MCP Tool Description Token Overhead

A development team had 8 MCP servers connected to their Claude Desktop: separate servers for Google search, Amazon lookup, YouTube search, Reddit search, a weather API, a database, a file system, and a calculator. Tool descriptions consumed 12K tokens per message. They consolidated the four search servers into Scavio's single MCP server, dropping tool description overhead to 5K tokens -- saving $18/day in token costs across their team.

MCP Tool Description Token Overhead is relevant to Google, Amazon, YouTube, Walmart, Reddit, TikTok. Scavio provides a unified API to access data from all of these platforms.

When an MCP client (Claude Desktop, Cursor, a custom agent) connects to MCP servers, it includes each server's tool descriptions in the system prompt sent to the LLM on every turn. A typical MCP server exposes 3-10 tools, each with a name, description, and parameter schema. This adds 500-2000 tokens per server to every LLM call. With 5 MCP servers connected, you are paying for 2,500-10,000 extra input tokens on every single message, even if the user's question has nothing to do with those tools. At Claude's pricing ($3/million input tokens), 10K extra tokens per message across 1K messages/day costs $30/day in pure overhead. The compounding effect is worse: more tools in context also degrades the LLM's tool selection accuracy, as the model must parse through more options. The solution is server consolidation: use fewer servers that each cover more surface area. Scavio's MCP server (mcp.scavio.dev/mcp) covers Google, Amazon, YouTube, Walmart, Reddit, and TikTok search in a single server, replacing what would otherwise be six separate search-related MCP servers. One server's tool descriptions instead of six means roughly 5x reduction in search-related token overhead.