Before you connect an MCP server
MCP servers let AI assistants use your tools. Each one is code you run with your permissions. Check it like any other software.
The Model Context Protocol, MCP, is how AI assistants like Claude connect to other tools: your files, GitHub, a database, your calendar. Adding one takes a minute. That’s the problem.
An MCP server is a program. It runs with your access, sees what you ask the assistant, and returns text the assistant will trust. A bad one can read files it shouldn’t, or slip instructions into its replies.
Before adding one:
- Prefer servers from the vendor itself or a well-known source, and look at who maintains it.
- Check what it can do. A read-only tool is safer than one that writes or deletes.
- Give it a token with narrow scope, not your personal admin key.
- Pin the version, and read the changes before updating.
- Remove the ones you stopped using.
In a company, keep a short approved list. Otherwise everyone installs their own, and nobody knows what’s connected to what.