Open-Source Project ldraw-nova Lets AI Agents Design Buildable LEGO Models in LDraw

A developer has released an open-source project called ldraw-nova that lets AI agents design physical LEGO models, outputting files in LDraw, the long-established assembly-style language used for LEGO CAD work. The tool was shared on Hacker News under the title "Show HN: Made an open-source Lego AI generator," with the code hosted at github.com/anteloc/ldraw-nova.

The premise is straightforward: a user gives an AI agent a model idea, guides it through the process, and receives a finished LDraw model. According to the project's documentation, the output includes the model's LDraw source code, multiple viewing options (a 3D viewer, a 3D player, and VR support for Meta Quest 3), a Blender-editable glTF file, and a record of the agent's chat history and reasoning process.

The creator says the project went through several earlier, less successful iterations — including ldbuilder-ai, py2bricks, and py4bricks — before arriving at its current approach. Two conclusions from that earlier work shaped the design: AI agents struggle with the geometry math required to position LEGO parts directly, but they are much better at writing Python code that generates that math. Agents also learn more effectively from Python code that produces models than from inspecting LDraw files themselves, which the creator describes as geometrically unintuitive for language models.```

That led to the current design, in which an agent doesn't place parts directly. Instead, it gathers information from documentation, prior examples, and planning steps; builds a detailed plan describing a model's geometry; and writes a Python generator script that, when executed, produces the final LDraw file. The creator describes the result as "a compiler of sorts" — an agent that writes a generator that produces a 3D model in an assembly language.

To find parts and reference models, agents rely on a semantic search and re-ranking tool called jev-rerank, also built by the project's author, which is powered by an AI model described as TypeSafe's Jev System One. Users need a TypeSafe API key for this re-ranking search to work; without one, the system falls back to full-text search, which the creator says may produce weaker results.

The creator credits the project's breakthrough to newer AI models, writing that the tooling had proven difficult to get right until "GPT-6 Astra and Claude Opus 5.5 arrived." These model names, along with other specifics in the writeup, could not be independently verified from the material provided and should be checked against the primary GitHub source before being treated as confirmed.

The software runs as a self-hosted web app via Docker, requiring two sibling repositories — ldraw-nova and ldraw-nova-docker — cloned at matching version tags. Setup instructions describe building a Docker image (requiring roughly 5GB of disk space on first build) and accessing the app over HTTPS on port 8443 for VR support, or plain HTTP on port 8765. The app has no login system, and the documentation warns it should only be run on trusted networks.

The project remains an early release with acknowledged limitations. The documentation lists several open issues: VR support on Meta Quest 3 has performance and handling problems, only expensive high-end models can currently generate large, correct outputs, the generation process is slow, and certain model categories — including minifigs, Technic machines, and spaceships — are described as not yet well supported. The creator also says the tooling and documentation need adaptation to work better with smaller, lower-cost AI models.

The author credits several members of the broader LDraw community, including contributors to LDView, LeoCAD, LDCad, ldraw.rs, and pyldraw3, for tools and resources used in building the project.