We introduce \method\, a mutual evolution framework that closes the loop between Model Context Protocol (MCP) agents and their training data. Conventional supervised fine‑tuning relies on a static data distribution, which cannot keep pace with a student’s growing capabilities nor address newly exposed weaknesses. \method\ first assesses the student’s current skills and gaps from rollout feedback, thereby defining learning targets. Task evolution then converts these targets into executable tasks by adapting the instructions and observable states of compatible verified seed tasks while keeping their verifiers aligned. To supply reliable and efficient demonstrations, trajectory evolution reuses modular skills to recover failed teacher rollouts and refine successful ones. The resulting task‑trajectory pairs fine‑tune the student, and accepted tasks seed the next evolution round. After each fine‑tuning cycle the updated student is re‑evaluated, and its feedback guides the subsequent data evolution. Experiments on Toolathlon, MCPMark, and MCP‑Atlas show consistent gains of 16.5%, 10.2%, and 12.6% respectively. The codebase is publicly available at https://github.com/xagao/EvoMCP.
Review: \method\ demonstrates that dynamically generating targeted tasks and high‑quality trajectories can continuously adapt a student model, offering a practical route toward robust MCP agents.