Show HN: Made an open-source Lego AI generator

What it claims

ldraw-nova gives an LLM agent the tooling, examples and instructions to design LEGO models from real LDraw parts. It ships as a dockerized web app. Each finished build produces:

• LDraw source
• a 3D viewer, a 3D player and VR on Meta Quest 3
• rendered images
• a Blender-editable .glb file, with metadata stored as Blender Custom Properties
• the full chat history and the agent's thinking

How it works

The design comes from three earlier attempts by the same author: ldbuilder-ai, py2bricks and py4bricks. The author drew two conclusions from them:

• Agents should not do geometry math directly. They do better writing Python that computes rotations and positions than emitting that math themselves.
• Agents learn better from generator code than from finished models.

So the agent does not place parts directly, except when prototyping or modifying an existing example. Instead it:

1. Gathers information from experiments, the docs and its own planning.
2. Writes one or more plan.json files describing each model and submodel, including geometry.
3. Writes generate.py, which outputs the .mpd file.

flowchart TD
    P[prompt + instructions.md] --> S{TYPESAFE_API_KEY set?}
    S -- yes --> R[jev-rerank semantic search]
    S -- no --> F[full-text search fallback]
    R --> PL[plan.json]
    F --> PL
    PL --> G[generate.py] --> M[model.mpd]
    M --> V[headless render + collision/gap check]
    V -- adjust --> PL
    V -- done --> D[deliver]

The author also wrote jev-rerank, the parts search tool. Without the key, the page says the fallback "could (maybe) yield worse models".

How to use it

You need Git and Docker. Clone both repos side by side at the same tag. The image build takes a while and needs about 5 GB of disk.

git clone --branch v0.6.0 https://github.com/anteloc/ldraw-nova.git
git clone --branch v0.6.0 https://github.com/anteloc/ldraw-nova-docker.git
cd ldraw-nova-docker
docker compose build
docker compose up -d

• https://localhost:8443 uses a self-signed certificate and is required for VR on Quest 3.
• http://localhost:8765 is plain HTTP; VR does not work over it.
• Other devices on your network can connect using your machine's IP.
• Set TYPESAFE_API_KEY in the app's Settings section.
• Stop the app with docker compose down.

Limits

These are the author's own limitations for this first release:

• There is no login. Only run the app on networks you trust.
• Only expensive, high-end models can generate large, correct models. The models named are Claude Opus 5.5 and GPT-6 Astra. Adapting it for low-end models (e.g. Haiku) is still open work.
• Generation is slow.
• VR on Quest 3 is rough. It has model-handling and performance issues.
• Several model families are weak. Spaceships are "not very good", and minifigs and Technic models still need work.
• Building from manuals only partly works. It does better when manual pages are given as images.
• No benchmarks or success rates are published. Quality evidence is limited to the demo video and example files such as atlas-crane.mpd.
• LEGO trademark. The LEGO Group does not sponsor, authorize or endorse the project.