Open-source ldraw-nova lets AI agents build editable LDraw LEGO models with VR output

What ldraw-nova does

A developer posting as anteloc has released ldraw-nova, an open-source project that surfaced on Hacker News's front page, which hands LEGO design work to an AI agent. Give the agent a prompt describing a model and it produces a complete LDraw file, the long-standing format used across the LEGO CAD community.

According to the project's README, the delivered build includes more than the model file. You also get a 3D viewer, an animated player view, rendered images, an interactive VR mode for the Meta Quest 3, and a glTF export in .glb form that opens in Blender with model metadata carried as custom properties. The chat history and the agent's reasoning trail are kept as well, so the design process can be audited afterwards.

How the agent builds

The pipeline is the notable part. The agent starts by reading bundled instructions and LDraw documentation, then plans the parts, submodels and overall aesthetics it needs. From there it loops: render images of the current state, inspect them, adjust positioning and appearance, and repeat until it considers the model done.

Crucially, the agent never hand-places geometry. It writes a plan file in JSON that fully describes the model and its submodels, then writes Python generator scripts that, when executed, emit the actual LDraw source, such as an .mpd file. The developer likens the setup to a compiler: an agent that builds a generator, which in turn builds the model.

The README traces this architecture to earlier experiments (ldbuilder-ai, py2bricks and py4bricks) and two conclusions: agents write Python code that computes geometry far more reliably than they produce geometry math directly, and they learn better from code that generates models than from the finished models themselves. The final tooling was largely built through AI-assisted vibe coding with GPT-6 Astra and Claude Opus 5.5, the developer notes.

Supporting tools include semantic search for suitable parts and example models, powered by jev-rerank, a re-ranking search tool from the same author that relies on TypeSafe's Jev System One model, plus collision and gap detection and headless rendering for self-inspection. Reranked search requires a TypeSafe API key; without one, agents fall back to plain full-text search, which the developer warns may yield worse models.

Running it yourself

ldraw-nova ships as a Dockerized web app. Setup means cloning two sibling repositories at the same release tag, v0.6.0 in the published instructions, and building the image, a first-time job the README says takes a while and needs about 5 GB of disk space. The app then serves HTTPS on port 8443 with a self-signed certificate, which is required for Quest 3 VR, and plain HTTP on 8765, where VR does not work. There is no login, so the developer advises running it only on networks you trust.

What is still rough

As a first release, the project is candid about gaps. VR model handling on the Quest 3 still has significant problems, including performance. Generation is slow, and only expensive high-end models currently manage large models that come out correct; adapting the tooling for cheaper models, with the README naming Haiku and Luna as examples, is on the roadmap. So are new model families such as minifigs, Technic machines and better spaceships, plus building from manuals, which partially works and improves when pages are supplied as images, and finer step-by-step inspection of submodels. The project also credits years of LDraw community work, thanking the maintainers of LDView, LeoCAD, LDCad, ldraw.rs and pyldraw3 by name.

Why it matters

The stated goal is getting agentic LLMs to design buildable, physical things, and ldraw-nova's central trick is a transferable one: when a model struggles with a formal output format, have it write a program that emits the format instead. The results are not throwaway renders either. LDraw files are editable in established LEGO CAD tools, viewable in VR, and map to real parts, so an agent's design can be checked, modified and physically built. The project's own caveat is cost: until the tooling is adapted to smaller models, generating a correct large model means paying for frontier-tier inference.