Text-to-CAD

Main approaches#

Text-to-CAD tools reach editable geometry in a few different ways:

• Generate CAD code. An LLM writes a script in OpenSCAD, CadQuery, build123d, or a vendor language such as Zoo's KCL, and a CAD kernel runs it to produce the solid.
• Drive a CAD program. The model calls a CAD program's API step by step, often through MCP, and builds features in a real feature tree.
• Predict CAD operation sequences. Research models such as Text2CAD generate sketch-and-extrude sequences directly from text, trained on datasets of CAD construction histories.

Commercial tools often combine these and do not always publish how their pipeline works.

What makes it different from text-to-3D#

Text-to-3D tools generate a visual shape, usually a mesh. Text-to-CAD aims for B-Rep geometry or code that produces it, so the result has exact faces, real diameters, and edges that can be filleted. A useful quick test is the export menu: STEP or native CAD output points to text-to-CAD, while STL, OBJ, or glTF only points to text-to-3D.

Where it stands#

Vendors and startups are both working on it. Zoo offers text-to-CAD through its Design Studio and API. Autodesk's 2026 Fusion roadmap describes work toward neural CAD experiences that turn natural language prompts into editable design geometry.

Current tools do best on simple mechanical parts such as brackets, plates, spacers, and enclosures. They struggle with assemblies, precise engineering rules like gear geometry, tolerances, and manufacturing constraints. Dimensions stated in the prompt are not always honored, so generated parts should be measured before use.

Using it well#

Specific prompts with units, key dimensions, and features produce better results than vague descriptions. Treat the output as a starting model: open it in your normal CAD program, check it, and add the engineering detail that the prompt could not carry.