Your MCP tool surface has a token bill — here's how to read it

If you're building MCP servers, here's a cost that's easy to miss: Every tool you expose is re-sent to the model on every call.

Not once at startup, every. Single. Turn. The tool's name, its description, and its full JSON input schema all go into the model's context each time it decides what to do. Expose 30 tools with verbose schemas and you can burn a thousand-plus tokens on tool definitions alone before the agent has done a single useful thing. Worse, a bigger, noisier tool menu measurably degrades tool-selection accuracy — the model has more near-identical options to confuse.

Most of us add tools to an MCP server with zero feedback on what it's costing. So I built a small CLI to make the bill visible.