MAC (Multi-Agent CAD)
Four-agent text-to-CAD framework that splits spec planning, geometric design, Python coding, and validation across separate LangGraph agents to generate printable build123d models with a fraction of the tokens single-model systems use.
What it does and how it connects
MAC is a decoupled multi-agent framework for text-to-CAD generation via constrained test-time compute, built on LangGraph with build123d as the geometry engine, from the Tsinghua University College AI IEI Lab. It turns an ordinary design brief into editable engineering geometry: a single printable part, or a mechanism decomposed into separately generated components, assembled, validated and exported as STEP, STL and GLB, with a URDF handoff for assemblies. Different OpenAI-compatible models can be configured for the planning, geometry, coding and repair stages, and a browser UI with 3D preview serves the single-part pipeline.
Generation is auditable: structured briefs, geometry plans, generated Python, measurements, QA reports and repair history can all be inspected, and generated CAD is executed and checked with bounded feedback loops for recoverable failures. On the documented ten-prompt, 141-feature single-part benchmark, MAC v2 with visual and geometric verification kept a 140 of 141 feature pass rate at 1.40M recorded tokens and about 4.85 dollars estimated model cost, against 62.7 times fewer recorded tokens than the reproduced Skill baseline. Assembly generation is an experimental technology preview with curated gallery outputs, not a measured prompt-success rate.
The README is candid about install friction: a fresh install can fail at the coder stage (see the voice below), and build123d must stay below 0.13 because the bundled STEP runtime needs OCP 7.x while 0.13 switches to OCP 8.
What engineers report
Fresh install fails at the coder stage: cadpy/step_scene.py imports TDF_LabelSequence, which no longer exists in OCP 8.x (pulled in by unpinned build123d>=0.8)
2026-10-14Read the original
Alternatives
Compare side by side- Chamfer
Writes build123d Python and measures the result against the request, with an optional live Fusion backend, rather than splitting the work across LangGraph agents.
- Text23D Mechanical
Turns chat prompts into CadQuery or FreeCAD models in a web workbench, rather than a multi-agent build123d pipeline.
- cad-agent (kerryshi)
Turns English into CadQuery through a frozen spec with spec-derived verification, rather than a multi-agent build123d pipeline.
- Forgent3D
Pairs coding agents with a desktop viewer to verify parametric models against real geometry, rather than a multi-agent pipeline.
- vibecad (JoeMattie)
A Claude Code skill that edits build123d models through a browser canvas, rather than a multi-agent pipeline.
Sources
Not verified
- Pricing: no product price is stated; the README's dollar figures are estimated model API costs for benchmark runs, not a price for MAC.
- A named individual builder beyond the Tsinghua University IEI Lab named in the README.
Last verified November 6, 2026. Listings are informational. The evidence score measures public, verifiable evidence and is not an endorsement or a compatibility claim.