GNAP is a protocol for coordinating AI agents using only git, with no servers or databases required.
GNAP addresses the problem of orchestrating multiple AI agents without requiring centralized infrastructure. It works by storing the entire protocol state in four JSON files within a `.gnap/` directory in a git repository: agents, tasks, runs, and messages. Agents operate on a heartbeat loop where they pull from the repository, discover their assigned tasks, execute work, and push results back. Git history serves as the built-in audit log, eliminating the need for separate databases or logging systems.
Teams should adopt GNAP if they want to coordinate AI agents without deploying servers or managing databases, and if their agents can perform git operations. It suits projects where audit trails matter, where human and AI participants need to work together as equals, and where teams prefer working offline with eventual synchronization. The README compares GNAP to AgentHub, Paperclip, Symphony, CrewAI, and LangGraph, noting that GNAP uniquely requires no server, uses git instead of a database, avoids vendor lock-in, and supports offline operation. Among the alternatives listed, only GNAP and AgentHub support both human and AI agents as first-class participants, and only GNAP works offline.
The project shows active development with regular commits addressing protocol refinement and documentation. The maintainers are responsive to issues and pull requests, indicating ongoing engagement with users. The codebase demonstrates attention to protocol versioning and backward compatibility through explicit version checking mechanisms.
