A neon-cyberpunk pixel-art office where every character is a real AI agent. Give the Boss a goal and it decomposes the work, delegates to role workers (engineer, designer, researcher…), they each do real work via the Claude / OpenAI API, and the Boss synthesizes a final answer — all visualized in a top-down office you can watch.
Single self-contained index.html — no build, no dependencies. Double-click to run (file://) or host anywhere static.
- Open
index.html(or the GitHub Pages URL). - ⚙ Settings → paste your Anthropic API key and/or OpenAI API key (stored only in your browser's localStorage; sent only to the provider).
- Type a goal in the top bar → DISPATCH. The Boss plans and delegates.
Keys are pay-as-you-go API keys (from
console.anthropic.com/platform.openai.com), not Claude.ai / ChatGPT subscriptions. A local companion server for subscription use is optional (seecompanion/).
- Orchestrator + workers — Boss decomposes a goal into role-tagged subtasks, delegates, then synthesizes.
- Multi-provider — mix Claude (Opus/Sonnet/Haiku) and OpenAI (GPT-4o / 4o-mini / 4.1) per agent.
- Live office — agents pathfind to desks, "code"/"search", gather to meet, take ☕ breaks; speech bubbles + activity log.
- Controls — + Agent (name/role/model/color), click an agent to chat or assign, task board (kanban), layout editor, ☕ break (per-agent / everyone), web search toggle, export/import, autosave.
The deliverable index.html is generated from src/*.js + src/shell.html + src/styles.css:
bash build.shEdit the modules in src/, then rebuild. Load order is fixed (config → pixelart → world → api → store → agents → orchestrator → ui → main). SPEC.md is the architecture contract.
Transient API failures ("Failed to fetch", rate limits, overload) auto-retry with exponential backoff; concurrency is capped and staggered to stay under rate limits.