CAD Agent

> Give your AI agent eyes for CAD work.

Description

CAD Agent is a rendering server that lets AI agents see what they're building. Send modeling commands → receive rendered images → iterate visually.

Use when: designing 3D-printable parts, parametric CAD, mechanical design, build123d modeling

Architecture

Critical: All CAD logic runs inside the container. You (the agent) only:

1. Send commands via HTTP
2. View the returned images
3. Decide what to do next

YOU (agent)                     CAD AGENT CONTAINER
─────────────                   ───────────────────
Send build123d code      →      Executes modeling
                         ←      Returns JSON status
Request render           →      VTK renders the model
                         ←      Returns PNG image
*Look at the image*
Decide: iterate or done

Never do STL manipulation, mesh processing, or rendering outside the container. The container handles everything — you just command and observe.

Setup

1. Clone the Repository

git clone https://github.com/clawd-maf/cad-agent.git
cd cad-agent

2. Build the Docker Image

docker build -t cad-agent:latest .

Or using docker-compose:

docker-compose build

3. Run the Server

# Using docker-compose (recommended)
docker-compose up -d

# Or using docker directly
docker run -d --name cad-agent -p 8123:8123 cad-agent:latest serve

4. Verify Installation

curl http://localhost:8123/health
# Should return: {"status": "healthy", ...}

> Docker-in-Docker caveat: In nested container environments (e.g., Clawdbot sandbox), host networking may not work—curl localhost:8123 will fail even though the server binds to 0.0.0.0:8123. Use docker exec cad-agent python3 -c "..." commands instead. On a normal Docker host, localhost access works fine.

Workflow

1. Create Model

curl -X POST http://localhost:8123/model/create \
  -H "Content-Type: application/json" \
  -d '{
    "name": "my_part",
    "code": "from build123d import *\nresult = Box(60, 40, 30)"
  }'

2. Render & View

# Get multi-view (front/right/top/iso)
curl -X POST http://localhost:8123/render/multiview \
  -d '{"model_name": "my_part"}' -o views.png

# Or 3D isometric
curl -X POST http://localhost:8123/render/3d \
  -d '{"model_name": "my_part", "view": "isometric"}' -o iso.png

Look at the image. Does it look right? If not, modify and re-render.

3. Iterate

curl -X POST http://localhost:8123/model/modify \
  -d '{
    "name": "my_part", 
    "code": "result = result - Cylinder(5, 50).locate(Pos(20, 10, 0))"
  }'

# Re-render to check
curl -X POST http://localhost:8123/render/3d \
  -d '{"model_name": "my_part"}' -o updated.png

4. Export

curl -X POST http://localhost:8123/export \
  -d '{"model_name": "my_part", "format": "stl"}' -o part.stl

Endpoints

build123d Cheatsheet

from build123d import *

# Primitives
Box(width, depth, height)
Cylinder(radius, height)
Sphere(radius)

# Boolean
a + b   # union
a - b   # subtract
a & b   # intersect

# Position
part.locate(Pos(x, y, z))
part.rotate(Axis.Z, 45)

# Edges
fillet(part.edges(), radius)
chamfer(part.edges(), length)

Important

• Don't bypass the container. No matplotlib, no external STL libraries, no mesh hacking.
• Renders are your eyes. Always request a render after changes.
• Iterate visually. The whole point is you can see what you're building.

Design File Safety

The project has safeguards against accidentally committing CAD outputs:

• .gitignore blocks .stl, .step, *.3mf, etc.
• Pre-commit hook rejects design files
• User's designs stay local, never versioned

Links

• Repository
• build123d docs
• VTK