Turn text into 3D printable CAD models with MAC

MAC (Multi-Agent CAD) is an open-source framework that turns simple natural-language design briefs into editable, 3D-printable CAD models.

Instead of outputting non-functional mesh approximations, MAC produces standard STEP, STL, and GLB files. It can also break down multi-component mechanisms, generate each part separately, assemble them, and export simulation-ready URDF files for robotics workflows.

What it can do

You type an ordinary description, like a hinged twin-claw gripper, a telescopic cinema robot crane, or a three-axis gantry metrology platform. MAC plans the geometry, generates individual parts, and handles the assembly interfaces automatically.

It handles standalone mechanical models just as well. Examples include honeycomb organizers, gyroscopes, Geneva mechanisms, and print-in-place articulated chains that come off the print bed already moving.

The system derives mirrored and repeated components from a single source rather than regenerating them. Every run leaves behind a structured brief, geometry plans, raw Python code, measurements, and a full repair history.

How it works and benchmark numbers

Traditional text-to-CAD workflows often trap large language models in expensive trial-and-error loops that burn tokens without converging on valid geometry.

MAC decouples the pipeline across dedicated agents for planning, geometry design, code execution, and bounded repair. It also introduces test-time compute constraints and an independent Judge.

In a standardized single-part benchmark covering 10 prompts and 141 geometric features using Qwen 3.7, a replicated CAD Skill baseline consumed 103.95M tokens at an estimated cost of $18.59, hitting a 97.9% pass rate.

MAC v1 completed the same task using only 0.90M tokens and $1.43, boosting the pass rate to 99.3%. That translates to 116 times fewer recorded tokens and a 13-fold reduction in cost.

In MAC v2, adding visual verification and an independent Judge with Qwen 3.8 cost $4.85 across 1.40M tokens while maintaining the same 99.3% pass rate. That is still 62.7 times fewer tokens and 6.1 times cheaper than the corresponding baseline.

Curated multi-part assemblies scale predictably. Generating a vision inspection arm costs around $1, a three-axis gantry cell runs about $2, and a heavy-duty mobile manipulator costs around $13 in API tokens.

Getting it running locally

The code is hosted on GitHub under Pan-Chera/Multi-Agent-CAD.

It is model-agnostic. The default configuration hooks into OpenAI-compatible DashScope endpoints for Qwen models, but you can point it to OpenRouter, DeepSeek, or other providers just as easily.

You will need a Python 3.11 environment. Because build123d and aider-chat have conflicting NumPy version constraints, the project documentation provides a straightforward Conda setup that installs aider without strict dependencies.

A built-in FastAPI web UI lets you test prompts, rotate 3D previews directly in your browser, and download the resulting CAD files. Because the tool runs generated Python locally, keep the service on localhost rather than exposing it to untrusted networks.

Why it matters

If you work with 3D printers, design physical mechanisms, or build simulated environments for robotics, MAC is genuinely worth pulling down to test.

Its real strength lies in breaking the problem apart. Instead of asking one model to hallucinate an entire mechanical assembly in a single shot, it splits planning, coding, and inspection into distinct bounded steps that protect your API budget.

Assembly generation remains an experimental technology preview, so you will still want to manually verify clearances before cutting metal. But as a fast track from text prompt to functional engineering geometry, this workflow sets an impressive standard for open-source CAD tooling.