Getting started: MCP servers
A beginner's guide: understand what an MCP server is, add it to your AI client, and above all check it's safe before connecting it.
1. What is an MCP server?
The Model Context Protocol (MCP) is a standard that lets an AI agent (Claude, Copilot, internal agents…) call tools: read a file, query a database, search the web, send a message. An "MCP server" is a program that exposes a set of tools to the AI, in a standardized way.
Concretely, connecting an MCP server means giving your AI the right to run actions through those tools. Hence the importance of knowing what it really does.
2. Remote or package?
Remote: the server is a URL hosted by the publisher; your client connects to it (nothing to install).
Package (npm, PyPI, Docker…): code you install and run yourself. More control, but your machine runs third-party code.
3. Direct or with configuration?
Some servers work directly; others need configuration (API key, environment variable). A server's MCP Hub fiche shows what it needs.
4. Adding it to Claude Desktop
Claude Desktop reads a config file. Add your servers under mcpServers.
A package server (npm)
{
"mcpServers": {
"my-server": {
"command": "npx",
"args": ["-y", "the-mcp-package-name"]
}
}
}
A server with an API key
{
"mcpServers": {
"my-server": {
"command": "npx",
"args": ["-y", "the-mcp-package-name"],
"env": { "API_KEY": "your_key" }
}
}
}
A remote server (via the mcp-remote bridge)
{
"mcpServers": {
"my-web-server": {
"command": "npx",
"args": ["-y", "mcp-remote", "https://example.com/mcp"]
}
}
}
Restart Claude Desktop after editing. The server's tools then appear in the interface.
5. Verify BEFORE adding (the security reflex)
- its verdict (did it pass the behavioral pentest?);
- the list of its tools and their parameters;
- its token cost (the context it consumes);
- give the AI the least access possible and prefer verified servers.
6. Reading an MCP Hub fiche
Each server fiche shows: the verdict (Verified / Flagged / Not tested), the context cost, and the list of tools — for each: parameters (type, required, allowed values), return schema, a sample call, and an estimated cost. Reminder: only the behavioral pentest yields a verdict; description analysis alone is merely indicative.