LDraw Nova is an open-source toolkit that lets AI agents design LEGO models with thousands of real pieces, outputting detailed LDraw CAD files. According to Tom's Hardware, it works with frontier models such as GPT-6 Astra and Claude Opus 5.5, which generate Python programs that produce the builds without placing every brick manually. The project's gallery includes a 2,175-piece Sakura Garden designed by Claude Opus 5.5 from a single prompt.
The workflow has an agent write a JSON plan describing the model and submodels, then create a Python program that generates the LDraw source. The model is rendered as an image, inspected, adjusted, and rendered again until the agent considers it complete. This approach avoids "geometry math" by relying on code generation. However, none of the designs have been built with real bricks yet; developer Carlos Antelo says he didn't even try.
The tool has notable limitations. It lacks physics modeling, so it handles collisions but not stability, and only frontier models currently produce large, accurate builds. Running it isn't free: Antelo estimates around $5 in token costs for one Technic mechanism. This contrasts with Carnegie Mellon's LegoGPT (now BrickGPT), which was trained on over 47,000 LEGO structures and checks designs for validity and physical stability as it generates them.
Antelo built the project to get AI agents "capable of designing buildable, physical things." He is considering 3D-printing a small model, and the project's roadmap includes minifigures and Technic machines. For now, the bricks exist only in CAD.