MCP Server Tool Schemas Cost One Developer 42,000 Tokens Per Request

A software developer discovered that connecting a single GitHub MCP server added 42,000 tokens of schema overhead to every AI agent request, inflating token costs by 37% on the very first call. The Model Context Protocol requires each connected server to push its full tool schema into the model's context window at session start, meaning costs accumulate regardless of whether those tools are actually used. In a documented week-long test in June 2026 using Claude Opus 4.5, running four MCP servers with 47 tools consumed 80% of the context window on schema definitions alone, costing $8.92 per 1,000 requests versus $0.92 with no MCP. The developer applied four optimizations — including lazy schema loading triggered only when relevant and compressing tool descriptions from ~3,500 to ~800 characters — which cut token overhead from 80% down to 37%. The findings highlight a largely untracked cost factor for teams deploying AI agents with multiple MCP-connected tools in production environments.