AI-CAD: An OSS Multi-Agent Harness for Mech. Eng. CAD

An open-source multi-agent harness for mechanical CAD automates the engineering workflow: give a plain-English brief or sketch and a team of LLM-driven agents plans the machine, decomposes it into parts, writes parametric CadQuery Python for each component, renders engineering views, scores designs against DFMA/assembly rules, proposes repairs, and iterates until the design passes review. Outputs are concrete engineering artifacts - named-dimension Python part programs plus STEP/STL exports and eight engineering-view renders per part - stored in a self-describing project directory that can be diffed and versioned like any codebase. The system emphasizes manufacturability as a hard gate, preserves design intent, and presents a live read-only dashboard and an agent activity waterfall to watch runs in real time.

The architecture separates deterministic verification from LLM choice: bash-callable Python tools execute geometry, render views, run measurements, and score DFMA rules so validation is reproducible and independent of the models. Agent roles (orchestrator, planner, cad_designer, validator, dfma_inspector, repair, assembly_resolver, sourcing, reviewer) draw on explicit, versioned skills (cadquery cookbooks, anti-hallucination lists, dfm rules, run-reflection schema). An orchestrator is mortality-proof - state can be rebuilt solely from the filesystem - and every run ends with a structured run_reflection.md capturing failures, improvisations, and lessons. The project is harness-agnostic, Apache-2.0 licensed, and designed so generated parts survive engineering review and downstream manufacturing workflows.