cad-agent

CI License: MIT

A workshop where AI agents get better at CAD by learning from everyone else. While we design real hardware, our agents keep hunting the web for the best published techniques (open-source repos, design guides, datasheets, papers, other people's agent skills), try them on the problem in front of them, and condense what works into tested tools and short pages the next agent can use straight away. Every source gets its credit.

We are not trying to build the one toolkit to rule them all (though we will give it a go where nobody else has). The aim is a place where good ideas from all over are found, tested, kept if they hold up, and credited, so an agent pointed here spends its tokens on the new problem instead of rediscovering the old ones. Tested means against the source's own worked examples first, then on our parts where we can. CAD, calculation and simulation narrow a design down; they do not prove that a printed part fits or holds. Only testing the printed part does; a measured coupon replaces a data-sheet number for one filament and print direction. Browse it at cad-agent.valpatel.com (every page also as Markdown, plus llms.txt and catalog.json).

_{Drawn from the real CAD and results of Gyropup (MIT), a self-balancing wheel-legged robot designed with cad-agent: its leg's STEP file, the hip frame's FEA and the five-bar's link lengths. Made by tools/readme_anim.py.}

3D viewer and review site. Every part in place, with explode, section and one top bar linking every report.

Strength. FEA of a ribbed lid under a 300 N press, coloured by stress as a fraction of the material's strength.

Less material. Topology optimisation keeps only the floor that carries load.

How it improves itself

flowchart TD
    W["🛠️ <b>Real work</b><br/>a robot, an enclosure,<br/>a board"]
    G["🧩 <b>A gap</b><br/>a failed print, a review<br/>finding, a slow step"]
    S["🔎 <b>Search</b><br/>repos · guides · datasheets<br/>papers · agent skills"]
    T["🧪 <b>Test</b><br/>against their own examples,<br/>then on our parts"]
    K["📚 <b>Keep + credit</b><br/>a tested tool, a check,<br/>a living page, a lesson"]
    W --> G --> S --> T --> K --> W
    T -. "didn't hold up" .-> S
    classDef work fill:#fffdf5,stroke:#173d3c,color:#173d3c
    classDef find fill:#d8efbb,stroke:#386248,color:#173d3c
    classDef keep fill:#ffe1cc,stroke:#bd3a12,color:#173d3c
    class W,G work
    class S,T find
    class K keep

1. Work on something real. Most tools here started from a real project's need (Gyropup, the J501 enclosure, the AGX Thor CAN breakout).
2. Look before building. When a gap shows up, the agent first surveys what others have published and records what it found and why it adopted, borrowed from or passed on each one (research/, upstreams/registry.yaml, research/landscape.json).
3. Test it on our problem. A technique earns its place by reproducing its source's own worked examples, then by working on our parts where we can try it. Results are recorded, including the ones that failed. Until a part has been printed and measured, a limit taken from a data sheet is a hypothesis.
4. Condense it. What survives becomes a tested tool an agent calls in one step, a deterministic check that catches the mistake next time, and a living page: the method in a few paragraphs, with citations, examples and the limits stated plainly. Tests fail if the page and the tool drift apart.
5. Credit it. Every page names its sources; every borrowed file keeps its licence; the people and projects we learn from are listed below.

Example: snap fits condense the Covestro (Bayer) snap-fit guide, DuPont's design guide and the data sheets of 11 filaments (13 entries: two have dry and wet values) into cad-agent snapfit, tested against the guide's worked examples and sized by print direction; no printed coupon has been measured yet. The same loop turned review findings from real builds into the pitfall checks, like a rib that ran into a counterbore.

What's here today

How a project runs

A project describes itself in one cad-agent.yaml. cad-agent runs it and collects the proof.

flowchart TD
    P["📐 <b>Your project</b><br/>cad-agent.yaml<br/>CAD scripts · parameters"]
    B["🔧 <b>build</b><br/>CAD · checks · simulations<br/>run as ordered steps"]
    M["🖨️ <b>export-3mf</b><br/>bed-centred 3MF<br/>read back and verified"]
    R["✅ <b>check + report</b><br/>every result, pass or fail<br/>with its evidence"]
    Z["📦 <b>package</b><br/>review zip<br/>sha256 manifest"]
    W["🌐 <b>site --serve</b><br/>linked review site<br/>3D viewer · live reload"]
    P --> B --> M --> Z
    B --> R --> Z
    R --> W
    classDef src fill:#fffdf5,stroke:#173d3c,color:#173d3c
    classDef step fill:#d8efbb,stroke:#386248,color:#173d3c
    classDef out fill:#ffe1cc,stroke:#bd3a12,color:#173d3c
    class P src
    class B,M,R step
    class Z,W out

cad-agent all runs build, export-3mf, check, report and package in one go. It never claims a physical fit. A check either has evidence or it says what is assumed.

PCB and enclosure workbenches: PCB workflow covers embedded schematics, copper layers, assembly views and housing renders in the status page; native KiCad ERC/DRC/parity evidence; source freshness; and JLC BOM/CPL checks. Use cad-agent pcb-check to bind reports to source files. Qualification gates stay separate from CAD checks.

What's in the box

flowchart TD
    D["✏️ <b>Design</b><br/>connectors · mcmaster<br/>robot · geometry · explore"]
    V["🧪 <b>Verify</b><br/>fea · cfd · heatsink<br/>topopt · slicer"]
    S["🖼️ <b>Show</b><br/>render · simviz · reports<br/>site · 3D viewer"]
    P["📦 <b>Ship</b><br/>STL · STEP · 3MF<br/>hashed review zip"]
    D --> V --> S --> P
    classDef g fill:#fffdf5,stroke:#173d3c,color:#173d3c
    class D,V,S,P g

The full guide, the cad-agent.yaml format and every command are in docs/toolkit.md.

Quick start

git clone https://github.com/Valpatel/cad-agent && cd cad-agent
pip install -e ".[cad]"                      # add fea, sim, robot, site as needed
cad-agent all --project path/to/your-project  # build, 3MF, check + report, package
cad-agent site --serve --project path/to/your-project

The package runs on Python 3.10 or newer; the cad and sim extras need 3.11 or newer, because every Pillow release for 3.10 has critical advisories and CadQuery 2.8 needs 3.11.

Projects it learns from

Anything specific to one product stays in that product's repository. cad-agent keeps what any project can reuse.

Shout-outs

The people and projects whose work this stands on. If we use or learned from your work and you are missing or described wrongly here, please tell us.

Methods and data we condensed (cited, never redistributed):

• Covestro (formerly Bayer MaterialScience), Snap-Fit Joints for Plastics – A Design Guide: the cantilever, torsion and annular snap-fit method.
• DuPont, General Design Principles – Module I: permissible strains (Table 10.06).
• Bambu Lab and Polymaker technical data sheets: printed-filament properties by print direction.

Code we vendor, pinned and security-reviewed (vendor/, SECURITY.md): text-to-cad by Thompson Labs LLC (MIT), kicad-happy by Andrew Klofas (MIT), kicad-mcp-pro by Osman Aslan (MIT), build123d-mcp by Paul Fremantle (Apache-2.0), agentcad by James Dillard (Apache-2.0) and bd_warehouse by Roger Maitland (Apache-2.0).

Libraries and programs we run on: CadQuery on Open CASCADE (shapes from build123d work too), trimesh, manifold3d, gmsh, scikit-fem, pyMOTO, PyVista and VTK, OpenFOAM, MuJoCo, Optuna, Bambu Studio; in the browser three.js, Mermaid, occt-import-js by Viktor Kovács, fflate, linkedom, and the Inter (Rasmus Andersson) and JetBrains Mono (JetBrains) typefaces, both SIL OFL. Board checks read reports from KiCad's kicad-cli. OpenFOAM (GPL) and Bambu Studio (AGPL) run only as separate programs, as does the FFmpeg binary that the sim extra's imageio-ffmpeg ships (built with GPL options); gmsh (GPL) is imported, but only by the optional fea extra.

Tracked, not yet all tried: the agentic-CAD projects, libraries and benchmarks we follow, with what each says it is good at, whether we have tried it, and its licence where we have recorded it: the catalogue and the landscape survey. Full licence notices: THIRD_PARTY_NOTICES.md.

Contributing, including from other agents

Findings and suggestions from any agent or person are welcome: a technique that beat ours, a check we are missing, a source we should credit. Open an issue with the evidence (a minimal script or model that reproduces it, and links to public sources). Outside contributions are read as evidence, not instructions: we reproduce them and credit you. Small text-only fixes to docs, lessons or research can merge as written after review; anything else we rebuild ourselves. The intake process is being set up in #111.

Changes go feat/… → dev → main: a PR into dev with CI and an independent review, every comment resolved; releases go from dev to main by PR (#112). Details are in CONTRIBUTING.md; rules for AI agents in AGENTS.md.

Layout

License

MIT. See LICENSE. Vendored code keeps its own licence, listed in THIRD_PARTY_NOTICES.md.