The LDraw Nova project, developed by Antelo, utilizes AI agents to design complex LEGO structures consisting of over 2,000 pieces. The system generates LDraw-compatible files, which can be opened in software such as LeoCAD, LDView, and Studio. The process involves an agent creating a JSON plan, writing a Python program, and iteratively refining the design through image rendering.
The tool operates as a Docker web application compatible with AI providers including OpenAI, Anthropic, and OpenRouter. It supports models like Claude Opus 5.5 and Astra. The project currently lacks physics modeling for stability, and running the tool incurs costs, with an estimated $5 per Technic mechanism generated by the Astra model.
While the project currently exists only in CAD format, the developer is exploring 3D printing options. Integration with Model Context Protocol servers like Kiln and OctoEverywhere allows AI agents to manage 3D printing tasks. The software also supports optional semantic search via the Jev System One model to improve part selection.
LDraw Nova differs from the BrickGPT project, which was trained on over 47,000 LEGO structures and incorporates physical stability checks. Although the tool currently requires high-end AI models for accurate results, the developer aims to enable iterative building for lower-end agents. VR support for the Meta Quest 3 headset is currently available but remains in an early, unstable state.
