Text-to-CAD

LLMs are good at code but bad at keeping a 3D model coherent across turns. A model generated in one shot rarely survives a follow-up like "make the neck taller", because the second generation starts from scratch or edits the wrong geometry. CAD also demands exact, reproducible artifacts: a mesh that regenerates differently is worse than none.

A text-to-CAD harness built on coding agents with a Python CAD runtime on build123d over OpenCascade. The agent writes real, source-controlled model code instead of returning a blob of geometry. The harness exports STEP, STL, DXF, GLB, and topology outputs, plus URDF robot descriptions through a bundled URDF skill. Every artifact is generated by a vendored CLI tool, so regeneration is a command, not a hope.

Vendored agent skills (CAD and URDF) expose CLI tools: gen_step_part, gen_step_assembly, gen_urdf, cadref, and snapshot. The agent calls these to build and preview, and @cad[...] prompt-reference handles let a follow-up edit point at the exact part it means, with stale references detected and flagged instead of silently applied. Generated artifacts are checked into the repo, so a model is as reviewable as a pull request. A local CAD Explorer (React, Three.js, Vite) renders the results, and node:test suites cover geometry, URDF parsing and kinematics, DXF parsing, and scene scale logic.

The viewer came after the generation harness, which made debugging geometry slow. I would ship a minimal renderer first, even a crude one, because seeing the part while iterating on the prompt loop saves more time than any export format.