What they did
antelocnova began around December 2025 by having ChatGPT and Claude emit LDraw source, where each line places one Lego part, and iterating every time the models came out broken. Using GPT-6 Astra and Opus 5.5 he built a Python toolset, agent instructions and docs so that an agent can write a model, render it from several angles, examine the renders, run collision checks and correct itself. Connector data from LDCad's shadow library lets the tools verify that bricks actually join. It ships as a dockerized web app with OpenAI, Claude and OpenRouter as providers, and he published a gallery of sample models.
How it went
Categories with enough examples, such as houses, come out well but not perfectly; sign flips on rotations are the common failure and the validation cycles are what fix them. His rough cost estimate is about $5 for a Technic mechanism on Astra, with Technic models dearer than brick ones because first placements land further from correct. A test with Qwen 3.8 27B through OpenRouter was very slow and still had defects, though starting in a plan mode helped. The tooling checks collisions only; there is no physics or stability check, and the thread does not establish whether any model was built in real bricks.
Worth knowing
Generation alone is unreliable; renders, collision checks and connector data are what make the loop work, and the author's plan is for cheap models to iterate while expensive ones plan. A commenter reported a parallel result with FreeCAD, freecad-mcp and Claude, three 3D-printed or milled parts that worked on the first hardware revision, with assembly order the weak point.
GitHub