MCP's tool explosion problem limits its effectiveness for coding agents¶
Insight: Despite being available for nearly a year (as of mid-2025), MCP has not gained widespread usage in coding workflows. The core problem is the static nature of tool definitions: the protocol doesn't allow selective or dynamic enabling of tools within a server. If an MCP server has 20 tools, all 20 definitions are stuffed into the context window, degrading model performance through reduced signal-to-noise ratio, tool confusion, increased costs, and reduced usable context. For coding agents specifically, CLI-based approaches may be more effective than MCP.
Detail: Zhu Liang assesses MCP as "Emerging, Limited Effectiveness" — a notably cooler assessment than the hype. Workarounds exist (Cursor allows selective tool enabling, you can fork servers to remove tools) but they don't address the underlying protocol limitation. Shrivu Shankar adds in a footnote: "I'm not knocking MCP generally... For non-coding agent use cases, MCP is critical for bridging the gap between agent interfaces and third-party data/context providers." This suggests MCP's value proposition differs significantly between coding agents (where CLI tools already work) and general-purpose agents (where MCP solves a real integration problem). The key guidance: "be selective and strategic about the MCP servers and tools you enable."
Sources
Related: existing entry "MCP has fundamental security and UX problems" in external/mcp.md — CORROBORATES