explain mcp

Every AI app today needs its own special plug just to reach your files or calendar. Imagine if every sport used a different whistle. MCP is the one whistle every referee learns — and it changes how AI connects to your world. Think of MCP as a team huddle.

The Host is the coach — your AI app, like Claude Desktop or an IDE. The Client is the voice link to one player. That player is the Server, which holds a tool or piece of data. The protocol is the shared playbook every player knows.

Inside that playbook, three primitives: Resources are the scoreboard, Tools are the plays, and Prompts are the formations. Here is the full play, step by step. First, you ask the assistant a question. Next, the model inside the host looks at the request and chooses which move fits best.

Then the client sends a standardized JSON-RPC request to the server that owns the right tool or data. The server executes the real action on your files or account. The result returns along the same channel. Finally, the model uses that result to answer you.

Notice what did not happen: the model never touched your files directly. It called the play, and the player executed it. Under the hood, every message is JSON-RPC 2. 0 sent over standard input-output or HTTP, so the same playbook works whether the server lives on your laptop or in the cloud.

Each server exposes only what it chooses, so the rest of your data stays behind its defensive line. Here's the part that matters most: MCP is a universal language. One protocol, any host, any tool. The model calls the play; the server runs it.

Here's a typical MCP round trip. The host, Claude, asks: 'What did I write about MCP last week? ' It sees a search_notes tool, so it calls search_notes('MCP'). The client relays that call as JSON-RPC to the Notes server.

The server scans the notes file and returns the matching paragraph. The model quotes it straight back. The coach never walks to the bench—it just requests the scouting report. Here is the edge case.

A Notes server exposes only read-only tools. You ask the model to reply to a comment. The model offers to do it, but when it calls reply_to_comment, the server has no such tool. It returns an error.

The model adapts: it cannot post. The model never executes — the server owns the action, and here it says no. Here’s the map. Any AI app is the host, linked through an MCP client to an MCP server that runs the tool.

One protocol is the shared playbook between every app and every tool. The model calls, the server executes, and you stop writing custom plugs for every app. That is MCP.