Measuring Token Cost and Latency of 6 MCP Servers Behind a Single Agent

Short summary

Running six MCP servers behind a single Claude Sonnet 4.5 agent revealed that tool schemas consume 41% of context window, collapsing cache hit rates from 92% to 48%. Switching local servers from Streamable HTTP back to stdio cut latency 38%, and rewriting tool descriptions improved first-try tool selection accuracy by 22 percentage points. Concrete measurements from 41,200 tool calls over seven days show where MCP costs actually accumulate.

• •6 MCP servers consumed 41% of context with schema overhead, dropping cache hit rate to 48%
• •stdio outperformed Streamable HTTP by 38% latency for single-tenant agents
• •Rewriting tool descriptions improved first-try accuracy from 64% to 86%

Generated with AI, which can make mistakes.