MCP explained simply: what Model Context Protocol is and why IT people should care

MCP is why Claude knows your files without you pasting them in. Here’s how it works in 3 minutes.

Before MCP, integrating an AI with your systems was a custom job every time. You wanted Claude to read your files? A one-off integration. Query the database? Another integration. Every AI tool had its own way, every connection had to be rebuilt from scratch. The result was an AI living in a bubble: it knew everything about its training set, but nothing about what was actually happening in your infrastructure right now.

The right analogy: USB for AI integrations

Before USB, every device had its own cable. Monitor: DisplayPort. Keyboard: USB-A. Phone: Micro-USB. Mac: proprietary Thunderbolt. USB standardized everything: same connector, same protocol, any device.

MCP does the same thing for AI integrations. It defines a standard protocol that an AI client — Claude Desktop, Claude Code — uses to talk to any external system without custom integrations. If someone writes an MCP server for GitHub, it works with Claude and with any other client that supports MCP. The work is done once.

How it works: clients and servers

There are two pieces.

The MCP server is a process that exposes resources and tools. A filesystem server exposes “read this file”, “list this folder”, “write here”. A GitHub server exposes “list open PRs”, “read this issue”, “create a comment”. Anyone can write one in Python or Node.js using the official SDK.

The MCP client is the AI. When Claude needs to do something, it talks to the appropriate MCP server using the standard protocol. Communication happens via stdin/stdout for local processes, via HTTP for remote servers. You don’t need to know these details to use it.

Servers ready to use today

You don’t need to write anything to get started. Servers already exist for the most common use cases:

• @modelcontextprotocol/server-filesystem — reads and writes local files and folders
• @modelcontextprotocol/server-github — PRs, issues, repositories, code search
• @modelcontextprotocol/server-postgres — SQL queries on PostgreSQL
• @modelcontextprotocol/server-slack — messages, channels, users
• @modelcontextprotocol/server-gdrive — Google Drive files and documents

The full list is at github.com/modelcontextprotocol/servers. There are dozens between official and community-maintained.

To configure them in Claude Desktop, edit the JSON file in the appropriate location (~/Library/Application Support/Claude/claude_desktop_config.json on macOS, %APPDATA%\Claude\claude_desktop_config.json on Windows):

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/stefano/Documents"]
    },
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": { "GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_..." }
    }
  }
}

Restart Claude Desktop. From that point on you can say “read the file report.pdf in Documents” or “show me the open PRs on my-project repo” and Claude does it directly.

In Claude Code it’s even faster: claude mcp add filesystem -- npx -y @modelcontextprotocol/server-filesystem /path/to/folder.

Why a sysadmin should build their own MCP server

This is the interesting part. A company MCP server lets you expose any internal system to Claude in natural language.

System logs: a server that reads /var/log/ and lets you search for error patterns or do time-based correlations — instead of grep. Monitoring: if you use Prometheus or Zabbix, expose real-time metrics and ask “why did web01 slow down last night at 2:30 AM?”. Ticketing: an MCP server on Jira or Redmine and Claude creates tickets, updates them, searches without switching interfaces. Inventory: expose your CMDB or even just a CSV and ask “which machines still have Windows Server 2012?”

The official Python SDK installs with pip install mcp. Documentation is at modelcontextprotocol.io.

What to do

• Install Claude Desktop (free, claude.ai/download) and add the filesystem server pointing to a working folder: it’s the fastest way to see MCP in action
• Add the GitHub server if you work with repositories — the difference when doing code review is immediate: Claude reads comments and code without you pasting anything
• Identify an internal system you’d use more readily if you could talk to it in natural language: logs, ticketing, inventory — that’s your candidate for your first custom MCP server