agentcad
Hand off CAD to your agent.
An open-source MCP server and CLI that lets your coding agent design, render, and export real CAD output.
Open source · Free · No signup · Runs locally
StarNew
Specs and measurements
Renders are useful, but serious CAD work also needs measurement. agentcad now lets agents check the model they actually generated against explicit dimensional requirements.
- Measure— read dimensions, cylinder diameters, axes, and feature counts from STEP geometry.
- Check specs— compare measured features against a JSON checklist before calling the model done.
- Review visually— open Spec check mode in the browser to inspect pass/fail rows on the 3D model.
Try it
Paste this prompt into Claude Code, Cursor, or any coding agent. It sets everything up and designs your first model — no manual install:
Create a Python 3.12 virtual environment, then: pip install agentcad[mcp] agentcad skill install agentcad --help Read the --help output — it's your operational briefing. Then design me a phone stand: a simple angled cradle that holds a phone at 60 degrees. About 80mm wide, 50mm deep, with a 5mm lip at the bottom to keep the phone from sliding. Show me a preview when you're done.
The prompt creates a Python 3.12 environment, installs agentcad with its MCP server, installs the agent skill, and reads agentcad --help — its operational briefing — before designing anything.
Prefer to install it yourself? pip install 'agentcad[mcp]' (Python 3.10–3.12), or with uv: uv tool install --python 3.12 'agentcad[mcp]'.
Source: github.com/jdilla1277/agentcad (Apache-2.0).
See it in action
Watch one Claude session build the 1903 Wright Flyer end to end.
Case study · 2026
234 parts. Built with prompts.
A 3D reconstruction of the 1903 Wright Flyer, faithful to the Christman measured drawings, written in Python with agentcad.
See it in action →Made with agentcad
Each of these was designed by an AI agent using a single prompt.
30 faces, 72 edges, 39cm³ volume, valid geometry
15 faces, 28 edges, 107cm³ volume, valid geometry
59 faces, 140 edges, 19.2cm³ volume, valid geometry
10 faces, 24 edges, 42cm³ volume, valid geometry
How it works
What the agent sees, runs, and hands back.
{
"command": "run",
"status": "success",
"metrics": {
"dimensions": {
"x": 6480.8,
"y": 12293.6,
"z": 1999.9
},
"volume": 305738441.28,
"face_count": 684,
"is_valid": true
}
}
How it works
Designed for human–agent collaboration.
Every run returns metrics and a render. The agent reads both — and catches its own mistakes before they reach you.
See the loop →What it does
Agents write bad geometry on the first try. agentcad gives them a tight feedback loop so they converge on printable geometry without you babysitting.
- Execute
- — run scripts, produce versioned STEP files + geometric metrics
- Render
- — PNG views from any angle for visual verification
- Export
- — STL, GLB, OBJ for 3D printing and web viewers
- Validate
- — pre-execution checks catch errors before anything runs
- Inspect
- — topology report for debugging geometry issues
- Diff
- — compare versions to track design iteration
No boilerplate
Scripts need zero imports. build123d primitives, show_object, and agentcad edit helpers are pre-injected:
box = Box(10, 20, 5) show_object(box)
MCP integration
Add to .mcp.json for Claude Code, Cursor, or Windsurf. Works without a prior install — uvx caches the environment on first run:
{
"agentcad": {
"command": "uvx",
"args": ["--from", "agentcad[mcp]", "python", "-m", "agentcad.mcp"]
}
}Agent skill
agentcad ships with an installable agent skill. Run this in your project and Claude Code auto-discovers it:
agentcad skill install
Source: github.com/jdilla1277/agentcad
Get release updates
Optional. New releases and agent-facing features. We’ll send a confirmation link before adding you.