Multi-Agent-CAD

MAC — Multi-Agent CAD for Parts and Assemblies from Natural Language

Tsinghua University, College AI, IEI Lab

简体中文

MAC turns an ordinary design brief into editable engineering geometry. It can generate a single printable part or decompose a mechanism into separately generated components, assemble them, validate the result, and export the artifacts needed for downstream inspection and simulation workflows.

MAC was originally released as a single-part Text-to-CAD workflow and has since received nearly 1,000 GitHub stars. This release preserves that workflow while adding visual verification, an independent Judge, and experimental multi-part assembly generation.

Assembly Gallery (Technology Preview)

Building on the original single-part workflow, MAC v2 introduces experimental natural-language assembly generation. The examples below are curated outputs, not a measured prompt-success rate, and the assembly workflow is not presented as outperforming CAD-agent skills. Validate generated geometry and joints before manufacturing or simulation use.

Complex assemblies

Three-Axis Gantry Metrology Cell	Telescopic Cinema Robot Crane
A bridge-style inspection platform with three orthogonal motion stages.	A multi-stage crane with an articulated support and telescopic payload arm.

Heavy-Duty Mobile Manipulator	Advanced Vision Inspection Robot Arm
A mobile base carrying a visually detailed articulated manipulator.	A multi-joint inspection arm with a framed vision payload.

Complex-gallery generation usage

Compact functional assemblies

Hinged Twin-Claw Gripper	Guided Linear Plunger	Rotary Fork-Key Tool	Reusable Five-Digit Hand

These curated compact examples exercise natural-language decomposition, joint interfaces, repeated-part reuse, and mirrored geometry. The complex gallery shows how the same workflow scales to larger visual assemblies. See the assembly workflow documentation for implementation details and current limitations.

URDF Simulation

The assembly workflow can export generated mechanisms as URDF files for downstream robotics and physics simulation. The demonstrations below use an AI-generated hand assembly to execute object rotation and pick-and-place tasks in simulation.

Object Rotation	Pick and Place

Part Gallery

MAC also generates standalone mechanical and creative parts. The examples below are editable CAD results, not image-only generations.

Honeycomb Organizer	Gyroscope Ornament	Lighthouse	Smartphone Stand	Ball-in-Cage

Articulable Gyroscope	Multi-Link Chain	Geneva Mechanism	Plasma Reactor	Brake Disc

The single-part workflow documentation covers configuration, execution, caching, and QA. Benchmark prompts and per-model data are linked from its benchmark-details section.

Real-World Results

The printed collection above includes benchmark parts and original showcase models generated by the single-part workflow. MAC can also produce print-in-place mechanisms containing separate bodies and functional clearances:

Single-Part Benchmark

The following numbers apply specifically to the documented ten-prompt, 141-feature single-part-workflow benchmark. They are not assembly success-rate or assembly-cost claims.

Under the Qwen 3.7 setup, MAC v1 used 116× fewer recorded tokens and had a 13× lower estimated cost than the reproduced Skill baseline. Under Qwen 3.8 with visual verification, MAC v2 used 62.7× fewer recorded tokens and had a 6.1× lower estimated cost.

MAC v1 refers to the original single-part workflow and its published benchmark run. MAC v2 refers to the updated single-part workflow with visual verification and the independent Judge. Assembly generation is presented separately as a technology preview and is not included in these benchmark figures.

Across the two documented model configurations, MAC retained a 140/141 feature pass rate while using substantially fewer recorded tokens and lower estimated cost than the corresponding reproduced Skill runs.

Methodology and raw breakdowns: single-part README, English evaluation, and Chinese evaluation.

Why MAC?

• Natural-language input — describe geometry, dimensions, interfaces, and motion without writing CAD code or an internal schema.
• Parts and assemblies — use one project for standalone printable parts and articulated, multi-component mechanisms.
• Editable engineering outputs — export STEP, STL, and GLB rather than a render-only result; assemblies also provide a URDF handoff.
• Auditable generation — inspect structured briefs, geometry plans, generated Python, measurements, QA reports, and repair history.
• Automatic execution and repair — generated CAD is executed and checked, with bounded feedback loops for recoverable failures.
• Reuse instead of regeneration — repeated and mirrored components can be derived from one source geometry.
• Model-flexible stages — configure different OpenAI-compatible models for planning, geometry, coding, and repair.
• Visual workflow — the single-part pipeline includes a browser UI with 3D preview and downloadable results.
• Built on a proven workflow — MAC v2 extends the original nearly 1,000-star single-part pipeline with visual verification and assembly generation while retaining the existing part-generation workflow.

Quick Start

Install

git clone https://github.com/Pan-Chera/Multi-Agent-CAD.git
cd Multi-Agent-CAD
conda env create -f environment.yml
conda activate multi_agent_cad
pip install --no-deps "aider-chat==0.82.3"
export DASHSCOPE_API_KEY="your-key"

The final pip install is needed because aider-chat pins NumPy 1.x, which conflicts with build123d’s NumPy 2.x requirement. It is therefore omitted from environment.yml; --no-deps avoids replacing the working NumPy version.

The default configuration targets an OpenAI-compatible DashScope endpoint. Other providers are covered in the single-part configuration guide.

The bundled cadpy runtime currently requires OCP 7.x, so installation constrains build123d to <0.13. Version 0.13 switches to OCP 8 and cannot load the current STEP runtime. If an existing environment has already upgraded, run python -m pip install "build123d>=0.8,<0.13" to restore compatible dependencies.

pip users (no conda): aider-chat pins numpy==1.26.4, but build123d>=0.8 requires numpy>=2,<3 — these conflict in pure pip. Use this workaround (verified on macOS arm64 + Python 3.11):

python3.11 -m venv .venv
source .venv/bin/activate          # Windows PowerShell: .venv\Scripts\activate
pip install --upgrade pip
# Install aider first (pulls numpy 1.26.4 + transitive deps), then force-upgrade numpy.
# Verified: aider 0.82.3 imports cleanly on numpy 2.x — the pin is over-cautious upstream.
pip install "aider-chat==0.82.3"
pip install --no-deps --force-reinstall "numpy>=2,<2.3"
pip install "build123d>=0.8,<0.13" "langgraph>=0.2,<0.3" "langgraph-checkpoint>=2.0,<3.0" \
            "pydantic>=2.5" "openai>=1.20.0" "anthropic>=0.30" \
            "trimesh>=4.0" "rtree>=1.1" "scipy>=1.10" "scikit-learn>=1.3" \
            "fastapi>=0.110" "uvicorn[standard]>=0.27" "ipython>=8.15" "pytest>=7.4"
# --no-deps skips re-checking the numpy pin in pyproject.toml; fastapi+uvicorn
# are already installed by the previous step, so the [web] extras resolve.
pip install --no-deps -e .

The last step registers the mac-config-reset console script and lets you run python -m multi_agent_cad.graph from any directory. See requirements.txt / pyproject.toml for the canonical dependency list.

Windows: the same conda env create + pip install --no-deps aider-chat==0.82.3 flow works — trimesh and rtree come from conda-forge prebuilt; OCP is pulled in transitively by build123d (via its PyPI dep cadquery-ocp-novtk). Don’t use the pure-pip workaround below on Windows — native wheels for trimesh/rtree can be unreliable. Set the API key in PowerShell as $env:DASHSCOPE_API_KEY = "sk-..." (or set DASHSCOPE_API_KEY=sk-... in cmd.exe). For the Web UI under conda, pip install -e ".[web]" inside the activated env works — uvloop auto-skips on Windows. CI covers Windows with Python 3.11 and Linux with Python 3.11/3.12, including an isolated wheel installation check.

Generate one part

Set USER_REQUEST in multi_agent_cad/config.py, then run:

python -m multi_agent_cad.graph

For the single-part browser UI:

pip install -e ".[web]"
python -m multi_agent_cad.web

The provider selector includes OpenRouter (https://openrouter.ai/api/v1). Switching presets fills suggested stage and Aider model names; models you have edited manually are preserved. Existing endpoint and model settings are loaded from config.py when the page opens. For OpenRouter, Aider uses an OpenAI-compatible model name such as openai/qwen/qwen3-max.

If the key field is blank, the Web UI uses the server’s DASHSCOPE_API_KEY, then DS_API_KEY from config.py. Explicit form keys take priority; API clients can also supply the existing config.DS_API_KEY request field. Server keys are never included in the configuration schema response or written to job config files. Use a key for the selected provider.

The UI listens on 127.0.0.1 by default. It executes generated Python with your user account’s permissions, so do not expose it directly to an untrusted network. Copying results to an arbitrary local directory is disabled unless MAC_WEB_ALLOW_DEST_PATH=1 is explicitly set.

Web UI security and environment variables

State-changing API endpoints (POST, PUT, PATCH, DELETE) require a custom X-MAC-CSRF: 1 header. The browser frontend sends it automatically; non-browser clients (curl, HTTP libraries) must add it manually or the server returns 403:

curl -X POST http://127.0.0.1:8000/api/run \
  -H "X-MAC-CSRF: 1" \
  -H "Content-Type: application/json" \
  -d '{"prompt": "a 30 mm cube", "api_key": "YOUR_API_KEY"}'

GET endpoints (/api/health, /api/jobs, status polling) are unaffected.

The following environment variables control Web UI behavior:

• MAC_WEB_HOST — bind address. Loopback values (127.0.0.1, localhost, ::1, the default) also enable a loopback Host allowlist as a defense-in-depth CSRF layer. Binding to a non-loopback address (LAN deployment) disables that layer; the X-MAC-CSRF header remains required.
• MAC_WEB_ALLOW_CUSTOM_ENDPOINT=1 — opt in to non-loopback OpenAI-compatible DS_BASE_URL hosts: vLLM on a private hostname, a corporate gateway, or Ollama / LM Studio on a non-loopback address. Default off; otherwise the host must match the provider allowlist (loopback, api.openai.com, api.deepseek.com, openrouter.ai, generativelanguage.googleapis.com, .aliyuncs.com, .googleapis.com). Scheme (http/https) is always enforced.
• MAC_WEB_ALLOW_DEST_PATH=1 — opt in to the dest_path artifact-copy feature. When set, MAC_WEB_DEST_ROOT must also be set to an absolute directory, and each request’s dest_path must be relative and resolve under that root. Absolute dest_path and .. traversal are rejected.
• MAC_CHILD_ENV_ALLOW=VAR1,VAR2 — pass additional non-secret environment variables through to the generated-Python subprocess (for example, HTTP_PROXY,HTTPS_PROXY on a corporate network). The default allowlist is PATH, HOME, TMPDIR, TMP, TEMP, LANG, LC_ALL, LC_CTYPE, PYTHONPATH, PYTHONIOENCODING, LD_LIBRARY_PATH, DYLD_LIBRARY_PATH, SYSTEMROOT, WINDIR, plus ITERATION set per child. This applies to deterministic / Aider-assisted cadpy generation as well as the LLM fallback and repair runners. Never list API keys or credentials here — env-var stripping reduces secret leakage but does not sandbox the subprocess; see SECURITY.md.

Generated CAD execution, including the cadpy runner, uses CAD_SCRIPT_TIMEOUT from multi_agent_cad/config.py. Increase it for models whose geometry or export needs more time; execution failures continue through the existing repair flow.

Generate an assembly

Assembly input is also ordinary natural language. Run an included request:

MAC_ASSEMBLY_REQUEST="$(cat mac_assembly/assembly_prompts/natural_language_benchmarks/01_hinged_twin_claw_gripper.md)" \
python -m mac_assembly

Or provide a brief directly:

MAC_ASSEMBLY_REQUEST="Create a two-part hinged clamp with a fixed base and one rotating jaw." \
python -m mac_assembly

Assembly jobs are written under assembly_jobs/job_<timestamp>/, including separate-part geometry, assembled STEP/STL/GLB, URDF, manifests, and QA artifacts. See the assembly guide before relying on an output for manufacturing or simulation.

Documentation

• Single-part workflow: setup, architecture, benchmarks, and Web UI
• Assembly workflow: decomposition, mating, reuse, QA, and exports
• 单零件中文文档

Citation

If you find this project useful for your research, please consider citing:

@misc{mac2026,
  author       = {Guanxing Qu and Xueyan Zou},
  title        = {MAC (Multi-Agent CAD): A Decoupled Multi-Agent Framework for Text-to-CAD Generation},
  year         = {2026},
  publisher    = {GitHub},
  journal      = {GitHub repository},
  howpublished = {\url{https://github.com/Pan-Chera/Multi-Agent-CAD}}
}

The quantitative single-part evaluation uses earthtojake/text-to-cad (CAD Skills) as the comparison baseline. If you cite the benchmark comparison, please also cite that project:

@misc{texttocad2026,
  author       = {earthtojake},
  title        = {CAD Skills: A skills library for CAD, robotics, and hardware design agents},
  year         = {2026},
  publisher    = {GitHub},
  journal      = {GitHub repository},
  howpublished = {\url{https://github.com/earthtojake/text-to-cad}}
}

License

MIT — see LICENSE.

The vendored packages/cadpy STEP/GLB runtime is derived from earthtojake/text-to-cad (CAD Skills) and is redistributed under its original MIT license — see packages/cadpy/LICENSE.

Acknowledgements

• Tsinghua University, IEI Lab — the lab where this project was developed; provided the research environment and advisor guidance.
• earthtojake/text-to-cad (CAD Skills) — source of the comparison baseline and the ten shared single-part benchmark prompts. The vendored packages/cadpy runtime also derives from this project and retains its original MIT copyright.
• build123d — algebraic B-rep CAD kernel.
• LangGraph — stateful agent orchestration.
• Aider — LLM-driven code repair.
• Qwen Model Studio — OpenAI-compatible models used in the documented experiments.