MCP Design in the Real World

When we started Project Brain's MCP server, we followed a common pattern: every new need got a new tool. It felt productive at first. But after a while, the tool menu became crowded, the rules around each tool grew, and the agent got slower at choosing what to call. That led to more retries, higher token usage, and slower progress on simple tasks. The big lesson was straightforward: after a certain point, adding more tools hurts more than it helps. In this post, I’ll share what changed for us and what we learned from running this in production. Motivation: why reduce the number of tools? 1. Tool selection gets harder for the model Every new tool is another decision branch. Instead of focusing on the task, the model spends effort deciding which tool is "most correct," whether parameters are supported, and whether an older tool has been replaced by a newer one. Those extra decisions show up as wrong calls and wasted turns. 2. Context gets bloated Each tool adds descriptions, argument rul