mesh2cad

Turn 3D meshes (STL/OBJ/PLY) into real, editable CAD (STEP) — fitted planes, cylinders, spheres, and cones instead of triangle soup. CLI + MCP server for AI agents. MIT, pure Python on top of OpenCascade.

Left: 21k-triangle mesh with 0.1 mm noise. Right: the STEP mesh2cad writes — 10 planes + 2 cylinders, max deviation 0.05 mm from the true part.

Install

pip install "mesh2cad[mcp] @ git+https://github.com/emirmrkaljevic/mesh2cad"

(PyPI release coming.)

Related MCP server: TianshangCAD

Use

mesh2cad convert part.stl -o part.step --tol 0.1 --report report.md
mesh2cad report part.stl --json

--tol ≈ the noise / accuracy of your mesh in model units (default 0.1 % of the bounding-box diagonal).

Output: STEP + fidelity report — method (exact / faceted-hybrid / faceted), face types, solid validity, two-sided Hausdorff / RMS deviation, volume error, % area per surface type.

MCP server (Claude Desktop, Cursor, …)

{ "mcpServers": { "mesh2cad": { "command": "mesh2cad", "args": ["mcp"] } } }

Tools: convert(input_path, output_path?, tol?, cylinders?, spheres?, cones?), analyze(input_path, tol?, cylinders?, spheres?, cones?), report(input_path, tol?, cylinders?, spheres?, cones?). spheres and cones default on.

Listed on Glama and the TensorBlock MCP catalog.

How it works

1. Load + clean mesh (trimesh); Taubin-smoothed copy for normals.

2. Region growing into planes; strip chains and leftovers are fitted as cylinders, spheres, or cones (normal-ray / algebraic / constant-normal-angle fits, robust nonlinear refinement). A coaxial cylinder + cone component (countersink, chamfered pin) is split by normal-to-axis angle first. Small noisy patches that tile a sphere are merged into one spherical region. Split/redundant regions are merged and absorbed.

3. Exact B-rep (OCP / OpenCascade): plane polygons from triple-plane intersections, circles from plane ∩ cylinder and plane ∩ sphere, plane ∩ cone, cylindrical faces, spherical caps and zones, conical frustums, sewn into a solid and validated.

4. Fallbacks: hybrid faceted solid (vertices snapped to fitted surfaces, coplanar triangles merged), then raw faceted solid.

5. Fidelity report against the input mesh.

Self-test benchmark

Ground-truth cadquery parts → tessellated + subdivided (≤ 2 mm edges) → Gaussian vertex noise → mesh2cad → compared with the truth. 3 seeds, worst case shown (python benchmarks/run.py); the same cases run in CI.

"exact" = valid B-rep with the true face count. All rows are the worst of 3 seeds from python benchmarks/run.py (v0.3.0). Rows with fewer than 3/3 exact fell back to a faceted solid (face count = triangles): noisy narrow countersink/chamfer bands don't constrain a cone yet, so CI checks those parts only where they are exact.

Limits (v0.3)

Planes, cylinders, spheres, and cones. Cylinders must end on planes perpendicular to their axis (through-holes, bosses, pins). Spheres are exact when each boundary loop is a single plane ∩ sphere circle (domes, spherical bosses, spherical pockets, and zones between two parallel planes). A sphere that runs into an edge or a cylinder falls back to the hybrid solid — vertices are still snapped onto the fitted sphere, but the face stays faceted. Cones are exact when they run between two planes perpendicular to their axis: countersinks, chamfered holes and bosses, drafted/tapered bosses. Cones may also meet a coaxial cylinder (countersunk hole, chamfered pin). A cone that ends in its apex falls back to the hybrid solid, and countersinks are reliable up to ~0.05 mm noise, narrow chamfers on clean meshes only for now (wide drafted walls up to 0.2 mm). No fillets or free-form yet.

Development

pip install -e .[dev]
pytest
python docs/make_demo.py   # regenerate images (needs matplotlib, pillow)

License: MIT