MCP results cost tokens. Fetches, logs, and unbounded SQL are the usual way to fill the window until the model starts dropping filenames.
Bound the call: path, line count, fields, time window. Summarize before you return a body. Fetch details in a second call by offset or name.
Install fewer servers. Every tool schema sits in the prompt from the first turn. A filesystem server, version control, and one search tool is enough to start.
Detail pages here show category and description so you install less. Pair this with “when not to use MCP” and how to pick coding MCP.