Text to CAD is a library of agent skills for computer-aided design, engineering and manufacturing — packaged instruction files that let an AI coding agent produce real mechanical geometry, robot description files and printer-ready toolpaths instead of only describing them. It is aimed at mechanical engineers, robotics developers and 3D-printing tinkerers who already drive an agent in their editor or terminal and want that agent to reach into hardware work rather than stopping at source code.
What it does
Most coding agents are fluent in text files and helpless in front of a solid model. They can write a Python script about a bracket, but they cannot hand you a part file an engineer or a printer will accept. This repository fills that gap by shipping ready-made skills an agent loads and then uses.
The pieces the project advertises are:
- A CAD skill that turns a plain-text description into geometry and exports it as STEP, STL and 3MF files.
- Skills that write robot description files in the URDF and SDF formats used by simulators and robotics stacks.
- A CAM side that slices a mesh into G-code a 3D printer can actually run.
The demo animation at the top of the README shows the CAD skill generating and previewing geometry straight from a prompt, which is the clearest statement of the project's ambition: one agent taking a request from idea to a physical, printable part, with the intermediate files real enough to open in other tools.
STEP output matters more than it first looks. STL and 3MF are mesh formats — fine for printing, awkward for editing — while STEP is the interchange format that downstream CAD and CAM packages read. Emitting all three means the agent's output can go to a printer or back into a conventional engineering workflow.
How it works
This is not an application with a window, and not a hosted service. It is a collection of skills laid out in a skills/ directory, each with a SKILL.md file — a Markdown document that tells the agent what the skill is for and how to use it — alongside the Python code and dependencies that do the geometry work. The badges in the README point at skills/cad/SKILL.md and skills/cad/requirements.txt, and at a separate skills/urdf tree, so the skills are separable rather than one monolith.
Python is the main language and the README asks for Python 3.11 or newer. The repository runs a GitHub Actions test workflow on its main branch, and the code is MIT licensed, so it can be vendored into commercial work without friction. Because each skill is fundamentally a Markdown instruction file plus a Python environment, the same library is not tied to a single agent product — whatever agent can read a skill file and run Python is the intended host.
Getting started
The README's entry point is a documentation site at texttocad.dev rather than a one-line install command, so that is where to go for the current setup steps and the list of skills. From the repository itself, the practical prerequisites are a Python 3.11+ environment and the dependencies listed under the CAD skill; from there you point your agent at the skill directory you want it to use.
There is also a Discord server linked from the README, which is worth noting for a project this young: the repository was created in April 2026 and was still being pushed to in September 2026, so the fastest route to an answer about a rough edge is probably the chat rather than the issue tracker. With roughly 15,000 stars and 1,500 forks it has attracted a large audience quickly, but the star count says more about interest in the idea than about how settled the API is.
When to use it / when not
Reach for this if you already work with an agent and your work touches physical parts: enclosures, brackets, fixtures, robot links, anything you were going to print or model anyway. The appeal is that a text description becomes a file you can inspect, edit and manufacture, and the URDF and SDF skills extend the same idea to robot models that simulators can load.
Skip it if you do not run a coding agent at all — there is no standalone editor here, and nothing to click. Skip it, too, if you want a replacement for parametric CAD in serious design work: generated geometry still needs a human to check dimensions, tolerances and fits before anything gets cut or printed, and the project makes no claim to engineering validation. And because it is only months old, expect the skill layout and interfaces to keep moving.
Anyone building agent tooling around hardware should take this repository seriously, because it is one of the few that treats CAD, CAE and CAM as things an agent can be given competence in rather than subjects it can only discuss. Mechanical and robotics engineers get the more immediate payoff: a way to go from a sentence to a STEP file or a sliced print without leaving the agent they already use. Treat the output as a first draft to be checked, not as a finished design, and it is a genuinely useful addition to a hardware workflow.