CAD Agent

Send build123d CAD commands via HTTP to render images, allowing visual iteration on 3D models entirely within a containerized CAD environment.

Security Scan

The skill's behavior (running and building a CAD server container that will execute arbitrary build123d Python sent by the agent) is coherent with a CAD-rendering purpose, but the manifest omits required host tooling and the instructions ask you to build and run code from an external repo—this mismatch and the need to execute unreviewed container code are worth caution.

Purpose & Capability

The SKILL.md describes a CAD rendering server that accepts build123d code over HTTP and returns renders; that matches the name. However, the manifest declares no required binaries or environment variables even though the runtime instructions require git, docker/docker-compose, network access, and a local port. The omission of these required host capabilities in the manifest is an inconsistency.

Instruction Scope

Runtime instructions tell the agent (or operator) to clone a GitHub repo, build a Docker image, run the container, and then send arbitrary build123d Python code to endpoints. Executing arbitrary code inside the container is expected for a CAD runner, but the instructions also encourage running host docker commands (docker run, docker exec) which may require privileged access and can be risky if the image or Dockerfile is untrusted. The SKILL.md also asserts 'never do STL manipulation outside the container' but gives no technical enforcement mechanism—this is guidance only.

Install Mechanism

There is no formal install spec in the manifest (instruction-only). The SKILL.md directs cloning from GitHub (a well-known host) and building a local Docker image. Cloning and building from a third-party repo is a moderate risk: the source should be reviewed (Dockerfile, entrypoint, scripts) before building or running. No opaque or shortened download URLs are used.

Credentials

The skill requests no environment variables or credentials in the manifest, which is consistent with a local/offline CAD container use-case. There are no unexpected credential demands listed. That said, the SKILL.md assumes the operator can bind ports and run Docker, which are host privileges rather than environment variables.

Persistence & Privilege

always is false and the skill is user-invocable and may be invoked autonomously (default). The skill does not request permanent platform-wide presence or claim to modify other skills' configs. The main privilege implication is that running Docker commands grants the container host-level capabilities depending on how Docker is used—this is an operational risk rather than a manifest privilege flag.

What to consider before installing

This skill appears to implement a local CAD render server and expects you to build and run a container from a GitHub repo. Before installing: - Review the repository (especially Dockerfile, entrypoint, and any scripts) to see what the image will run and whether it mounts the Docker socket or host paths. - Verify you have git and Docker/docker-compose available on the host; the manifest does not list these but the README requires them. - Treat the container as running untrusted code: run it with least privilege, avoid mounting sensitive host paths, and avoid exposing the Docker socket to the container. - Understand that the server will execute arbitrary build123d Python you send to it; limit who/what can send code and inspect the server code to ensure it restricts filesystem/network access as appropriate. - If you plan to export or 3D print parts, validate exported files before use and ensure no sensitive data is written out inadvertently. If you want a lower-risk option, request a vetted prebuilt image from a trusted publisher or ask the author for a security review of the Dockerfile and entrypoint.

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

Endpoint	What it does
POST /model/create	Run build123d code, create model
POST /model/modify	Modify existing model
GET /model/list	List models in session
GET /model/{name}/measure	Get dimensions
POST /render/3d	3D shaded render (VTK)
POST /render/2d	2D technical drawing
POST /render/multiview	4-view composite
POST /export	Export STL/STEP/3MF
POST /analyze/printability	Check if printable

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