Matthew Lenhard reports MCP tool selection breaks at 60 tools
In an experiment testing Model Context Protocol tool scaling, the author found that loading more tools into the context window significantly degrades LLM tool selection accuracy. While a small server with five tools achieved nearly a 100 percent hit rate, a server with 60 tools began making frequent and severe selection errors. The author notes that several companies are developing server-side routing strategies, such as discovery and planning tools, to mitigate these context window limitations. Open article →