Text-CAD connects prompt to 3D print with 11 engineering skills

Hardware prototyping usually means hours of clicking inside SolidWorks, double-checking dimensions with digital calipers, and hunting across the web for random motor brackets.

Text-CAD wants to automate that entire chore. Instead of merely hallucinating a pretty 3D mesh, the project hooks natural language prompts directly to STEP files, slicing tools, and local Bambu Labs 3D printers.

11 engineering skills from prompt to physical part

The framework packages 11 distinct skills to cover the hardware manufacturing pipeline end to end.

Feed it plain text, reference images, or 2D technical drawings, and it outputs production-ready STEP models, with exports for STL, 3MF, and GLB formats. When an assembly requires off-the-shelf hardware, it queries step.parts, an open-source library packed with over 16,000 STEP parts including Adafruit boards, Arduino units, Feetech servos, and standard linear rails.

The pipeline goes well beyond basic modeling. It can export 2D DXF files and validate specs before sending parts to SendCutSend for laser cutting. For 3D printing, it handles G-code slicing and connects directly to local Bambu Labs printers, running a dry-run check before laying down the first layer.

Robotics engineers get native support too, with generation of URDF, SRDF, and SDF formats ready for MoveIt2 motion planning and simulation environments.

Writing Python code instead of raw geometry

Having an LLM output raw geometric binary formats directly is notoriously unreliable. Text-CAD dodges this trap by never asking the model to build CAD binaries directly. Instead, the model writes clean Python scripts using the build123d CAD framework, which then compiles down to STEP.

This gives the language model a structured text layer it already understands well. The engineering harness prioritizes geometric accuracy above all else.

A dedicated inspection script measures faces, alignments, and clearances. Even if deterministic geometric checks pass, the pipeline enforces a visual snapshot verification step. When errors pop up, the agent patches only the faulty code segment and reruns the failing test instead of regenerating the entire model from scratch.

There is a trade-off. By using thin wrappers and progressive reference documentation, the system saves context window space, but its capabilities remain bound to underlying CLI tools like OpenCASCADE, build123d, and local slicers.

Why it matters

We think Text-CAD points to a much smarter direction for generative hardware design. Most text-to-3D models focus on mesh generation for games, producing non-manifold assets that fail the moment you send them to a CNC mill or slicer. Text-CAD sidesteps generative chaos by relying on parametric code.

For makers, indie hackers, and robotics engineers, this dramatically shrinks the friction between an idea and a tangible prototype. Sourcing a servo bracket or designing an enclosure around an existing board becomes a quick prompt and an automated print job.

Its limitation, however, is that the LLM is orchestrating rather than solving geometry. If build123d or the underlying geometry kernel cannot solve a complex fillet or loft, prompting adjustments will only go so far. Still, turning probabilistic text generation into a strict, verifiable engineering workflow is exactly how AI earns a place on the factory floor.