breptile

Convert STL triangle meshes into clean, editable STEP BREP models — not just a tessellated dump. Planar regions become single planar faces; cylindrical holes and bosses become true analytic cylinders; spherical countersinks and domes become real spheres. Anything that can't be fitted within tolerance falls back to faceted geometry and is flagged in a JSON report for manual (or LLM-assisted) rebuild.

Left: input STL (3,476 triangles). Right: converted STEP — 249 BREP faces with 35 planes and 17 true cylinders. Every hole selects as a single cylindrical face in CAD.

Why

STL files carry no topology and no analytic surfaces, so most "STL to STEP" converters emit one planar face per triangle — a file that opens in CAD but is unusable for editing, CAM, or feature recognition. breptile reconstructs analytic surfaces instead, with a hard guarantee: every fitted surface passes through the mesh vertices within a configurable tolerance, and dimensions are never snapped.

Install

python3 -m venv .venv
.venv/bin/pip install breptile        # or: pip install -e . from a checkout

Requires Python 3.10–3.13 (needs OCP/build123d wheels). Dependencies: build123d, trimesh, numpy, scipy, manifold3d, rtree, networkx.

Use

breptile input.stl output.step                      # auto: fit primitives, per-region fallback
breptile input.stl output.step --mode tessellated   # guaranteed success, faceted STEP
breptile input.stl output.step --mode prismatic     # planes only
breptile input.stl output.step --tol 0.01 --report report.json --verify

Or from Python:

from breptile import convert
report = convert("input.stl", "output.step", mode="fit", tol=0.01)

How it works

1. Load & repair (trimesh + manifold3d): normals, holes, degenerate faces; clear error on non-watertight input.
2. Segment: region-grow smooth patches by dihedral angle; classify each by least-squares fit — plane → cylinder → sphere — validated against --tol. Cylinder axes come from the facet-normal cloud; fits are refined with Levenberg–Marquardt.
3. Rebuild BREP (OpenCascade via OCP):
   • planar regions → single faces with hole wires;
   • full-wrap cylinders and spherical bands/caps → analytic faces with exact circular rims shared with neighboring faces, so sewing closes analytically;
   • partial cylinders/sphere patches → trimmed patches with the parametric seam rotated into the region's angular gap;
   • a u/v coverage check prevents a fit from claiming surface the mesh doesn't cover;
   • everything else stays faceted (honest fallback).
4. Sew → solid → ShapeUpgrade_UnifySameDomain → ShapeFix → BRepCheck → STEP (AP214 or AP242).

Benchmark

Run on the trimesh model corpus (python benchmark/run_benchmark.py):

Highlights — 17/18 models produce a valid STEP solid (the 18th is deliberately random triangle soup, which degrades to a flagged open shell):

Verified deviation stays at or below the input mesh's own chord error in all cases — on coarse meshes the analytic surface is more accurate than the STL that described it.

Hybrid LLM workflow

.claude/skills/breptile/SKILL.md teaches Claude (or any agent) to run the pipeline, read the report, and rebuild the regions the fitter couldn't handle as build123d code — then verify the result against the original mesh. Regions carry their fitted parameters (axis/center/radius) even when rejected, giving the agent measured starting points.

Limitations

• Cones, tori, fillet blends, and freeform surfaces fall back to facets (cone and NURBS fitting are the roadmap).
• Organic/scanned shapes convert tessellated — a valid STEP, but not parametric.
• Coplanar-but-disconnected regions are not merged across bodies.

Development

.venv/bin/pytest            # round-trip tests on generated fixtures
python benchmark/run_benchmark.py

MIT licensed. Issues and PRs welcome.