An ant farm for AI agents.
Instead of watching a wall of scrolling terminal output, your Claude Code instances become little characters running around a living village — sprinting to the terminal when they run a command, ducking into the library to read a file, heading to town hall to spawn a subagent. It's the same work, just alive.
Run multiple Claude instances and each one gets its own character, moving independently through the same village. Think of it as a gamified window into what your agents are actually doing.
Try it instantly at vibecraft.sh — still connects to your local Claude Code instances!
Claude Code → Hook Script → WebSocket Server → Browser (React)
Claude Code fires hook events on every tool call. A bash hook script captures those events and forwards them to a local WebSocket server, which broadcasts them to the browser. Character sprites move to the matching village location in real time.
No files or prompts leave your machine.
- macOS or Linux (Windows not supported — hooks require bash)
- Node.js 18+
- jq — for hook scripts (
brew install jq/apt install jq) - tmux — for session management (
brew install tmux/apt install tmux)
# 1. Install dependencies
brew install jq tmux # macOS
# sudo apt install jq tmux # Ubuntu/Debian
# 2. Configure hooks (one time)
npx vibecraft setup
# 3. Start server
npx vibecraftOpen http://localhost:4003 and use Claude Code normally. You'll see your agent move around the village as it uses tools.
From source:
git clone https://github.com/nearcyan/vibecraft
cd vibecraft && npm install && npm run dev
# Opens on http://localhost:4002To uninstall: npx vibecraft uninstall (removes hooks, keeps your data)
Run multiple Claude Code instances simultaneously and each one appears as its own character in the village, moving independently based on what it's doing. Characters spread out when they're at the same location so you can always tell them apart.
To send prompts to a specific instance, run Claude in a named tmux session:
tmux new -s claude
claudeThen use the session panel to select which instance to talk to.
Characters move between 9 locations based on the active tool:
| Location | Tools |
|---|---|
| Library | Read |
| Writer's Cottage | Write |
| Workshop | Edit |
| Terminal Tower | Bash |
| Observatory | Grep, Glob |
| Signal Tower | WebFetch, WebSearch |
| Town Hall | Task (subagents) |
| Notice Board | TodoWrite |
| Village Square | Idle / default |
- Per-session characters — each Claude instance gets its own sprite, moving independently
- Live character movement — agents walk to the building matching the active tool
- Subagent visualization — mini-agents appear at Town Hall when Task spins up parallel work
- Activity feed — full log of tool calls, prompts, and responses per session
- Session management — spawn and manage multiple Claude instances from one view
- Attention system — sessions pulse when they need input or finish a task
- Sound effects — synthesized audio feedback on tool calls and state changes
- Custom scenarios — generate entirely new visual themes with OpenAI + Stable Diffusion
- Per-asset regeneration — re-roll any individual character or building image if the generation misfired
Vibecraft can generate entirely new visual themes — different art styles, settings, and characters — using OpenAI and Stable Diffusion.
Vibecraft uses AUTOMATIC1111 (A1111) with the --api flag enabled.
1. Install A1111 if you haven't already — follow the AUTOMATIC1111 installation guide.
2. Enable the API. Add --api to your launch arguments:
- macOS/Linux: edit
webui-user.shand add toCOMMANDLINE_ARGS:export COMMANDLINE_ARGS="--api"
- Windows: edit
webui-user.bat:set COMMANDLINE_ARGS=--api
Then launch A1111 normally. The API will be available at http://localhost:7860 by default.
3. Install required models and LoRAs into your A1111 models folder:
TODO: paste model and LoRA download links here
4. Set SD_URL in your .env to point at your A1111 instance:
SD_URL=http://localhost:7860
If A1111 is running on a different machine or port, update the URL accordingly.
Setup (one time): add to your .env:
OPENAI_API_KEY=sk-...
SD_URL=http://localhost:7860
Then open the ☰ menu in the app and click Create New Scenario. Describe the world you want:
"A cozy underwater kingdom with coral buildings and bioluminescent lighting" "A retro 80s cyberpunk city with neon signs and rain-slicked streets"
OpenAI plans the scenario and generates Stable Diffusion prompts. SD generates the background, all 9 building exteriors, and 7 characters with 5 animation states each (idle, walking, working, thinking, finished). Background removal runs automatically on characters and buildings.
If a generation goes sideways (LoRA didn't trigger, wrong composition), hit ↻ on any individual asset to regenerate just that image. OpenAI automatically rewrites the prompt with fresh phrasing before re-running SD.
Run Claude in tmux to send prompts directly from the browser:
tmux new -s claude
claudeThen use the prompt input at the bottom of the activity feed. The selected session receives the prompt via tmux.
vibecraft [options]
Options:
--port, -p <port> WebSocket server port (default: 4003)
--help, -h Show help
--version, -v Show versionSee CLAUDE.md for full technical documentation.
Website: https://vibecraft.sh
MIT