img2cad

Turn a photo, sketch or engineering drawing into editable, coloured CAD — with your coding agent.

Website & 3D examples · Install · Commands · Examples · How it works

img2cad gives Claude Code, Codex, Cursor, Gemini CLI or any MCP host a staged, checked workflow for image → CAD. The agent reads the image, records what it sees and plans the parts, writes build123d source, and the local runner builds it, measures it, and overlays it on the image — refusing to move on until geometry, dimensions and a visual review pass.

You get a parametric model.py you can keep editing, plus STEP (named, coloured parts), GLB, STL and 3MF.

Highlights

• 🧭 Agent workflow, not a one-shot guess — observe → plan → build → compare → revise, with hash-bound gates and resumable revisions.
• 📐 Real dimensions — fit the photo's camera to known features (img2cad camera), register the model on the image, trace outlines in millimetres.
• 📷 Any input — product photos, several photos from any angle (visual hull from fitted cameras), hand sketches, dimensioned drawings.
• ✂️ Masks in one click — colour seeds, polygons, or optional SAM 2 segmentation.
• 🎛️ Live parameter panel — img2cad view model.py opens a 3D viewer with sliders; changes rebuild locally and save back to the source.
• 🏭 Ready to make — 3D-print check, slicer hand-off, CNC / sheet-metal DFM, engineering drawings (SVG/DXF), standard parts library, URDF for robots.

Install

In your agent

{
  "mcpServers": {
    "img2cad": { "command": "uvx", "args": ["--python", "3.13", "img2cad", "mcp"] }
  }
}

Command line

pip install img2cad              # or: uv tool install img2cad
img2cad doctor                   # checks the CAD stack (build123d / OCP)

uvx img2cad doctor               # run without installing
pip install "img2cad[sam]"       # optional: SAM 2 segmentation for `img2cad mask --sam`

Python 3.10–3.13. Latest from GitHub: pip install git+https://github.com/YUI-TIEN/img2cad.

git clone https://github.com/YUI-TIEN/img2cad && cd img2cad
python3 -m venv .venv-cad && .venv-cad/bin/pip install -r requirements-cad.lock -e .
.venv-cad/bin/img2cad doctor

Quick start

Give your agent an image and ask:

Turn mug.jpg into editable CAD. It is 95 mm tall. Use img2cad.

The agent follows skills/img2cad/SKILL.md: it runs img2cad workflow init, records observations and a part plan, writes model.py, builds it, compares it with the photo, and revises until the review passes. Open the result:

img2cad view runs/mug/model.py      # 3D viewer + parameter sliders

Commands

img2cad <command> --help shows every option. Full CLI reference: docs/WORKFLOW.md.

Examples

Each one keeps the whole chain — reference, observations, plan, model.py, comparison and review. Browse them in 3D on the website.

How it compares

Text-to-CAD tools start from words; mesh generators give you triangles. img2cad starts from your image and ends at editable B-rep source with measured checks — a workflow your agent runs, not a hosted model. See docs/DESIGN.md for the design and related projects.

Documentation

• docs/DESIGN.md — problem, design principles, acceptance criteria
• docs/WORKFLOW.md — workflow CLI reference and learning material
• skills/img2cad/SKILL.md — the agent contract · MODELING.md · LESSONS.md
• skills/cad-output/SKILL.md — drawings, printing, DFM, parts, URDF
• benchmark/ · CHANGELOG · ROADMAP · CONTRIBUTING · SECURITY

Status: beta. Reviews are agent reviews (human_acceptance: not_verified); check the measured geometry and previews before you manufacture anything.

License

MIT