# Trusting Servers — Agent Frameworks and MCP Basics

Source: https://www.geekswithgeeks.com/en/agent-frameworks-mcp/sec-trust-supply-chain

> Treat MCP servers as code you run and vet them like any dependency.

## A server is code

A local MCP server is a program on your machine with your file and network access. A remote one may see everything you send it. Prefer official or well-reviewed servers, pin versions, read what a server does before installing, and avoid copy-pasting setup commands from unknown posts.

## Every boundary is a risk

Servers run code, tool results enter the model's context, and the model can be steered by text. Each link needs a defence.

![Four checkpoints: trust, scope, approve, log.](assets/figures/agent-frameworks-mcp/section-6-map.svg) — Figure 6.1 — Trust, scope, approve and log.

## Beware tool descriptions that change

A server can change its tool descriptions after you approved it. Treat description text as untrusted input to the model, and re-review servers after updates.

**Quiz:** What is the safest attitude toward a third-party MCP server?

- [ ] Install anything with many stars blindly
- [x] Treat it like any code dependency: review, pin and limit it
- [ ] Run it as administrator for compatibility
- [ ] Disable all logging

*Answer:* Treat it like any code dependency: review, pin and limit it. A server executes code with your permissions, so the usual supply-chain care applies.