Four LLM Agents Design Printable Parts for 1% the Token Cost

What it does

MAC is a decoupled multi-agent framework that converts a natural-language description into a printable 3D model, emitting STEP or STL files. It routes the task through four LLM roles—spec planner, architect, coder, and repair—coordinated by LangGraph and backed by the build123d CAD kernel. The pipeline handles everything from simple brackets to articulable print-in-place assemblies with sub-millimeter clearances.

The interesting bit

The authors claim that slicing the job across specialists drops token usage from 103.9 M to 896 k and shrinks inference cost from ¥125.69 to ¥9.67 on their 141-feature benchmark suite, while nudging the pass rate from 97.9 % to 99.3 %. The system speaks OpenAI-compatible APIs, so the default Qwen backend can be swapped for DeepSeek, Ollama, or others without touching the graph logic.

Key highlights

• Benchmarks against the CAD Skills baseline report 116× fewer tokens, 13× lower cost, and a 99.3 % pass rate (140/141 features).
• Supports multi-body print-in-place kinematics—ball-in-cage fidget toys and gyroscopes—with 0.4–1 mm clearance gaps generated as a single STEP file.
• Offers both a terminal mode with mid-run steering checkpoints and a Web UI with an in-browser 3D viewer and one-click downloads.
• Provider-agnostic: defaults to Alibaba Cloud DashScope (qwen3.7-max) but accepts any OpenAI-compatible endpoint via config fields.
• Uses an open-source geometry stack: build123d, OpenCASCADE, LangGraph, and Aider for automated repair.

Verdict

Try it if you need programmatic, fabrication-ready mechanical parts from prose and want to keep API bills low. Look elsewhere if your workflow centers on organic sculpting or manual parametric CAD.