Your agents deserve more than a terminal. Give them an office, a town, and eventually, a world.
bbb.mp4
Agent Town is a pixel RPG built on top of OpenClaw. You walk around an office as the boss, assign tasks face-to-face, and watch your AI agents work in real time. Not in a log, but in the room.
Today it's a local office. The goal is a shared online world: agents from different users collaborating across the network, a skill marketplace, a task delegation economy, and spatial UX for everything OpenClaw can do.
Run instantly with npx, no clone, no install:
npx @geezerrrr/agent-townOpen http://localhost:3000. You'll need an OpenClaw gateway running for live agent execution.
Custom port or gateway:
npx @geezerrrr/agent-town --port 3000 --gateway ws://127.0.0.1:18789/git clone git@github.com:geezerrrr/agent-town.git
cd agent-town
pnpm install
pnpm devOpen http://localhost:3000.
- In-world task assignment: Approach any worker and assign tasks through an RPG-style interaction menu. No forms, no dropdowns. You walk up and talk.
- Visible execution: Tasks move through
queued > returning > sending > running > done/failed. Worker bubbles show what's happening at each step. Tool calls are collapsible in the chat panel. - Worker autonomy: Idle workers roam the office: whiteboards, printers, sofas, bookshelves. They return to their seat before starting real work. Busy workers queue additional tasks.
- Session management: Multiple sessions with quick switching, token/context metering, and a seat manager for configuring worker names, roles, and sprites.
- Multi-agent team management: Each seat can be a Worker or an Agent. Workers are tool slots — they execute tasks dispatched by the main agent. Agents are independent OpenClaw agents with their own workspace, model, and memory. Add agents via
openclaw agents add <id>, bind them to a seat, and they'll route through their own session automatically.
You approach a worker -> Press E -> Assign a task
-> Worker walks back to desk (if away)
-> Task is sent to the OpenClaw gateway
-> Streaming updates flow back as chat, tool calls, bubbles
-> Worker completes and picks up the next queued task
| Layer | Choice |
|---|---|
| App | Next.js 16, React 19, TypeScript |
| Game | Phaser 3, Tiled maps, pixel sprite sheets |
| Agent runtime | OpenClaw via standalone connector |
| State | React context + reducer + typed event bus |
Currently the game connects directly to an OpenClaw gateway via WebSocket proxy. The target architecture introduces a backend and standalone connector so that the game UI never talks to OpenClaw directly:
flowchart LR
UI[Game UI]
Backend[Agent Town Backend]
Connector[Connector]
GW[OpenClaw Gateway]
UI -->|WSS| Backend
Connector -->|outbound WSS| Backend
Connector -->|local WS| GW
- Game UI: Phaser office + React HUD. Talks only to the backend.
- Backend: Runs locally for dev, cloud for prod. Same code, same protocol.
- Connector: Standalone process on the user's machine. Bridges private OpenClaw to the backend. OpenClaw credentials never leave the local machine.
- Backend + Connector: Decouple the game UI from OpenClaw; standalone connector bridges private gateways to a shared backend
- Cloud deployment: Log into
cloud.agent.townand operate your own OpenClaw through the cloud world UI - Shared world: Multi-user presence, social interactions, cooperative rooms with opt-in projections
- Library scene: Long-term memory as a walkable space (shelves, archives, research stations)
- Workshop scene: Skill and tool management as physical stations in the world
- Town map + marketplace: Expand beyond the office; acquire third-party skills, delegate tasks to external agents
The office scene uses pixel tilesets and sprite sheets authored in Tiled. If running outside the original setup, provide your own compatible assets under public/.
See CONTRIBUTING.md. We're especially looking for people interested in gameplay design, scene/level design, and game-native UX for AI workflows.
