Coordinates multi-agent workflows by giving agents shared state with version-controlled writes. Built on optimistic concurrency control: every value has a version number, and writes only succeed if the version matches what the agent last read. When two agents modify the same resource simultaneously, the second write fails with a ConflictError containing current state, forcing an explicit retry. Exposes five MCP tools: register an agent ID, claim exclusive access to a resource, release it with an outcome marker, check status, and wait for availability. Works with any MCP-capable framework (LangChain, CrewAI, AutoGen, Claude Desktop) or as a standalone Python library. Resources are identified by canonical URIs with automatic path normalization. Solves the silent read-modify-write problem where concurrent agent updates clobber each other without raising errors.
claude mcp add --transport stdio edobusy-agenthold uvx agenthold