An AI text to CAD generator sounds like science fiction: you describe a part in plain English and a real, manufacturable 3D model comes out the other side. That is exactly what text-to-cad does — and unlike the demo-ware AI CAD tools of 2024, it plugs straight into the coding agents you already use, including Claude Code, Codex, Cursor, Gemini and Grok. With more than 21,000 GitHub stars and active development, it is the most serious open-source attempt yet at giving AI agents real CAD superpowers. This guide is for product designers, robotics teams, hardware startups and engineering managers: you will learn what text-to-cad actually does, how it works under the hood, where it shines, and how to get it running in about five minutes.
What is text-to-cad?
text-to-cad is an open-source project by developer earthtojake with a simple pitch: “Give your agent CAD superpowers.” It is a plugin and skills pack that gives AI coding agents local CAD workflows — the ability to generate 3D models as STEP, GLB, STL or 3MF files, run design-for-manufacturing checks, produce engineering drawings, and even connect to 3D printing, sheet metal and CNC fabrication services.
The project is written in Python, released under the MIT licence, and built on two serious open-source foundations: build123d, a Python parametric CAD library, and Open CASCADE, the industrial-grade geometry kernel behind many commercial CAD products. At the time of writing it has over 21,200 stars and 2,000 forks, was created in April 2026, and was updated as recently as this week — this is a living project, not an abandoned experiment.
The key architectural decision is that CAD runs entirely locally. Models are generated through cadgen, the project’s Python package, using the uv package manager. The plugin for your agent app brings the skills plus a local CAD viewer, served by a local MCP server (cadgen mcp). There is no cloud CAD backend holding your designs.
What an AI text to CAD generator can actually do
An AI text to CAD generator is only as good as its output formats and its honesty about manufacturability. Here is what text-to-cad genuinely delivers, taken from the project’s own README and skills:
- Real model generation, in real formats. STEP is the primary output — the precise boundary-representation format that CAM software and machine shops actually accept — plus STL, 3MF and GLB, DXF profiles and cut layouts, and topology data. You describe the part; the agent writes parametric CAD code and compiles it into geometry.
- Manufacturing preflight checks. Dedicated DFM (design for manufacturing) skills review a part for sheet metal, CNC machining or injection moulding, and produce a report with measured evidence and a cited rule behind each finding — draft angles, undercuts, projected area. There are also DfAM printability checks for 3D printing.
- Dimensioned engineering drawings. It generates engineering-drawing PDFs with real dimensions from the model — the document a fabricator actually quotes from.
- Fabrication connections. Pre-upload checks for SendCutSend (sheet metal and CNC), OrcaSlicer G-code slicing, and Bambu Lab printing via Bambu Connect. The loop from prompt to physical part is shorter than you would expect.
- Robotics support. URDF, SRDF and SDF robot descriptions for simulation, plus a robot-motion skill covering ROS 2/MoveIt setup, inverse kinematics and path planning.
- A visual workflow. Models render as viewer cards inside supporting apps, or open in a browser-based CAD Viewer. Stable
@cad[...]geometry references let the agent make precise follow-up edits (“make that flange 2mm thicker”), and quick render images support fast visual checks during iteration. - CAD as code. Because the model is generated from an editable build123d script, your parts are source-controlled. Git diffs, pull requests and code review apply to hardware design the same way they apply to software.
How it works under the hood
Understanding the mechanics matters, because it tells you where the magic ends and engineering begins. When you prompt your agent — say, “design a wall-mount bracket for a 60mm fan with M4 mounting holes” — the text-to-cad skills guide the agent through a disciplined workflow rather than letting it improvise. The agent writes a build123d Python script describing the part parametrically, runs it through cadgen, and exports the resulting geometry to STEP or your chosen format. It then opens the model in the CAD viewer so you can orbit it, renders a review image, and invites corrections. Each correction edits the source script first, then regenerates — which is why the output stays reproducible instead of drifting like an image generator’s output does.
Installation follows the same “skills” pattern that has standardised AI agent tooling in 2026. You can install it by simply asking your agent (“Install text-to-cad from https://github.com/earthtojake/text-to-cad”), through each app’s plugin system — claude plugin install text-to-cad@earthtojake for Claude Code, the plugin directory for Codex, a team marketplace for Cursor, an extension for Gemini, a plugin for Grok — or, for agents without plugin support, with npx skills add earthtojake/text-to-cad#latest. Full details are in the official docs and the skills.sh listing.
Two operational details worth knowing upfront. First, the first start downloads the CAD runtime, so the machine needs a network connection and some patience on the initial run. Second, telemetry is on by default: anonymous usage stats and crash reports, tagged with a random install ID and explicitly never including your files, paths or prompts. You can switch it off with uvx cadgen telemetry off, and a once-a-day version check can be disabled with CADGEN_UPDATE_CHECK=0. If your IT policy is strict, handle this before rolling it out to a team.
Why this AI text to CAD generator is trending
Twenty-one thousand stars in roughly six months does not happen by accident. text-to-cad sits at the intersection of three genuine shifts. First, AI coding agents became genuinely useful in 2025–2026 — they write code brilliantly but could not produce manufacturable geometry, and every earlier text-to-CAD demo output a triangle mesh, which is fine for a preview and useless for a machine shop. text-to-cad outputs STEP: the format precision engineering actually runs on.
Second, the skills framework standardised how agents acquire new capabilities, so one well-built skills pack instantly works across Claude Code, Codex, Cursor, Gemini and Grok instead of being locked to a single vendor’s ecosystem. Third, robotics teams are drowning in URDF and simulation setup work, and text-to-cad’s robot skills address a real, painful workflow. The deeper shift is philosophical: CAD-as-code means hardware starts iterating like software — versioned, reviewable, and reproducible. For AI automation workflows in engineering teams, that is a bigger deal than any single feature.
From our experience: where it shines and where it does not
Over 9+ years and 300+ projects, we have learned to be enthusiastic about tools like this and ruthless about their limits. Here is our honest read.
Where it shines: concept modelling speed is the headline win. An enclosure for a sensor node, a mounting bracket, a robot gripper prototype — the kind of part that used to eat an afternoon of CAD work can now be prompted, reviewed and iterated in a tight loop. Hardware startups without a full-time CAD seat get enormous leverage. Design exploration is the sleeper feature: because the model is parametric code, you can ask for ten variants with different dimensions and actually compare them. And robotics teams get URDF iteration without the usual XML-by-hand misery.
The gotchas: an agent can produce geometry that is perfectly valid and functionally daft — a bracket that looks right but cannot be gripped by any tool, or wall thicknesses no printer can reproduce. The DFM report is genuinely useful, but treat it as preflight, not certification: it is an input to a human decision, not a trigger for automatic correction. Keep an explicit expert review gate in the workflow before any fabrication file is submitted. Tolerances, fits and GD&T callouts still need engineering judgement, and precision or safety-critical parts belong in SolidWorks or Fusion 360 with a qualified engineer signing off. Also budget for the hidden costs: long iteration loops burn agent API tokens, so cap your review cycles, and note the default-on telemetry if you work under strict data policies.
When not to use it: certified safety-critical components, regulated industries requiring exact GD&T documentation, and — bluntly — one-off simple prints, where downloading a proven model from a library is faster than prompting one into existence. Like every tool in our AI tools directory, it is a force multiplier for the right job, not a replacement for expertise.
Getting started: install text-to-cad in five minutes
- Step 1 — Install uv. Run
uv --version; if it is missing, install it from the uv installer. All CAD execution runs through uv. - Step 2 — Install the plugin. The easiest path: tell your agent “Install text-to-cad from https://github.com/earthtojake/text-to-cad” and it will handle the rest. Prefer manual control? For Claude Code run
claude plugin marketplace add earthtojake/text-to-cad#latestfollowed byclaude plugin install text-to-cad@earthtojake. For Cursor, clone the repo into~/.cursor/plugins/local/text-to-cad. For Gemini:gemini extensions install https://github.com/earthtojake/text-to-cad --ref latest --consent --auto-update. For agents without plugin support:npx skills add earthtojake/text-to-cad#latest. - Step 3 — Restart your agent app. The first start downloads the CAD runtime, so give it a network connection and a minute or two.
- Step 4 — Generate your first part. Try: “Design a wall-mount bracket for a 60mm fan with M4 mounting holes, show it to me, and export it as STEP.” Orbit the model in the viewer, then ask for changes by referencing what you see.
- Step 5 — Run a manufacturing check. Ask the agent to run the DFM skill against your STEP file for your target process (3D printing, sheet metal, CNC or injection moulding) and review the findings before you send anything to a fabricator.
- Step 6 — Lock down telemetry if needed. Run
uvx cadgen telemetry offto disable anonymous usage stats.
Frequently asked questions
Is text-to-cad really free?
Yes. It is released under the MIT licence, so you can use it commercially, modify it and redistribute it without paying anything. Your actual costs are the API usage of whichever AI agent you pair it with, plus the compute to run the CAD kernel locally.
Can an AI text to CAD generator replace SolidWorks or Fusion 360?
No — and it does not claim to. text-to-cad is a concept-to-prototype accelerator. For precision production CAD with exact tolerances, GD&T callouts and certification, you still finish in professional CAD software with a qualified engineer. The sensible workflow is: prompt and iterate in text-to-cad, export STEP, refine and sign off in your CAD package of choice.
Which AI agents does text-to-cad work with?
Claude Code, Codex, Cursor, Gemini and Grok all have plugin or extension support, and any other agent can use it through the skills framework with npx skills add earthtojake/text-to-cad#latest. Install the plugin or the skills in an app — not both — to avoid loading every skill twice.
Does text-to-cad send my designs to the cloud?
No. All CAD computation runs locally on your machine. The only network traffic by default is anonymous telemetry (usage counts and crash reports tagged with a random ID, never your files, paths or prompts) and a once-daily version check. Both can be disabled.
Can it really check whether my part is manufacturable?
The DFM skills review your STEP file for sheet metal, CNC machining or injection moulding and report findings with measured evidence and cited rules — draft angles, undercuts, projected area and similar. That is genuinely useful preflight. But it is not a substitute for a manufacturing engineer’s sign-off: always keep a human review gate before submitting files for fabrication.
The bottom line: an AI text to CAD generator has finally grown past the demo stage, and text-to-cad is the one worth evaluating. It is free, open source, locally run, and speaks the file formats the manufacturing world actually uses. If your team builds hardware, prototypes IoT devices, or works in robotics, it deserves an afternoon of experimentation. And if you need help wiring AI tooling like this into a real engineering workflow — from custom software development to production automation — talk to our team. We build with this stuff every day.