mcp-spatial-critic

An MCP server (Model Context Protocol, stdio JSON-RPC 2.0) that gives headless CAD agents a deterministic 3D-printing critic. Zero dependencies — pure Python standard library, no numpy, no trimesh, no Blender, no OpenSCAD at runtime.

An LLM can write OpenSCAD that looks correct in a screenshot and still exports a mesh that is inside-out, floating 5 mm above the plate, split into three bodies, or full of zero-area slivers. mcp-spatial-critic fails the build before the printer does.

agent → generate STL → mcp-spatial-critic → {"approved": false, "errors": ["NON_MANIFOLD"]}
                                          → agent repairs → re-audit → {"approved": true}

Why agents need this

Text-only agents cannot see geometry. Vision models see a render, not a manifold. The failure modes below all render beautifully and all fail in a slicer:

Related MCP server: ReverseCAD MCP

The five invariants

Every audit returns a critic_score from 0.0 to 10.0 and a list of critiques.

approved is true only when the error list is empty, the mesh is watertight and there is a single shell. The score is a ranking signal for repair loops; approved is the gate.

Install

# from GitHub — works today
pip install git+https://github.com/vanovri/mcp-spatial-critic

# or a checkout you can hack on
git clone https://github.com/vanovri/mcp-spatial-critic
cd mcp-spatial-critic
pip install -e .

Requires Python 3.10+. No third-party runtime packages.

Register the MCP server

Add this to your agent's MCP configuration (Claude Desktop claude_desktop_config.json, Claude Code .mcp.json, or any MCP-compatible client):

{
  "mcpServers": {
    "spatial-critic": {
      "command": "mcp-spatial-critic",
      "args": []
    }
  }
}

The server resolves relative paths against the directory it was launched from. Set MCP_SPATIAL_CRITIC_PRESETS to override where list_presets looks.

Tools

audit_stl

{
  "name": "audit_stl",
  "arguments": { "path": "build/bracket.stl", "min_wall_mm": 2.0 }
}

Response payload (also usable directly as a Python dict via audit_stl_file):

{
  "file": "bracket.stl",
  "path": "build/bracket.stl",
  "triangle_count": 12,
  "vertex_count": 8,
  "watertight": true,
  "non_manifold_edge_count": 0,
  "shell_count": 1,
  "bounding_box_mm": {
    "width_x": 20.0, "depth_y": 20.0, "height_z": 10.0,
    "min_z": 0.0, "max_z": 10.0
  },
  "volume_mm3": 4000.0,
  "degenerate_triangle_count": 0,
  "critic_score": 10.0,
  "approved": true,
  "critiques": [],
  "errors": []
}

isError is set on the MCP result whenever approved is false, so a tool-using agent sees the failure as a tool error rather than a success.

generate_reference_box

Writes a known-good watertight box and audits it. Use it as a control experiment when an agent claims its exporter is broken.

{ "out_path": "reference_box.stl", "width": 20, "depth": 20, "height": 10 }

list_presets

Lists the parametric presets bundled in presets/:

python presets/pagoda_roof_tier.py > pagoda.scad
openscad -o pagoda.stl pagoda.scad

Talk to it without an agent

The transport is line-delimited JSON-RPC 2.0 on stdio, so you can drive it by hand:

printf '%s\n' \
  '{"jsonrpc":"2.0","id":1,"method":"initialize"}' \
  '{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"generate_reference_box","arguments":{"out_path":"box.stl"}}}' \
  | mcp-spatial-critic

Library use

from mcp_spatial_critic import audit_stl_file, audit_stl_mesh, generate_watertight_box_stl

report = audit_stl_file("build/bracket.stl", min_wall_mm=1.6)
if not report["approved"]:
    for critique in report["critiques"]:
        print("FAIL:", critique)

# or audit an in-memory mesh
report = audit_stl_mesh(generate_watertight_box_stl(height=1.0))
assert "THIN_FEATURE" in report["errors"]

How it works

• ASCII + binary STL parsers (core.py) — binary uses struct with a 50-byte stride; ASCII tolerates CRLF, tabs and malformed vertex lines.

• Manifold test — canonicalised undirected edge keys (vertices rounded to 1e-5 mm) counted in a dict; any count ≠ 2 is a non-manifold edge.

• Shell count — union-find with path compression over shared vertices.

• Orientation — signed volume via the divergence theorem, Σ dot(v1, cross(v2, v3)) / 6.

• Deterministic output — no timestamps, no random ordering; the same STL always produces the same bytes on the wire.

• Server — handle_request() is a pure function (dict → dict | None), so the whole protocol is unit-testable without starting a process.

python -m unittest discover -s tests -v   # 40 tests

Known limitations (deliberate, tracked as issues)

1. min_wall_mm is compared against the bounding box, not per-wall ray casts. A 40 × 40 × 20 mm hollow box with 0.4 mm walls passes this check. Real thin-wall detection needs slice-plane sampling — see GOOD_FIRST_ISSUES.md #3.

2. Only STL is supported. 3MF and OBJ are open tasks.

3. The critic diagnoses, it never repairs. Auto-repair belongs in the agent loop.

4. Binary STL headers that begin with solid are disambiguated by looking for the facet keyword in the first 512 bytes. A pathological header could fool it.

5. No build-volume check yet (the bbox is reported, the comparison is not).

Contributing

See GOOD_FIRST_ISSUES.md for scoped, test-covered starter tasks. Every PR needs a test that fails before the fix. Run the suite with python -m unittest discover -s tests -v.

License

MIT — see LICENSE.