What MCP Actually Is, and Why "Model Context Protocol" Is a Misleading Name

You will see MCP mentioned constantly in AI tooling discussions. The name "Model Context Protocol" makes it sound like something about context windows or memory. It isn't. That confusion causes people to misread what MCP is actually useful for.

The jargon

Protocol — an agreed format two pieces of software use to talk to each other. HTTP is a protocol. USB is a protocol. A protocol means you don't need a custom cable for every device.

MCP server — a small program that wraps a tool (a database, an API, a file system, a calendar) and exposes it in the standard MCP format.

MCP client — the thing consuming those tools. Claude Code is an MCP client. So is Claude Desktop. So is any app built on the API that implements the spec.

The lesson

MCP is a plug standard, not a memory system. Think USB-C. Before USB-C, every device had its own cable. After USB-C, any device works with any charger that supports the spec. MCP does the same thing for AI tools.

Before MCP, if you wanted Claude to query your database, you wrote custom tool definitions in your API code. If you then wanted your n8n workflow to do the same thing, you wrote the integration again. Every AI client needed its own wiring.

With MCP, you write the integration once as an MCP server. Any MCP client — Claude Code, Claude Desktop, a third-party agent framework — can pick it up without you touching it again.

How it works

An MCP server exposes three things:

• Tools — actions the AI can call (run this query, send this email).
• Resources — readable data the AI can pull in (this file, this database record).
• Prompts — reusable prompt templates the server makes available.

The client connects to the server at startup, asks "what can you do?", and gets back a list. From there it works exactly like the tool-use loop you already know — define, detect, execute, return.

Claude Code connects to MCP servers via your config file (claude_desktop_config.json on Desktop, or the mcpServers block in Claude Code's settings). Each entry points to an executable. Claude Code launches it, queries its capabilities, and treats those capabilities like built-in tools.

"mcpServers": {
  "my-database": {
    "command": "node",
    "args": ["C:/tools/my-db-mcp/index.js"]
  }
}

When to reach for it / when not to

Reach for MCP when you want the same tool available across multiple AI clients, or when you're building something you'd share with others. The one-time setup cost pays off quickly if you'll reuse it.

Don't bother for a one-off task. If you need Claude to do something once, a direct API tool definition is faster and simpler than standing up a server.

Try it

Open the MCP server registry on GitHub. Find one server that connects to a tool you already use — Obsidian, a database, a browser. Read its README and estimate how long the setup would take. You don't need to install anything today. The goal is to make MCP concrete by attaching it to something real in your stack.