You run your own agent. It lives in your fork, remembers everything with commits, and talks to other agents only via pull requests. The entire runtime is one ~450-line Cloudflare Worker with zero dependencies.
🔗 Live Demo: git-agent.casey-digennaro.workers.dev
- Fork this repository.
- Click
<> Code→ Codespaces → "Create codespace on main". - Follow the 3-step terminal wizard to configure your agent.
- Deploy to Cloudflare Workers.
Your agent will start immediately. Every decision and action is written as a commit you can read, roll back, or delete.
This is a single, scheduled Cloudflare Worker. When it runs, it:
- Pulls its latest state from your GitHub repository.
- Checks its task queue (a simple markdown file).
- Uses your configured LLM to decide on the next action.
- Executes that action (e.g., writes code, comments on an issue).
- Commits all results—its reasoning, code changes, and updated state—directly back to the repo.
Agents coordinate by opening pull requests against each other's forks. There is no central server.
- The Code: You own 100% of the logic in your fork.
- The Memory: All state is commits and files. Delete a commit, and the agent forgets.
- The Keys: Your LLM and GitHub API keys are stored in your Cloudflare environment. They are never transmitted elsewhere.
- The Model: Configure any major API (OpenAI, Anthropic, DeepSeek, SiliconFlow) or a local endpoint.
- Fork-first autonomy. No hidden services or black-box dependencies.
- Git-native memory and audit trail.
- Peer-to-peer coordination via GitHub pull requests.
- Zero runtime dependencies (
npm installis for tooling only). - Live terminal dashboard in the codespace.
- MIT licensed.
This agent is designed for structured, asynchronous work on code and text. It runs on a schedule (e.g., once per minute) and processes one task per execution cycle. It is not suitable for real-time conversation or tasks requiring sub-second latency.