Overview
ldraw-nova is an experimental agent tool for designing buildable LEGO models in the LDraw language. A user supplies a model idea and guides the agent while it constructs a digital model, combining generative AI with a low-level 3D CAD format focused on parts, positioning, and rotation.
The project grew from experiments with ChatGPT and Claude writing LDraw as they would other programming languages. Its goal is to explore whether agentic language models can design objects intended to become physical builds.
Key Features
The completed workflow can produce several related artifacts:
- LDraw source code for the generated model
- Multiple presentation modes, including a 3D viewer, 3D player, images, and interactive VR on Meta Quest 3
- An editable glTF file in GLB format, with metadata stored as Blender Custom Properties
- The chat history and the agent's recorded thinking process
Part and example-model discovery uses semantic search with reranking through jev-rerank and Jev System One. If that reranking service is not configured, the agents fall back to full-text search, which may produce weaker model results.
How to Use
The web application runs through Docker and requires Git plus two sibling repositories: ldraw-nova and ldraw-nova-docker at matching tags. After cloning both, build the Docker image and start the services with Docker Compose.
The initial image build requires about 5 GB of disk space. The app exposes both HTTPS and plain HTTP browser access; HTTPS is required for the Meta Quest 3 VR experience, while HTTP avoids self-signed certificate warnings.
Deployment Considerations
The application has no login, so it should only be operated on a trusted network. Other devices on that network can access it through the host computer's local IP address, enabling a headset or another browser to connect to the running instance.