LEGO from My Bricks

Photograph a LEGO build: Meta AI recognises it, then shows exactly how to take it apart and what else to build. Snap a photo of a LEGO model you already built. Meta SAM 3.1 outlines it and Muse Spark 1.3 describes what it sees and suggests which of your saved models it is. You confirm, and the app shows the exact parts list, an exploded view, step-by-step disassembly and a disassembly film, straight from the model's real data. Then it turns the same bricks into new models (cat to spaceship) and lists any missing parts as a BrickLink wanted list. It runs inside ChatGPT and Claude as an MCP app (live: https://lego-mcp.vercel.app/mcp). Problem: kids and parents rebuild, trade and reuse LEGO constantly, but instructions only go one way. Taking a model apart cleanly, knowing exactly which bricks it holds, or deciding what else those bricks can become means guessing. A photo alone cannot see hidden bricks, so the honest answer comes from the model's own data, with AI doing what it is good at: recognising and describing. Flow: 1. Photo (up to 3 angles) of a built model in ChatGPT or Claude. 2. Meta SAM 3.1 outlines the model and boxes the visible bricks. 3. Meta Muse Spark 1.3 (structured output) describes colours and parts and picks up to 3 matches only from the user's saved models (server-side allowlist). Marked "AI estimate". 4. The user confirms. No confirmation, no instructions. 5. Exact data: parts list, exploded view, reverse steps, disassembly film, one switch back to Build. 6. Rebuild: new models from the same bricks; close ideas get a BrickLink wanted list. 7. Unknown build (real photo of a dragon): honest "no match", no made-up steps, and how to add it. Honest limits: the positive match in the video uses a render of the built-in cat; the "no match" case uses real iPhone photos. Disassembly order is the build reversed, consistent with the data but not physically tested. Pre-existing work is tagged pre-hackathon in the repo; everything Meta-related was built during the event.