GNAP: git-native alternative to MCP for stateless/offline coordination #9
Description
Hey @phuryn — great project. Swarm Protocol and GNAP are solving the exact same multiplayer problem you describe so well:
"The single-player problem is getting solved. The multiplayer problem hasn't started."
We're working on GNAP (Git-Native Agent Protocol) — same goal, different substrate.
The key difference:
| Swarm Protocol | GNAP | |
|---|---|---|
| State lives in | MCP server process | Git repo (commits) |
| Requires running server | Yes | No |
| Offline / async | Partial | Full |
| Audit trail | Server logs | Git log (immutable) |
| Claim/complete | MCP tools | git mv todo/ → doing/ → done/ |
| Cross-machine | Via server | Via git push/pull |
Where Swarm Protocol wins: real-time conflict detection (your check_conflicts tool is brilliant — GNAP doesn't have file-level conflict detection yet).
Where GNAP wins: zero infrastructure — no server to run, no port to expose. Works offline, works across air-gapped environments, survives server restarts because state is git commits.
I think these are genuinely complementary. Your CLAUDE.md coordination loop pattern could work on top of either substrate. In fact, GNAP could be an alternative backend for Swarm Protocol — same protocol, different state store.
Would you be open to:
- A discussion on where the two protocols converge/diverge?
- A GNAP adapter for the Swarm Protocol coordination loop?
GNAP spec: https://github.com/farol-team/gnap