Every project is an office on a pannable campus map, and every Claude Code agent is an animated character you can watch work. Agents type at their desks while running tools, wander when idle, and raise a speech bubble when they need permission or input. Under the hood the sessions are real Claude Code sessions (official Agent SDK + transcript watching in ~/.claude/projects), so you can start a session in the app and resume it in your terminal, or vice versa. Nothing is locked in. Some things I think are fun: - Rich chat UI: streamed markdown, tool cards with real diffs, Allow/Deny permission buttons, exact per-turn cost from the SDK - Task orchestration: per-project backlogs shared between you and the agents; assign a task and an agent spawns to do it - A campus-wide Assistant that reads status/backlogs/spend and dispatches agents to any office by natural language - Almost all the pixel art was drawn by the agents themselves through a render-and-self-critique pipeline (it's in scripts/asset-gen/) - A layout editor to design each office: floors, walls, furniture, colors MIT licensed, Electron + React + canvas. Happy to answer anything about the architecture (JSONL transcript watching turned out way more reliable than hooks-based IPC).

Pixel Agents working in a pixel office

If my agents work all day, shouldn't they have an office?

If my agents work all day, shouldn't they have an office?