Show HN: MCPgraphy – a proxy that learns when your agent's MCP tools work (github.com)

MCPgraphy, a new proxy tool unveiled on HN, significantly enhances the operation of agents using Model Context Protocol (MCP) servers by learning from the agent's tool calls to determine which tools are effective in specific states. This proxy sits between the agent and the server, observing each call's outcome while maintaining the existing command structure. By creating a behavior graph, MCPgraphy categorizes tools into safe candidates, required approvals, and failures, actively reducing invalid calls while preserving agent efficiency. Users can install it via pip, and its direct integration with existing tools requires no changes to APIs or keys, promising secure local functionality. This tool is noteworthy for the AI/ML community as it smartly mitigates errors caused by agents making unsupported or erroneous calls, a common challenge faced during agent deployment. Benchmarked against traditional methods, MCPgraphy demonstrates superior performance, completing jobs with fewer invalid calls and achieving a higher overall task success rate. Its passive learning approach captures server behavior changes, thus informing agents of new state interactions without additional requests, making it an efficient solution for optimizing agent performance and reliability in complex operational environments.