Zoo's Text-to-CAD turns prompts into editable STEP files

AI can already generate impressive 3D models from simple prompts, but hand one of those files to a mechanical engineer, and they will likely sigh. Most text-to-3D models spit out dense polygon meshes that behave like frozen, uneditable digital clay inside engineering software.

Hardware infrastructure company Zoo wants to bridge that gap. The team just launched the alpha version of Text-to-CAD, a new tool designed to generate genuine, fully editable computer-aided design models from natural language.

Real CAD surfaces instead of gaming meshes

Most existing text-to-3D systems cater to game design. They generate point clouds and polygon meshes, which turn into amorphous, uneditable blobs the second you import them into standard CAD software.

Zoo is taking a fundamentally different technical path by generating Boundary Representation (B-Rep) surfaces. The platform outputs standard STEP files right out of the box, meaning you can drop the generated file into any major CAD tool and immediately adjust dimensions, add chamfers, or cut holes.

Looking ahead, Zoo also plans to output raw code in KCL (KittyCAD Language), allowing engineers to tweak the underlying parametric code directly within Zoo's own Modeling App.

Prompt like an engineer, not an artist

Teaching an AI to design mechanical parts is significantly harder than teaching it to paint a picture. While 2D image models rely on dense datasets with fixed representations, a single CAD part can be built using multiple valid feature trees.

Because of that complexity, Zoo noted that prompting requires a bit of discipline. Typing a generic noun like "car" will likely yield disappointing results. Instead, the model shines when users describe engineering features, such as an "involute helical gear with 36 teeth" or "a plate with 4 holes near each corner and rounded corners."

The tool is currently available via a web interface and a Discord bot. Both applications and the underlying API are open source, letting developers build their own generative CAD workflows.

Why it matters

We see this as a critical turning point where generative AI transitions from visual entertainment into physical manufacturing.

Until now, 3D generative AI has basically been stuck in concept art. Manufacturing does not run on meshes; it runs on precise geometric tolerances, smooth mathematical surfaces, and strict feature trees. By generating native B-Rep geometry and standard STEP files, Zoo is tackling the exact format real-world manufacturing requires.

For engineers and product designers, it is well worth testing today, even in its alpha state. While you shouldn't expect it to design a jet engine, using it to spin up brackets, basic plates, or standard gears could cut down tedious modeling setup. The main challenge will be reliability: in hardware, a hallucinated dimension means a scrapped part on the factory floor. But by pushing the generative conversation toward STEP and parametric code, Zoo is forcing the industry to focus on real manufacturing utility.