Multi-Agent-CAD

Why we included this project

Text-to-CAD is genuinely hard because a model has to turn plain language into geometry that will actually print, and MAC attacks it with four agents that pass structured state between steps instead of making one enormous model call. The output is editable STEP solids built on build123d and OpenCascade, real geometry you can machine or print rather than a rendered mesh. Splitting the work across agents keeps the context window small, which is why the project reports roughly a hundred times fewer tokens than single-agent baselines and a similar drop in inference cost. The articulable print-in-place examples are the most interesting part, since they show how clearance gaps are controlled so joints move freely right off the build plate. If you prototype mechanical parts or want a text-to-CAD pipeline you can run yourself at a fraction of the token spend, this is a good place to start.