I Installed 20 MCP Servers and Ended Up Worse Off
I Installed 20 MCP Servers and Ended Up Worse Off
The MCP ecosystem is exploding. There are servers for everything - GitHub, Slack, databases, file systems, browsers, email, calendar, and dozens more. The temptation is to install all of them. More tools means more capability, right?
Wrong. After installing 20 MCP servers, my AI agent got measurably worse at everything.
The Context Window Tax
Every MCP server registers its tools with the agent. Each tool has a name, description, and parameter schema. Twenty servers might register 150+ tools. That tool list alone can consume thousands of tokens from your context window before you've even asked a question.
The agent spends more time deciding which tool to use than actually using it. Tool selection accuracy drops because the model is choosing between too many similar options. "Should I use the GitHub MCP to read this file, or the filesystem MCP, or the project MCP?"
The Sweet Spot Is 3-4 Servers
After months of experimentation, the setup that actually works is minimal: a filesystem server, a browser server, and one or two domain-specific servers for whatever you're working on that day. Swap them based on the task, don't run them all simultaneously.
This keeps the tool list short, reduces context consumption, and lets the agent make confident tool selections without second-guessing itself.
Configuration Over Accumulation
The real power move is configuring a few servers well rather than installing many servers with default settings. A properly configured GitHub MCP server with the right scopes and filters is worth more than ten servers with default configurations fighting for the agent's attention.
Think of it like browser extensions. Everyone installs 30, disables 25, and actually uses 5.
Fazm is an open source macOS AI agent. Open source on GitHub.