In short MCP (Model Context Protocol) is an open standard by which an AI agent connects to external tools and data. One shared "socket": connect an MCP server once and any MCP-capable client (Claude Desktop, Claude Code, your app) can use it without reinventing the integration.
The problem MCP solves
Tools give an agent hands (the agent loop and tools). But before, each integration was written by hand and anew for each client: describe the schema, handle the calls, maintain it. Ten tools × three apps = thirty different wirings.
MCP standardises this. A server describes its capabilities once over a shared protocol, and a client (the agent) can talk to any such server. It's like USB: not a bespoke port per device, but one common one.
Analogy: a plug and socket to one standard. A device maker needn't know the building's wiring — they make a plug to the standard, and the device works everywhere.
How it works
- An MCP server — an "adapter" to a resource: a database, GitHub, the filesystem, a third-party service. It exposes a set of tools (and sometimes data) over the MCP protocol.
- An MCP client — the agent/app that connects servers and lets the model call them.
- The protocol — the shared rules of how they talk: how the server says "here's what I can do," how the client calls and gets the result.
Once the model has access to an MCP server, it sees its tools as ordinary tools and calls them in the agent loop.
Why it matters to you
- No hand-written schema (Claude Platform 101): connect a ready MCP server and the agent already works with the service.
- One integration, many clients: the same server works in Claude Desktop, in Claude Code, and in your code.
- An ecosystem: many services already ship MCP servers — you connect them, not write them.
- In Claude Code 101 and Claude Code setup MCP is the standard way to connect external systems to the terminal agent; in Cowork and connectors MCP is often what's "under the hood."
The safety line
An MCP server gets access to a real resource and can act. So the same rules apply as for any tools: permissions and review (the agent loop and tools). Connect trusted servers and understand what each one is allowed to do.
Related
The source is the Claude Platform 101, Claude Code 101 and Introduction to Claude Cowork courses. Related: the agent loop and tools, Claude skills, Claude Code setup.