LDraw Nova gets an AI agent to generate buildable LEGO models in LDraw, and hits 178 stars in 10 days: Reveneau

Why it mattersAn agent that outputs a specification a builder can inspect, correct and take to a 3D printer or a brick drawer removes the common failure of image models producing a picture that nobody can build.

An image model asked to draw a LEGO castle gives you a castle-shaped picture, and nobody can build it. LDraw Nova is an open-source project that gets an AI agent to produce the file a LEGO builder actually opens in LDView, LeoCAD or Studio: a plain-text list of part placements that another program turns into a real model. The agent reads a set of docs and tools, drafts a plan, writes a Python generator, and that generator writes the final LDraw file.

What was shipped

The repository is anteloc/ldraw-nova, created on 23 September 2026, tagged v0.6.0, licensed AGPL-3.0 and written in Python. GitHub lists it at 178 stars this morning. The author posted it to Hacker News the same week as a Show HN, where it reached 137 points and 49 comments on the thread at the time of writing. The project runs as a dockerised web app on a developer's own machine, with no account system, and talks to OpenAI, Claude or OpenRouter under the user's own API keys.

LDraw itself has been around since 1995 and is maintained by the LEGO community rather than by the LEGO Group. A .mpd file is a plain-text sequence of placement lines: part number, colour, rotation, position. That is why the author describes it as "an assembly language (pun intended)" in the README.

How the pipeline works

The agent does not try to place parts directly. The README is explicit that math-heavy coordinate work is where LLMs fail, so the pipeline avoids it by moving the math into Python. The agent reads the project's instructions.md and a set of workflow guides for vehicles, spaceships, Technic structures, mechanisms and modular buildings. It then writes a plan.json that describes the model and submodels, and a generate.py that executes to produce the .mpd. Collision and gap detection, headless rendering and a part-finding tool are provided for the agent to inspect its own work and iterate.

The model combination that produced a working version, in the author's own words, was GPT-6 Astra and Claude Opus 5.5. Earlier attempts in 2025 and the first half of 2026 did not solve the geometry reliably enough. Part search uses the author's own jev-rerank library, a semantic reranker that uses a TypeSafe Jev System One decision model; without a TypeSafe API key, search falls back to full-text search and the README notes that generated models "could (maybe) yield worse models".

Local only, and only on top-end models

A developer with Docker and a provider API key can run the web app on their own machine, type a prompt, watch the agent iterate, and get back a model with its source, a 3D viewer, a Meta Quest 3 VR preview and a Blender-editable glTF export. There is no cloud service: project data stays on the local disk.

The author lists the honest limits in the "Development" section of the README. VR on the Quest 3 has model-handling and performance issues. Low-end models are not yet adapted, and large or correct models today need the most expensive providers to run. Spaceships, in the author's phrasing, "are not very good" yet. Building a model from a step-by-step page manual works partially, and works better when the pages are supplied as images than as text.

Source

anteloc, ldraw-nova on GitHub, repository, v0.6.0, retrieved 2026-10-03.
Hacker News, Show HN: Made an open-source Lego AI generator, by antelocnova, 2 October 2026, 137 points and 49 comments.
anteloc, ldraw-nova sample gallery, model outputs, retrieved 2026-10-03.

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.