New ldraw-nova tool lets AI agents design 3D Lego models

Developer anteloc has launched ldraw-nova, an open-source agentic framework designed to let artificial intelligence construct physical, buildable 3D Lego models. Built using advanced models like GPT-6 Astra and Claude Opus 5.5, the system bypasses the complex geometry math that typically stymies large language models. Instead of writing raw LDraw CAD code directly, the AI agents generate a plan.json file and a corresponding generate.py Python script, which then executes to produce the final LDraw assembly files.

To run the web application, developers must clone both the ldraw-nova and ldraw-nova-docker repositories side-by-side at tag v0.6.0. The setup runs dockerized, requiring Git and Docker, with the initial build demanding approximately 5 GB of disk space. Once built, the application is accessible locally via port 8765 for plain HTTP or port 8443 for HTTPS, which is required to enable interactive virtual reality viewing on the Meta Quest 3.

The framework leverages semantic search with re-ranking powered by TypeSafe's Jev System One AI model, utilizing the author's jev-rerank tool to help agents find appropriate Lego parts. If users do not provide a TypeSafe API key, the system falls back to a standard Full Text Search strategy, which may produce lower-quality models. During the build process, the agent iteratively plans its construction, renders headless images to inspect its progress, and adjusts part positioning and aesthetics before delivering the final assets.

For practitioners, ldraw-nova provides a complete pipeline for generative physical design, outputting 3D viewer files, VR-compatible scenes, and Blender-editable glTF files in .glb format with custom properties. While the creator notes that generation is currently slow, expensive, and limited to high-end models, future updates aim to optimize efficiency and adapt the tooling for lower-end models like Luna and Haiku.