In the previous chapter you learned what agents are and that they draw their power from tools. But how do those tools actually reach the AI? This is exactly where a word comes in that you've suddenly been hearing everywhere in 2024 and 2025: MCP.

The problem before, everyone had their own plug

Until late 2024, the situation looked roughly like this: you wanted to connect an AI to your database? You had to build that yourself. You wanted to hook it up to Slack? A separate integration. To GitHub? Different again. To your own internal tool? Entirely by hand.

And the worst part: every AI platform had its own way of how these connections had to look. An integration you'd written for ChatGPT didn't work with Claude. One for Claude didn't work with Gemini. It was chaos, and a real pain for developers.

Analogy

Picture the time before USB: every device had its own plug. Printer, mouse, keyboard, external hard drive, scanner, all with different connectors. Your desk was buried in cable clutter. Then USB arrived: one plug that fits everywhere. Suddenly you could plug any device into any computer without checking whether the connectors matched. MCP is exactly that for AIs.

What exactly is MCP?

MCP stands for Model Context Protocol and is an open standard that Anthropic introduced in late 2024. By now it's also supported by other providers, and that's exactly what makes it so interesting.

A protocol is nothing more than an agreed-upon language in which two programs talk to each other. When two sides stick to the same protocol, they can understand each other even if they know nothing about one another. That's how email works (SMTP), that's how the web works (HTTP), and that's now also how AIs connect to tools.

Key takeaway

MCP is not an AI model. MCP is not a harness. MCP is only the language in which a harness talks to an external tool. So it sits between the AI application and the world of tools.

MCP is the USB port for AIs. Build it once, and the tool fits everywhere.

Server and client, the two sides

MCP defines two roles:

When you connect a client to a server, the server says: "Here are my tools. They're called this. They do that. They need these arguments." The client then passes these tools on to the AI model, which can use them just like its own tools.

MCP, a shared port for many tools
A central client (left) connects to any number of MCP servers (right).
Claude Desktop MCP client Filesystem MCP server GitHub MCP server Postgres DB MCP server Slack MCP server
Every MCP server provides tools or data. The client (e.g. Claude Desktop) can use all of them without a separate format having to be invented for each connection.

How does an MCP call actually work?

Step 1. Establish a connection

On startup, the MCP client (e.g. Claude Desktop) connects to one or more MCP servers. That can be local (a server on your machine) or over the network.

Step 2. Announce tools

The server sends the client a list of its tools along with a description. For example: "query_table, runs a SQL query on the database, expects one argument, query."

Step 3. AI receives the list

The client (harness) passes these tools on to the AI model, together with the tools the harness already has itself. To the model, both look the same.

Step 4. AI makes the call

If the AI wants to use a tool, it says so as usual: "Please run query_table with query=SELECT * FROM customers." The client recognizes: that belongs to the MCP server, and forwards the request.

Step 5. Server works, response comes back

The MCP server executes the request (database query, API call, file access, whatever it's capable of) and sends the result back to the client. The client passes it on to the model, which uses it for its answer.

What you can do with it

01

File system

A local MCP server gives the AI access to specific folders. It can read, write, and search, without your entire hard drive lying open.

02

Database

An MCP server that talks to your Postgres or SQLite database. "Ask the AI how many orders there were in Hamburg last month," and you get the answer from the real data.

03

Slack & Notion

Connect the AI to your team workspace. It can read Notion pages, summarize Slack threads, create new notes, as if it were a new team member.

04

GitHub & Jira

Search issues, open pull requests, comment on tickets. The AI becomes a coworker who really knows your project board.

Example

You use Claude Desktop and want to look at last week's newsletter send. Before: open Mailchimp, click through manually, copy it into a report. Now: you connect a Mailchimp MCP server once, then simply type into Claude "What was the open rate of the last three newsletters?", and Claude accesses it live, summarizes it, suggests improvements. No new UI, no new tab.

Why this is such a big deal

Until now, it was every individual AI tool's job to build its own integrations. Cursor needed a GitHub connector. Claude Desktop needed a GitHub connector. ChatGPT needed a GitHub connector. Three times the work for exactly the same tool.

With MCP, there's one GitHub MCP server. You connect it to anything that speaks MCP. That's a massive efficiency gain for the whole industry, and it means that even small tools and niche software can suddenly get AI connectivity that used to be unaffordable.

What MCP is not

Important, so you can place it correctly:

Security, important here too

Careful

An MCP server gives the AI real power over real systems. If you install a database MCP server that talks to your production database, the AI can theoretically do anything the server allows there, including deleting data. Configure MCP servers as narrowly as possible: read-only where write access isn't needed; only a specific table range; only approved folders. And don't install MCP servers from unknown sources.

There's a growing selection of official and community MCP servers. Stick to the established ones, or read the code before you let something unfamiliar run on your system. That applies not only to MCP, but especially here, because a malicious server could suddenly hand the AI tools nobody expected.

Where this is heading

We're only at the beginning. If MCP continues to catch on, and it looks pretty likely, in a few years every piece of software you work with daily will have an MCP server. Your AI will then no longer run "alongside" your tools, but inside them. It will know what's on your calendar, what's happening in your CRM, what's new in your code repository, without you having to copy or explain anything.

Key takeaway

If you hear today "Tool X is getting an MCP server" or "We're building this with MCP," that means: Tool X is becoming AI-capable, and not just for one specific AI, but for all that speak MCP. That's the real charm of this standard.

In three sentences

What you now know

  • MCP is an open standard from Anthropic that gives AIs and external tools a shared language, the "USB plug" of the AI world.
  • An MCP server provides tools or data (file system, database, Slack, GitHub), an MCP client (e.g. Claude Desktop) uses them.
  • MCP doesn't make AI smarter, but far more connected, and in the long run makes many platform-specific integrations unnecessary.