engine that behaves like real plastic:

• Catalog-only parts — brick_2x4, plate_1x6, tile_2x2, slopes (slope_2x2, slope_inv_1x2, …), baseplates (baseplate_32x32, …). Unknown part names are rejected with suggestions. No hallucinated 3×5 bricks.

• Solid bricks — placements that overlap existing pieces are rejected, with the exact colliding cells and the lowest free z reported back.

• Stud-to-tube connection required — a piece must sit on studs of an existing piece, on the baseplate, or press its own studs up under an overhang. Floating bricks are rejected with a valid z suggested.

• Tiles are smooth — nothing can stack on a tile.

• Slopes are directional — studs only on their non-sloped rows (nothing attaches to the sloped face); rotation is 0/90/180/270 and slopes descend toward +y at rotation 0. Inverted slopes grip below only on their tall rows.

• Baseplates are ground-only — studs on top, smooth underneath, z=0 only.

• No orphaning — removing a piece that would leave others disconnected from the baseplate is refused ("remove those first, top down").

• Stability hints — extreme cantilevers succeed but return a warning.

The error messages are written for the model: each rejection explains the violated constraint and suggests a fix, so an agent converges on buildable structures instead of drifting.

Built by a model, through the constraints

Every piece below was placed by an LLM whose only tools were this server's — each brick validated stud-by-stud, then exported to Blender and rendered. No piece floats, overlaps, or stacks on a tile.

The corbelled arch is the constraint engine at its most demanding: every course steps inward off the studs of the one below, and the two flanks have to meet at the crown without a single floating brick — the model gets there because each illegal step is bounced back with the reason and a legal alternative.

Related MCP server: mcp-toolbox

Coordinates

• x, y in studs, z in plates (1 brick = 3 plates tall).

• z=0 is the baseplate; a brick on top of a brick placed at z=0 goes at z=3.

• rotation is 0 or 90 (swaps the footprint).

Tools

State auto-saves to <state>/<name>.json after every mutation and the latest build is resumed on server restart. The state directory is, in order: $LEGO_MCP_HOME/builds if set, else the repo's builds/ when run from a writable checkout, else ~/.lego-mcp/builds when installed. Set LEGO_MCP_HOME to keep builds wherever you like.

Install

lego-mcp is a standard stdio MCP server, so any MCP-capable host can run it. It's packaged as a normal Python distribution — the only thing a host needs is a command that launches it.

Run it with no install (recommended)

uvx fetches and runs it in a throwaway environment, so there's nothing to maintain:

uvx --from "git+https://github.com/Axel-Jalonen/lego-mcp" lego-mcp   # straight from this repo
uvx --from ./dist/lego_mcp-0.1.0-py3-none-any.whl lego-mcp           # from a built wheel

Or install it as a tool

pipx install "git+https://github.com/Axel-Jalonen/lego-mcp"   # or: pip install ...
lego-mcp                     # the `lego-mcp` command is now on PATH

Or grab the prebuilt wheel from the latest release and pipx install lego_mcp-0.1.0-py3-none-any.whl.

Build the distributable yourself

uv build          # -> dist/lego_mcp-0.1.0-py3-none-any.whl  (+ .tar.gz)

The .whl is the portable artifact: copy it anywhere, uvx --from <the.whl> lego-mcp, done. No source checkout, venv, or absolute paths needed.

Wiring it into a host

Every MCP host takes the same shape — a command + args. Use uvx so the host doesn't need Python set up:

{
  "mcpServers": {
    "lego": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/Axel-Jalonen/lego-mcp", "lego-mcp"],
      "env": { "LEGO_MCP_HOME": "~/lego-builds" }
    }
  }
}

• Claude Desktop → claude_desktop_config.json (Settings → Developer).

• Claude Code → claude mcp add lego -- uvx --from git+https://github.com/Axel-Jalonen/lego-mcp lego-mcp (this repo also ships a dev .mcp.json).

• Cursor / Windsurf / VS Code MCP → the same mcpServers block.

• Ollama → Ollama's own runtime doesn't load MCP servers directly; run it through an MCP host that talks to Ollama, e.g. mcphost or Open WebUI. mcphost reads the identical config above, so pointing it at the same uvx command gives your local model the eight Lego tools.

Development

python3 -m venv .venv
.venv/bin/pip install -e . pytest          # editable install
.venv/bin/python -m pytest tests/          # 15 constraint tests

Pointing a small model at it

agent/build_agent.py runs an agent loop where a model's only tools are this server's — no filesystem, no bash, no repo access. The constraint engine rejects every invalid move with a corrective hint, so even a small model can only ever produce physically valid builds:

export ANTHROPIC_API_KEY=sk-ant-...
.venv/bin/python agent/build_agent.py "build a small watchtower"
.venv/bin/python agent/build_agent.py --model claude-sonnet-5 "build a bridge"

Defaults to claude-haiku-4-5. The final state is reported by the engine (get_build), not by the model's own claims, and builds land in builds/.

Layout

• lego_mcp/catalog.py — part definitions (footprint, height in plates, stud availability, LDraw ids) and colors

• lego_mcp/engine.py — occupancy grid, connection rules, orphan detection, ASCII views, save/load

• lego_mcp/export.py — LDraw + Blender exporters

• lego_mcp/server.py — FastMCP tool layer