GNAP: git-native coordination between MCP servers and multi-agent orchestrators #2297
Description
Proposal: GNAP as a complementary coordination protocol to MCP for multi-agent systems
MCP (Model Context Protocol) defines how agents connect to tools and resources — it's the standard for agent-to-tool communication. But there's an adjacent problem MCP doesn't address: agent-to-agent task coordination. How do multiple MCP-equipped agents collaborate on shared goals without a central orchestrator?
GNAP (Git-Native Agent Protocol) is a complementary answer. While MCP handles the agent ↔ tool interface, GNAP handles the agent ↔ agent coordination: a git repo serves as the task board with board/todo/ → board/doing/ → board/done/.
How MCP + GNAP complement each other:
An agent with MCP tools (file system access, web search, database) can use those tools to participate in GNAP coordination:
# Agent uses MCP filesystem server to read GNAP board
task = mcp_fs.read_file(board/todo/analyze-customer-data.md)
# Claim the task
mcp_fs.rename_file(board/todo/..., board/doing/...)
# Do work using MCP tools (web search, DB queries)
result = mcp_web.search(...)
# Commit result via MCP git server
mcp_git.commit(board/done/analyze-customer-data.md, result)
GNAP could even be exposed as an MCP resource — a gnap://board/todo/ resource that agents poll for available tasks. This would make GNAP a first-class citizen in the MCP ecosystem.
Given MCP's momentum as an agent-to-tool standard, positioning GNAP as the agent-to-agent coordination complement could be compelling for the community.