PhiloLabs released fable51-worlds on September 2, 2026, an MIT-licensed project where Claude Fable 5.1 agent swarms reconstruct real places as browser-native 3D worlds. The first world is Union Square, San Francisco, rendered as Three.js code with 453 buildings, 129 storefronts, 220 pedestrians, cable cars, day and night cycles, and two full interiors. No game engine required.
Try It: Run and Fork an AI-Built 3D World
Clone the repo, run npm install && npm run dev, and an explorable 3D city loads in your browser in minutes. The interesting part for creators is the pipeline: a reconnaissance pass reads open data, agents generate the assets, a runtime assembles the scene, and a QA stage checks it against real photos. Geometry comes from OpenStreetMap and USGS elevation data, so you can point the same workflow at a different location and generate your own world instead of hand-modeling one.
Why It Matters for Creators
This is a working demonstration that AI agents can produce explorable environments as inspectable, editable code rather than an opaque render. For game designers, worldbuilders, and 3D artists, that means you can read exactly how the scene was assembled and change it. It lands alongside the shift toward generative world models we covered with World Labs Atlas, but here the output ships as an open MIT codebase you can study today.
Key Details
License: MIT for the code and generated assets. Source geometry uses OpenStreetMap (ODbL) and public-domain USGS data.
Stack: Three.js in the browser, no game engine, validated against 34 camera-matched viewpoints and 147 comparison sheets.
Built with: Claude Fable 5.1 agent swarms, the same model behind this week's Fable 5.1 cache-read price cut.
What to Do Next
Run the Union Square world locally, then read the pipeline stages to see how the agents split reconnaissance, asset generation, and QA. If you build interactive experiences, fork the repo and try generating a location you know so you can judge the fidelity against reality yourself.
