A small, plain-Python agent harness. The LLM drives an external CAD MCP server to build a part from your prompt — then a deterministic kernel verifier decides PASS, never the model's own word.
# install the CLI (the command is `chamfer`) uv tool install chamfer-cad # or: pip install chamfer-cad # build a part from a prompt (needs a provider key in your env) chamfer run "a 2 L electric kettle with a rounded rim" --provider anthropic # deliver the verified STEP to a path of your choice chamfer run "a desk lamp" --provider openai -o out/lamp.step
On first run chamfer seeds ~/.chamfer with default skills and an
MCP config (build123d). Drop a CHAMFER.md or .chamfer/skills/ in your project to extend it.
After the agent says "done", a fixed kernel probe re-measures every solid and grades the checklist itself. No self-reported success.
The whole loop — agent, tools, verifier — reads top to bottom. Flat src/*.py, no magic.
Capabilities come from an MCP server + skills, not baked in. The harness itself is CAD-agnostic.
Anthropic, OpenAI, OpenRouter, or the Claude Code / Codex CLIs — one plain-text protocol.
Each GIF is a script-for-script replay of a real session log — actual geometry states, not illustrations. (These runs are from an earlier Fusion-backed version; the current release drives an external build123d MCP server through the same measured loop.)