earthtojake/text-to-cad

The problem

Engineers who want a part modeled usually open a CAD GUI and click through sketches and features. Agents can write text fast but have no way to produce a real, manufacturable solid: no geometry kernel, no export pipeline, no way to check whether a part can actually be printed or machined. text-to-cad installs a set of agent skills plus a local server around cadgen, a Python package built on build123d and Open CASCADE, so the agent can drive real CAD operations and write STEP, GLB, STL or 3MF files instead of describing them.

What's different

Most text-to-3D tools target visually plausible meshes, which is fine for art but useless for a bracket. This one aims at engineering geometry: parametric solids through a B-rep kernel, design-for-manufacturing checks, engineering drawings, and handoffs to 3D printing, sheet metal and CNC services. The distribution model is also unusual: instead of asking you to write code, it ships as a plugin or skill set for Claude Code, Codex, Cursor, Gemini and Grok, with a local MCP server providing a viewer. Compared with scripting build123d by hand, you give up fine control over each feature call; compared with a hosted text-to-CAD service, you keep the whole loop local and the files on your disk.

Who it's for

Mechanical engineers, robotics hobbyists and hardware teams already working inside a coding agent who want quick parametric parts without opening Fusion or SolidWorks. It suits small parts, fixtures, brackets and robot descriptions. People who need assemblies, tolerances, GD&T or production-certified drawings should stay with a conventional CAD package; agent-generated geometry still needs human review before it is cut or printed. Teams that cannot send prompts to a third party should check the telemetry story first.

Cost of adoption

You need Python 3.11+, Node.js 20+, and uv installed; the first start downloads the CAD runtime, so network access is required. Installation differs per agent, and the README warns that installing both the plugin and the skills duplicates every skill. Telemetry is on by default, tagged with a random ID and claimed to exclude files, paths and prompts, with an opt-out command. The project is MIT licensed and young, so expect install-path churn: the marketplace was already renamed once, and retired skills are not removed automatically.

Verdict

This is a brand-new project, created in April 2026, that has already pulled roughly 22.9k stars and added about 6.5k in the past month, ranking #12 on HotRepo's monthly board. That trajectory is steep enough to look like a spike rather than a settled tool, and its Hacker News moment in May 2026 is already months old. The approach is genuinely useful: local, kernel-backed geometry for agents is a real gap. But the churn in install paths, telemetry defaults and unproven manufacturing correctness mean you should treat it as an experiment on side projects, not as a replacement for your CAD seat.