I measured MCP vs. a CLI for agent search. The MCP used 17x more tokens per call.

A side-by-side test compared Google search via SerpApi's MCP server and a custom CLI, measuring token use for idle context, tool discovery, and results.

The CLI used far fewer tokens while returning the same underlying search output.

Most of the overhead came from always-loaded tool schema and bulkier MCP responses.

The finding suggests MCP tools can materially raise agent cost and latency when many tools are loaded.