LDraw Nova, an open tool published by Carlos Antelo, reached hardware coverage on October 4, 2026 with a direct claim: AI agents design LEGO builds using real LDraw catalog parts, including a 2,175-piece cherry garden. The author says he did not try to build any model with physical bricks.
The anteloc/ldraw-nova repository describes the flow. The agent gets a prompt, reads instructions and examples, plans submodels and writes a Python script. That script generates the LDraw file, the language used by programs such as LeoCAD and LDView. The point is to sidestep geometry: the model writes code that calculates positions instead of calculating every rotation itself.
What the tool delivers
- LDraw source;
- views, a viewer and a glTF file editable in Blender;
- agent history;
- a local Docker app on tag v0.6.0, with HTTP and an HTTPS path for Quest 3.
The README states the limits. Large generation still depends on expensive models. VR support has performance problems. Ships and human figures come out worse than structures and vehicles. Antelo cites GPT-6 Astra and Claude Opus 5.5 as the models that closed the tooling. That is the author's statement about his own project, not an independent benchmark.
What is not confirmed
As of this publication, there is no piece-by-piece physical build. An LDraw file can look stable on screen and fail on the table because of collision, gaps or unavailable parts. Tom's Hardware recorded Antelo's line: he "didn't even try." The project video is on YouTube.
The shared point with other recent agent experiments is the shortcut: instead of asking the model for the final part, you ask for a program that generates the part. With LEGO, that still does not replace the build table.
Sources
Image credit: anteloc/GitHub
Transparency: This content was created, edited or reviewed with the help of artificial intelligence. Information was cross-checked with public posts on X and sources available on the internet. Check the original sources for the full context.
Por GeekikiBot
