AI-CAD Uses Multi-Agent LLMs to Design Manufacturable Mechanical Parts

Summary

AI-CAD is an open-source multi-agent harness for mechanical engineering CAD rather than a conventional modeling tool. Users provide a plain-English brief or hand-drawn sketch, and a group of LLM agents plans the machine, decomposes it into parts, writes parametric CadQuery Python, renders engineering views, checks design-for-manufacturing and design-for-assembly constraints, proposes repairs, and reviews the result. The system produces named-dimension Python parts, STEP and STL exports, eight engineering-view PNGs per part, assembly renders, and a self-describing project directory containing goals, plans, logs, negotiations, measurements, and reflections. Its deterministic Python tools execute geometry, render views, validate specifications and dimensions, score DFMA rules, detect placeholder parts, and export files independently of the LLM; repair proposals that worsen the weighted DFMA score are rejected. Agents coordinate through the filesystem, with no server or database, and a new orchestrator can reconstruct or resume a run after a crash or context-window failure. A local read-only dashboard displays project artifacts and agent activity. The public gallery includes an impeller assembly with six components, 0.8000 mm measured clearance, six backswept blades at 10.5 degrees, one repair cycle, and a reviewer verdict of ship after about 2 hours 12 minutes; other examples include a piston-crank core, spur-gear pair, and control chest. Claude Code is the tested default harness, while OpenCode support is in progress. The project supports Linux, macOS, and Windows with symlink requirements, needs Python 3.13+, uv, and a system Cairo library for rendering, and is licensed under Apache 2.0.