Too many MCP tools make your AI worse. Here is how to fix it
Published November 4, 2026
An AI app sends the definition of every tool from every connected MCP server to the model, before your first message. Connect ten servers with twenty tools each and the model reads two hundred tool descriptions on every turn: it costs context, slows answers and makes the model pick the wrong tool more often. Fix it by connecting fewer, broader tools, merging accounts into one tool with an account argument, and letting the model search for rarely used tools.
Where the cost comes from
A tool definition is a name, a description and a JSON schema of its inputs. Detailed schemas are good for accuracy but long. They are sent with every request, so they take part of the context window that your documents and conversation could use.
Similar tools also compete. With “search_messages” from four servers, the model has to guess which one you mean.
Three fixes
- One tool per action, not per account: “search mail” with an account argument instead of a copy per mailbox. Reads can take “all” to search every account at once.
- Shared tools across services: one “search files” that covers Google Drive, Notion and Confluence, and one “list tasks” for every task manager.
- Search instead of a full list: show the model a few meta tools (“search tools”, “call tool”) and let it find the specific one when it needs it.
How Manyplug handles it
Manyplug groups accounts into one tool with an account argument, offers shared tools for mail, calendar, files, tasks, contacts and messages, and switches to a short search-based list when a workspace has many tools. You can choose the mode in the settings.