mesh2cad takes an STL and gives back CAD: a STEP file with real planes, cylinders and cones, and, when the part is simple enough, a CadQuery or OpenSCAD script with named parameters. The point of the script is that you can open it and change `hole_1_d` instead of fighting triangles. I test on six real Original Prusa i3 MK3S printer parts. Yesterday's post got five of them to exact STEP. 0.12.0 does two things: it sews the last one, and it makes the belt idler editable as a script. ## The last STEP: a bore that meets everything On y-rod-holder every surface was already fitted in 0.11.0, but the solid didn't close. The R=4.00 mm rod bore is one cylinder band whose boundary loop touches several planes and the coaxial 45 degree chamfer circles at both ends. The sewing code only handled a band that meets one or two simple caps, so it gave up and the output was the raw mesh: valid, but 3043 triangles and nothing to edit. 0.12.0 trims a cylinder or cone band against all the planes and coaxial circles in its loop. Two smaller fixes were needed on the same part: a hole end that is a full chamfer circle can now meet a cap that is split across coplanar pads, and a short chamfer that region growing had glued onto a plane is split back off when its vertices leave that plane. Result: 158 planes, 8 cylinders, 11 cones, Hausdorff 0.2178 mm, volume error 0.391%, 12.4 s. The cylinders and cones fit to under 0.005 mm. The remaining 0.22 mm is three shallow patches kept as single planes, with vertices 0.16 to 0.24 mm off the plane and no internal edge sharper than about 10 degrees. I'd rather report that than hide it. So all six real parts are now exact STEP solids. ## The editable part: chamfers, blind holes and pockets STEP is good for measuring and for other CAD tools, but you can't say "make this hole 2 mm bigger" to a STEP file. For that, mesh2cad writes a small script: intersected extrusions along X, Y and Z, cuts, and holes. Until now it refused any part with a chamfer on a hole, which ruled out most printed parts. 0.12.0 adds to that vocabulary: * a 45 degree axis-aligned chamfer on a hole, as `hole_N_chamfer` (the axial length) * a blind hole, as `hole_N_depth` * a blind pocket, and a shallow floor such as a hex nut pocket or an embossed logo y-belt-idler now exports both scripts. The solid is the X and Y silhouettes intersected, plus the belt grooves, hex pockets, logo floors, an Ø3.2 hole along X with an Ø5.6 counterbore, a 0.5 mm blind hole along Y with a 45 degree taper, and two Ø3.3 holes along Z with a 0.4 mm mouth chamfer. Every script is rebuilt and compared to the mesh before it is written. The gate is the same as before: max deviation at most max(4 x tolerance, 0.1 mm) and volume error at most 1%. For the idler the limit is 0.192 mm: script | max deviation | volume error ---|---|--- CadQuery | 0.1534 mm | 0.144% OpenSCAD | 0.1639 mm | 0.145% There is a test where a chamfer on an X hole through a slanted extrusion is edited from 0.8 mm to 1.2 mm, and the volume changes by exactly the analytic ring. That is the behaviour I want: change a number, get the geometry you'd expect. ## Still refused Three of six parts are parametric now (plug-aligner, endstop-block, y-belt-idler). The other three are refused, and the message says why instead of writing a wrong script: * **y-motor-holder:** "combined extrusion deviates by 7.924 mm (limit 0.266), volume error 30.07%". Its Ø5.6 / Ø3.2 screw holes run from z=0.4 to z=8 on a part that is 19 mm tall. They are stepped holes that don't go through, which the vocabulary doesn't have yet. * **y-rod-holder:** "combined extrusion deviates by 1.503 mm (limit 0.138)". The Ø3.3 hole along Y isn't one radius for its whole length, so a single cylinder sits about 1.5 mm off the mesh where the opening widens. * **y-belt-holder:** "combined extrusion deviates by 2.517 mm (limit 0.154), volume error 1.97%". Better than 24 mm in 0.11.0, but the outline still has a gap. These are the next three things to build, in that order. ## Try it pip install mesh2cad mesh2cad convert part.stl -o part.step mesh2cad convert part.stl --openscad part.scad It also runs as an MCP server (`mesh2cad mcp`), so an assistant can convert a part and then edit a parameter for you. If you have an STL of a simple mechanical part that fails, an issue with the file attached is the most useful thing you can send: github.com/emirmrkaljevic/mesh2cad.

mesh2cad turns an STL back into CAD: a STEP file with real planes, cylinders and cones, plus a CadQuery or OpenSCAD script when the part is simple enough. I test it on six real Original Prusa i3 MK3S printer parts, because synthetic test shapes lied to me before. After 0.10.0, two of the six came back as exact STEP solids. The other four fell back to a faceted solid (valid, but just triangles glued into faces). The tool reported a small "unfitted" area on each, and I wrote in the README that these were probably edge rounds: torus fillets and sphere corners that the fitter didn't know yet. That was wrong, and finding out how it was wrong is what made 0.11.0 work. ## Check the diagnosis against ground truth These Prusa parts are open source, and the STEP files the STLs were exported from are in the same repository at the same commit. So instead of writing a torus fitter, I first listed the surface types in those STEP files. There were no tori and no spheres at all. The surfaces were planes, cylinders, 45 degree cones (chamfers), and a few small B-spline patches for the belt tooth profile, about 0.1 to 0.2% of the area. So the "unfitted" regions were not a missing surface type. They were segmentation mistakes: * **Faces glued across a real edge.** A region grown across a 45 or 90 degree edge contains two planes, so no single surface fits it, and the whole patch was marked unfitted. * **Short chamfers rejected.** A 45 degree chamfer 0.6 mm tall has so few vertices that the cone test threw it out. * **Cylinders left behind after a chamfer was carved out of them.** The hole was fine, but its band was split into pieces that each looked too small. * **Flat polygons whose corners lie on a circle.** On the belt idler, eight bands with radius 11.1 mm looked like a cylinder. Every vertex is exactly on that circle, but the face centres are about 7 mm off it and the angle between neighbours goes up to 90 degrees. They are planes. A fitter that only checks vertices will call a polygonal boss a cylinder or even a sphere. ## What changed in 0.11.0 All of these only run when the mesh is a clean CAD tessellation, never on a noisy scan: * A region that crosses an edge sharper than about 35 degrees is split and each piece is refit. * A cylinder or sphere candidate must also match its face centres, not just its vertices, so polygons stay planes. * A short 45 degree chamfer is kept when its vertices lie on the cone to roundoff. * A ring of single-facet planes whose centres lie on one full cylinder is merged back into that cylinder. * A hole circle that ends on several coplanar pads that don't touch is sewn as arcs of one circle. * A plane corner whose three-plane intersection is unstable uses the mesh vertex, if that vertex lies on all three planes. A fillet between two planes is deliberately left alone. ## Numbers From `python benchmarks/real_parts.py` (H is Hausdorff distance from the mesh, in mm): part | 0.10.0 | 0.11.0 ---|---|--- y-belt-idler | faceted-hybrid, H 0.28, 12.3 s | exact: 197 planes, 6 cylinders, 3 cones, H 0.0082, 5.1 s y-belt-holder | faceted, 890 faces, 21.0 s | exact: 210 planes, 4 cylinders, H 0.16, 6.0 s y-motor-holder | faceted-hybrid, H 0.27, 13.4 s | exact: 396 planes, 6 cylinders, 2 cones, H 0.10, 12.2 s y-rod-holder | faceted, 38.6 s | still faceted, 34.2 s So five of six real parts are now exact STEP. The cylinders are the real screw and bearing holes (R 2.80, 1.65, 1.60, 1.50 mm) and the cones are the real chamfers. The 0.16 mm and 0.10 mm Hausdorff on the holder and motor holder come from the belt profile and a partial R=11.25 mm round that meets a chamfer, which stay as planes. ## What still doesn't work * **y-rod-holder doesn't sew.** Every region is now fitted (144 planes, 8 cylinders, 11 cones), but the R=4 mm rod bore is one loop that meets several chamfers, and the solid doesn't close. It falls back to the raw mesh: valid, Hausdorff 0, but not editable. * **Parametric scripts are still only 2 of 6.** plug-aligner and endstop-block get OpenSCAD and CadQuery scripts with real parameters. The idler, motor holder and rod holder are refused with "a countersink, chamfer or draft on a combined extrusion is not supported yet". The belt holder is refused because the combined extrusion would deviate by 24 mm. The tool says this instead of writing a wrong script. * **B-spline belt profiles stay planes.** That is fine for printing and measuring, less nice for editing. ## The lesson When a geometry tool fails on real data, the tempting move is to add the next surface type. Here the missing surface type didn't exist. If you have the original CAD for some of your test inputs, compare against it before writing new fitting code. It saved me writing a torus fitter nobody needed. Try it: pip install mesh2cad mesh2cad convert part.stl -o part.step mesh2cad convert part.stl --openscad part.scad It also runs as an MCP server (`mesh2cad mcp`) so an AI assistant can convert and edit parts for you. Next up are chamfers on combined extrusions, which would make the idler and motor holder parametric, and sewing the rod holder bore. Bug reports with STL files that fail are the most useful thing you can send: github.com/emirmrkaljevic/mesh2cad.

Abstract page for arXiv paper 2610.04262: CADForge: Agentic Single-View CAD Reconstruction with Explicit Geometry Reasoning

I had an idea I wanted to float. I was reading a children's book to my kiddos and it had a clause that it couldn't be used to train ai model

We gave six AI agents the same three printable parts to model, three in CadQuery and three in OpenSCAD, then verified every mesh independently. Both toolchains shipped. The difference is in how they fail.

We gave six AI agents the same three printable parts to model, three in CadQuery and three in OpenSCAD, then verified every mesh independently. Both toolchains shipped. The difference is in how they fail.