Control MCP tool definition token overhead in agent workflows
Updated September 19, 2026 · first published September 19, 2026
Connecting an AI agent to multiple Model Context Protocol (MCP) servers injects thousands of input tokens on every turn before the model reads a single word of user intent. Every registered tool appends a JSON schema containing function names, parameter descriptions, enum definitions, and type constraints. In an enterprise workspace with 6 connected MCP servers (GitHub, Postgres, Slack, Jira, Brave Search, Filesystem), this tool preamble consumes 4,500 to 12,000 input tokens per API call. At 30 turns across a task, that static overhead drains $0.35 to $1.20 in pure schema redundancy.
The anatomy of MCP schema tax
The Model Context Protocol establishes a standard client-server JSON-RPC interface for LLM capabilities. However, language models cannot infer tool availability without explicit schema injection in the system prompt or tool definition parameter block. When an engineer connects a database MCP server with 24 schema-inspection and SQL tools, the full JSON schema must be presented to the model on every inference turn so the attention mechanism can route tool calls. Because standard API billing meters total input tokens on every request, unmanaged MCP tool catalogues create an immediate, linear baseline tax across all agent operations.
Why do MCP server tool definitions cost so many tokens?
Each MCP tool definition requires verbose metadata: parameter names, nested properties, human-readable descriptions, and JSON Schema specification headers. A single detailed database query tool often consumes 350 to 600 tokens. Across dozens of tools spanning multiple departmental MCP servers, the initial context window is occupied by static tool contracts rather than dynamic conversational history.
Three architectures to eliminate MCP overhead
Engineering teams running production agent fleets deploy three distinct mitigation strategies to control tool token inflation:
- Prompt Caching Breakpoints: Place the static MCP tool definitions before any dynamic conversation history in the prompt structure, and set an explicit prompt caching breakpoint immediately after the tool block. On Anthropic, OpenAI, and Google Gemini models supporting 5-minute or 1-hour cache TTLs, cache hits reduce the input token cost of the tool schema by 75% to 90% after the first turn.
- Dynamic Two-Stage Tool Registration: Avoid exposing all MCP tools globally. Use a lightweight router model or semantic index to identify the specific tool domain required by the user prompt, dynamically injecting only the relevant 3–5 tool schemas into the execution agent's active payload.
- Schema Compression and Description Minification: Strip verbose formatting, eliminate markdown descriptions within JSON schema strings, and replace long field explanations with compact type definitions. Aggressive schema minification typically cuts tool definition payload volume by 35% to 50% without degrading tool selection accuracy.
How can engineering teams reduce MCP tool token overhead?
Teams reduce MCP overhead by setting prompt cache breakpoints after tool declarations, employing two-stage semantic tool routing, and minifying redundant JSON schema property descriptions.
Can prompt caching eliminate MCP schema token costs?
While prompt caching cannot eliminate token processing latency entirely, it slashes the billing cost of recurrent tool definition tokens by up to 90% on models that support prefix caching, provided the tool definition block remains identical across consecutive turns.
Related
Want this applied to your own LLM spend? FinOps LLM runs a free audit of your AI costs and shows where the savings are. Book free audit →