text-to-cad does not ship a text-to-CAD model. It ships eleven agent skills, each a SKILL.md workflow, that turn a general coding agent into one that writes real mechanical geometry and carries it all the way to a physical printer. Published as "CAD Skills," the repo installs with npx skills install earthtojake/text-to-cad or as a native plugin for Codex and Claude Code.
The core CAD skill takes a plain-language or image request and produces a STEP file, with STL, 3MF, and GLB as alternate exports. That is the entry point. The more interesting part is what sits downstream of it.
A chain from prose to a physical part
Read the skill table as a pipeline and the design becomes clear. The CAD skill makes the solid. The DXF skill flattens it into 2D drawings: profiles, gaskets, cut layouts. The SendCutSend skill checks those DXF and STEP files before you upload them to the fabrication service. For 3D printing, the G-code skill slices supported meshes into printer-profiled .gcode using real slicer CLIs, not a reimplementation of one. The Bambu Labs skill then dry-runs, uploads, and cautiously starts a print job on a local Bambu Lab printer from that validated gcode.
So the arc runs from a sentence to a running print, with a validation step before anything gets cut or extruded. Two supporting skills round it out: step.parts finds off-the-shelf STEP parts like screws, bearings, motors, and connectors, and CAD Viewer opens local browser previews for CAD, G-code, and robot files so a person can look before committing.
A second chain for robots
The repo carries a parallel track for robot-description files. The URDF skill writes robot structure: links, joints, limits, inertials, meshes. The SRDF skill layers MoveIt planning groups, end effectors, poses, and collision rules onto that URDF. The SDF skill builds simulator models and worlds with frames, physics, sensors, and lights. These are the standard formats mechanical and robotics work actually runs on, and the demo gifs in the README show a URDF example and an SRDF/MoveIt2 inverse-kinematics example rendered in the same CAD Viewer.
An eleventh skill, Implicit CAD, is marked experimental. It builds browser-native models from GLSL signed-distance fields and renders them with raymarching in CAD Viewer. It sits apart from the STEP-centered main path.
Benchmarks you can read
Most "text to CAD" projects show one hero render and stop. text-to-cad ships ten benchmark tasks, each with the exact prompt and an orbiting output gif. They climb in difficulty: a rectangular calibration block with four through-holes and a chamfer, a circular flange with a six-hole bolt circle, an L-bracket with gussets and two hole directions, a stepped shaft with a keyway, an open-top enclosure with internal standoffs, an aerospace-style clevis bracket with lightening cutouts, a radial-engine cylinder with twelve cooling fins and a 35-degree angled spark-plug boss, a centrifugal impeller with backward-curved blades, a spiral staircase with a helical handrail, and a simplified planetary gear stage with separate sun, planet, ring, carrier, and pin bodies.
Each prompt names dimensions and features, so the benchmark doubles as a spec you can grade the output against. The heavy gif and benchmark assets live in Git LFS and are excluded from default pulls, so a plain clone stays light; git lfs pull --include="benchmarks/**" hydrates them when you want to look.
What it is, and isn't
The README is explicit about scope. This is a skills library, Python 3.11+, MIT licensed, that generates, inspects, sources, slices, and hands off CAD and robot artifacts from local project files. It does not claim a novel model or accuracy numbers. The benchmarks are qualitative gifs, not scored results.
What it offers instead is coverage: the boring, load-bearing formats of mechanical and robotics work, STEP, DXF, gcode, URDF, SRDF, and SDF, wired into a single agent, with a validation gate standing between the geometry and any real hardware run. For anyone who has watched an AI produce a pretty mesh that no slicer or fab shop would accept, that emphasis on real CLIs, real file checks, and a cautious print start is the whole point. text-to-cad treats the output file, not the render, as the deliverable.
GitHub: https://github.com/earthtojake/text-to-cad
Curated by Agent Palisade — practical AI for small and mid-sized businesses.
Top comments (0)