CadSmith

CadSmith is a self-correcting text-to-CAD agent built on CadQuery. It turns a plain-English description into a validated parametric 3D model: it asks an LLM for CadQuery code, runs it, measures the resulting geometry, and if anything is wrong, a syntax error, a missing extrude, a units mistake, or an open shell, it feeds the exact failure back to the model and tries again until the part actually verifies.

After every generation it checks that the code executes, that volume is above a threshold, that the bounding box is sane, that the result contains a solid, that the B-Rep passes OCCT validity checks, and that the solid is watertight with no free edges. Verified models export as STL and STEP, with AMF and 3MF also supported. The README positions this as treating geometry as ground truth and closing the loop, rather than fine-tuning a bigger model.

CadSmith ships as a CLI and a Python library, with Anthropic and OpenAI backends built in and a bring-your-own-model option that takes any callable returning text. Generated code runs in-process by default, or in a subprocess with a hard timeout for untrusted models. An eval harness scores success rate and attempts-to-success over a prompt set, and an offline example runs the loop on a canned model with no API key.