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

What this post added

This post critically analyzes the limitations of the Model Context Protocol (MCP) for AI agent tool integration, specifically addressing context window bloat, passive architecture, and the inability of MCP servers to leverage the agent's LLM. It contrasts MCP with a CLI and Agent Skills paradigm, showcasing Zilliz's implementation with Zilliz CLI, Milvus Skills, and Zilliz Skills for managing Milvus and Zilliz Cloud. A key technical contribution is the memsearch memory layer, which uses a multi-stage process involving Milvus for initial retrieval, the agent's LLM for relevance evaluation, and a final drill-down, thereby reducing noise and improving context quality. The post provides a framework for choosing between MCP, CLI, and Agent Skills based on technical requirements.