PhiloLabs' open-source project that uses Claude Fable 5.1 agent swarms to generate explorable, browser-native 3D reconstructions of real places as plain Three.js apps.
Last verified · 2026-09-03 · by Moe Ameen
Fable 5.1 World Modeling is the informal name for PhiloLabs' open-source project "fable51-worlds" — "worlds via code, from fable 5.1." It uses autonomous Claude Fable 5.1 coding agents to research, model, and quality-check explorable reconstructions of real places, then ship them as plain Three.js applications that run in a browser. The important distinction up front: this is not a diffusion model that paints frames. It is a coding agent that writes the 3D geometry, the scene assembly, and the runtime. Every building, storefront, sign, tree, and traffic light is generated from open data and public reference imagery by code.
The pipeline has four stages. Research: parallel agents pull OpenStreetMap geometry (ODbL) and USGS 3DEP elevation (public domain), plus transit and storefront data, with source attribution. Asset generation: Blender scripts (bpy) emit optimized GLB kits — façade modules, street furniture, vehicles, vegetation. Runtime assembly: a pure Three.js app combines terrain, streets, façades, props, crowds, and traffic from JSON specs. Quality assurance: Playwright camera-matches fixed viewpoints against reference photographs, and independent reviewer roles (architect, geographer, technical artist) file reports.
Two worlds ship in the repo. Union Square, San Francisco has 453 building footprints, 129 named storefronts, 220 pedestrians, 109 vehicles, working traffic lights and cable cars, day/sunset/night modes, and two explorable interiors (Apple and Nintendo). Higashiyama, Kyoto has 266 buildings, 471 shopfronts, a 2.3 km continuously walkable route, and a cel-shaded anime look that deploys as a single self-contained HTML file. The creator reports a world running in roughly two hours for about 8 million tokens and around $33 in API cost.
Be clear-eyed about what it is and is not. It is a research-grade, developer-facing showcase — you run code and agents, not a consumer app with a signup. It is licensed MIT (code and generated assets). And its honest limits are known: topology can be messy and texturing is hard. It is not a generative video or image tool; the shareable visual it produces for a creator is a walkthrough video filmed inside the finished world. Treat any figure here as a snapshot of the repo and verify against the source.
The most useful way to think about Fable 5.1 World Modeling for content is as a virtual set builder. It hands you a moving image of a place — a walkthrough the agent filmed, or a screen-recorded fly-through, with day/sunset/night variants — and that footage is exactly the kind of raw source [Kompozy](/) turns into finished posts. The bridge is not "post the web app"; it is "use the world's footage as your backdrop and b-roll," then let Kompozy do the production.
Concretely: take a fly-through clip of the Union Square build and Kompozy uses it as B-roll behind a [Persona Short](/glossary/persona-shorts) — your HeyGen talking-head avatar narrates a script while the AI-built street rolls behind them, with auto-captions burned in. The same clip feeds [Clipped Shorts](/glossary/clipped-short), which cut the long fly-through into vertical moments for TikTok, Reels, and Shorts, and Marketing Shorts, which stitch a hook over the demo footage. Pull stills from the night variant into a brand-exact [carousel](/glossary/hyperframes) or quote graphics, spin a blog on how the world was built, and a newsletter — all governed by one [Persona Brief](/glossary/persona-brief) so the batch reads as one creator. Then [Autopilot](/glossary/autopilot) schedules and publishes the set across the eight social platforms plus blog and email, and its per-post review is where you add your AI-disclosure line. The model builds the set; Kompozy shoots the scene, cuts it, and ships it. (For the model behind the agents, see [Claude Fable 5.1](/ai-tools/fable-5-1).)
No. It is a coding-agent project (PhiloLabs' fable51-worlds) that builds explorable 3D scenes as Three.js web apps from open data. The video it produces is a walkthrough the agent films inside the finished world — not a diffusion-rendered clip like Sora or Veo. The geometry is written as code, not painted as frames.
PhiloLabs built it using Anthropic's Claude Fable 5.1 model. The repository is open source under the MIT license (code and generated assets). Running it costs Claude API usage — the creator reports about $33 and 8 million tokens for a single world — so it is free to fork but not free to generate a new world from scratch.
The web app itself does not post to social platforms, but the walkthrough video does. Screen-record or export a fly-through, then use a content engine like Kompozy to use it as B-roll behind a persona short, cut it into clips, pull stills for carousels, and schedule the set across platforms.
Claude Fable 5.1 is Anthropic's text-and-reasoning model. Fable 5.1 World Modeling is a third-party project that puts that model's coding ability to work generating 3D worlds. The model is the brain; the project is one striking thing people built with it.