We cut MCP token costs by 92% by not sending tool definitions to the model

If you're connecting Claude Code to MCP servers, every tool from every server gets injected into the model's context on every single request. 5 servers with 30 tools each means 150 tool definitions sitting in your prompt before Claude even starts thinking