Open-source toolkit LDraw Nova is turning text prompts into sprawling, thousands-of-piece LEGO builds—without touching a single brick—by having frontier AI models generate full CAD instructions instead of hand-placed studs.
As first reported by Tom’s Hardware, developer Carlos Antelo’s project uses large language models such as GPT-6 Astra and Anthropic’s Claude Opus 5.5 to plan and “code” LEGO builds that can surpass 2,000 real-world parts, including a 2,175-piece Sakura Garden generated from a simple prompt asking Opus 5.5 for “the most beautiful model” it could imagine.
Instead of asking the AI to place every brick directly, LDraw Nova treats LEGO design like a programming problem: an agent first writes a JSON plan that describes the overall model and its submodels, then turns that plan into a Python generator script, which finally emits LDraw source—the text-based format where each placed piece becomes a single line of code. That output can be opened in tools like LDView, LeoCAD, and Bricklink Studio, or transformed into rendered images the agent inspects and iteratively tweaks in a loop until the build looks “done,” neatly sidestepping the “evil geometry math” that LLMs struggle with while leaning on their strength in code generation.
Antelo has released LDraw Nova as a Docker-based web app that can talk to multiple AI providers, including OpenAI, Anthropic, and the OpenRouter ecosystem, and the project’s gallery credits each build to the model that created it along with the original prompt. Current showcase builds include Sakura Garden and an unfinished Atlas Crane from Claude Opus 5.5, a Cathedral and Tidal Observatory from Astra, and the Copper Bean apartment complex produced by Opus 5 (not 5.5), all of which exist today only as richly detailed CAD models rather than physical bricks.
Under the hood, Nova can tap an extra decision-making model called jev-rerank, a semantic search system backed by TypeSafe’s Jev System One, to improve part selection when Astra is hunting through thousands of available LEGO elements, though the tool can also fall back to plain full-text search if no TypeSafe key is configured. Jev’s reranking has already been battle-tested in a series of fully automated Pokémon Red playthroughs, and here it helps AI agents pick more appropriate parts for structural details like arches, supports, or Technic linkages while still leaving the final “build brain” to Astra or Opus.
There are clear limits, though, especially for anyone dreaming of printing instructions and dumping out bins of bricks the same day: Nova currently models collisions between parts but does not assess structural stability, meaning a gorgeous digital cathedral might crumble if assembled in gravity-bound reality. Antelo estimates that having Astra design a single Technic mechanism can run around $5 in token costs, and notes that only the newest frontier models reliably produce large, accurate designs, with VR support for Meta Quest 3 still experimental and suffering performance headaches even as the tool can export interactive glTF and .glb files for Blender.
That makes LDraw Nova an interesting counterpoint to work like Carnegie Mellon’s LegoGPT (now BrickGPT), a research project that takes text prompts and explicitly optimizes for physically buildable, stable LEGO structures while emitting instructions and LDraw files. Where LegoGPT focuses on engineering-sound designs trained on tens of thousands of existing creations, Nova leans into open-ended creativity and agent tooling, aiming for a future where even “low-end” AI models can iteratively design minifig-scale scenes and functional Technic engines with the help of Nova’s scaffolding and search components.
For LEGO fans and maker-culture geeks, the result is a new kind of collaboration: you dream up a model, an AI agent writes the code, and out the other end comes a fully editable digital build you can refine, dissect, or eventually translate into a real-world parts list if you’re brave enough to tackle thousands of pieces. Whether Nova’s designs ever become official sets or fan-built MOCs en masse, it marks a fascinating step toward AI systems that don’t just chat or draw, but actually design things that could exist on your shelf—and maybe topple off it if you don’t double-check the physics.
