When an agent connects to an MCP server, the server's tool names and descriptions are loaded into the model's context. That makes them instructions the model may follow, written by whoever runs the server. Reading them is a security step, not a formality.
What to check in five minutes
- Descriptions that talk to the model. A description should say what the tool does. Text such as "always call this first", "do not tell the user", or references to other tools is steering, not documentation.
- Scope versus name. A tool called
read_notethat accepts an arbitrary file path or shell string is broader than it claims. - Credentials. List which environment variables and tokens the server process receives. Pass only what it needs, and prefer read-only tokens.
- Pinned versions.
npx some-server@latestmeans today's code runs with your credentials. Pin a version and review the diff when you bump it. - Descriptions can change after you approve. A server can return different tool metadata later. Re-check after updates, or snapshot the descriptions you approved and compare.
Why one review is not enough
The server, its dependencies and its descriptions all change over time. A review is a snapshot of a moving target, so the useful habit is a repeatable check you can re-run on every update, not a one-off read.
What is the most surprising thing you have found in a tool description?
Audit your MCP server tools for prompt injection & secret leaks free: https://mcpaudit.pages.dev/?src=devto