Your CAD intern finally learned to read blueprints
A skill pack that lets AI agents generate, slice, and hand off real mechanical parts instead of just talking about them.

What it does CAD Skills is a library of agent “skills” — pluggable workflows for AI coding agents — that handle the full hardware design pipeline. Agents can generate parametric CAD models from plain text or images, export to STEP/STL/3MF, write URDF robot descriptions, create SDF simulation worlds, slice meshes into G-code, and even dry-run Bambu Lab print jobs. It is essentially a structured wrapper around OpenCascade (via build123d) and various hardware toolchain CLIs, exposed as installable skills for Codex, Claude Code, and similar agents.
The interesting bit The project treats mechanical design as a conversation: you describe a “rectangular calibration block with four holes” and the agent emits validated geometry, previews it locally, and prepares it for fabrication. The benchmarks are particularly telling — nine progressively complex parts from flanges to planetary gear stages, each with a natural-language prompt and a generated GIF. It is the closest thing yet to “vibe coding” for physical objects.
Key highlights
- Nine benchmarked tasks with natural-language prompts and generated geometry (calibration blocks, impellers, spiral staircases)
- STEP as the primary CAD exchange format, with STL/3MF/GLB export for downstream use
- URDF/SRDF/SDF skills for robotics and simulation pipelines (MoveIt2, Gazebo)
- G-code slicing via real slicer CLIs, plus Bambu Labs printer integration
- Local browser-based CAD viewer for inspecting geometry, G-code, and robot files
- Installable via
npx skills installor native plugins for Codex and Claude Code
Caveats
- Heavyweight benchmark assets live in Git LFS and are excluded by default; you must run
git lfs pull --include="benchmarks/**"to see them - The repo is mostly skill definitions and glue code around existing tools (build123d, OpenCascade, slicers); the heavy lifting is not original
- Requires Python 3.11+ and various toolchain dependencies; not a lightweight drop-in
Verdict Grab this if you are building AI agents for hardware or robotics and need structured, benchmarked CAD generation that goes beyond mesh soup. Skip it if you want a standalone end-user CAD tool — this is infrastructure for agent builders, not a replacement for Fusion 360.