Open-source Lego AI agent designs buildable LDraw models — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- ldraw-nova is an open-source Docker-deployed web app where an AI agent takes a prompt, plans the build, iteratively renders images, and outputs LDraw source code plus a Blender-editable .glb file with chat history and agent thinking traces
- Author anteloc built it after three prior attempts (ldbuilder-ai, py2bricks, py4bricks), concluding LLMs do better generating Python code that produces geometry math than doing the math directly — and that agents learn better from Python code than from finished LDraw models
- The project broke through when GPT-6 Astra and Claude Opus 5.5 arrived and could finally 'vibe code' the Python-based tooling correctly after earlier models failed to get it right
- Rather than placing bricks itself, it produces a generator script that when executed produces the .mpd file — the author describes it as 'a compiler of sorts,' with an agent → plan.json → generator.py → model.mpd pipeline
- Output includes a 3D viewer, images, and VR support for Meta Quest 3; semantic search uses the author's own jev-rerank tool backed by TypeSafe's Jev System One AI model, with Full Text Search as a fallback if no TypeSafe API key is configured
- The author flags current limits: VR on Meta Quest 3 has performance issues, only expensive high-end models can generate large correct models, the process is slow, and spaceship generation is weak
Why it matters: By routing geometry math through Python generators rather than asking the LLM to compute it directly, ldraw-nova shows buildable physical-design agents are achievable — but the tool currently requires expensive high-end models (GPT-6 Astra, Claude Opus 5.5) and runs slowly, constraining it to patient hobbyists rather than mainstream use.
Ask SkimNews
