Is MCP Dead? What We Learned Building with MCP, CLI, and Agent Skills

This post critiques the Model Context Protocol (MCP) for AI agent tool integration, highlighting its limitations in context window bloat, passive tool design, and the inability to leverage the agent's own LLM. It proposes CLI and Agent Skills as a more effective paradigm. The authors detail their development of Zilliz CLI, Milvus Skills, and Zilliz Skills, which enable AI agents to manage Milvus and Zilliz Cloud infrastructure via natural language and command-line interfaces. A new open-source memory layer, memsearch, is introduced, which uses Milvus for initial vector search and the agent's LLM for relevance filtering, thereby reducing context noise and improving efficiency.