Autopoe

A multi-agent collaboration framework where lightweight AI agents work together through structured coordination to accomplish complex software development tasks.

Pick one:

uvx autopoe                       # run without installing
uv tool install autopoe           # install via uv
pip install autopoe               # install via pip

To run Autopoe in Docker, download docker-compose.yml and run:

docker compose up -d

Autopoe will be available at http://localhost:6873.

• Bubblewrap (bwrap) — required for agents to execute shell commands in a sandboxed environment. Install it before running Autopoe if you want agents to be able to run code.

On first run, configure your LLM provider via the Settings panel (gear icon). Four API types are supported — any compatible endpoint works:

• OpenAI-compatible — OpenRouter, Ollama, ModelScope, vLLM, LiteLLM, or any /v1/chat/completions endpoint
• OpenAI Responses — OpenAI or compatible /v1/responses endpoints
• Anthropic — any endpoint following the Anthropic Messages API
• Google Gemini — any endpoint following the Gemini generateContent API

Autopoe stores instance data in ~/.autopoe/ by default, including settings.json, workspace snapshots, and image assets. You can override that app data directory before startup with AUTOPOE_APP_DATA_DIR=/path/to/data or autopoe --app-data-dir /path/to/data. The system working_dir can be changed at runtime from Settings without restarting.

# Clone the repo
git clone https://github.com/ImFeH2/autopoe.git
cd autopoe

# Backend (API + WebSocket only, http://localhost:8000)
uv sync
uv run fastapi dev app/dev.py

# Frontend (hot reload, http://localhost:6873, separate terminal)
cd frontend
pnpm install
pnpm dev

In development, open http://localhost:6873. The Vite dev server proxies /api and /ws to the backend on http://localhost:8000. In production, the backend serves the built frontend from app/static on http://localhost:6873.