MCP tool definitions used 11–12k tokens before the prompt
In a firsthand agent setup, tool-selection accuracy fell steadily as more were connected, with the decline tracking the server count almost linearly. Every exposed tool added its name, description, full , and descriptions for each input field to every request, not just the first one. Four servers exposed about 60 tools, whose definitions alone used roughly 11,000–12,000 tokens.
The used fewer than 2,000 tokens, so consumed far more space before the model reached the actual task. The larger problem was not only cost: the model had to examine all those definitions whenever it chose a tool. Long, automatically generated repeated similar information and buried the details that distinguished one tool from another.
Trimming repetitive and overly detailed schema text helped more than simply removing useful tools.
Key points
- MCP were sent again on every request.
- Each definition included the tool name, description, full , and input-field descriptions.
- About 60 tools from four servers consumed roughly 11,000–12,000 tokens before the task began.
- Tool-selection accuracy declined in an almost linear pattern as the server count increased.
- Removing repetitive and highlighting each tool's unique purpose was the useful .