What's MCP?
Back in Lesson 3.5 you learned that programs talk to each other through APIs — agreed-upon ways to ask for things. The problem: every API is different. Slack’s is not GitHub’s is not your database’s. If Claude had to learn each one by hand, connecting anything new would be a project.
MCP is the fix.
MCP (the Model Context Protocol) is a single, standard way to plug a tool or data source into an AI. Build one MCP “adapter” for a system, and any AI that speaks MCP can use it.
Think of it like USB-C. Before USB, every device had its own weird connector. USB made one shape that fits everything — so a single port on your laptop charges your phone, reads a drive, and drives a monitor. MCP is that universal port, but for connecting AI to the world’s tools.
The piece that does the work: a server
When someone wants to connect a system to Claude, they build an MCP server for it. Don’t let the word “server” scare you — here it just means a small program that exposes one system through the MCP plug.
- A GitHub MCP server lets Claude open issues, read PRs, leave comments.
- A Notion MCP server lets Claude search your workspace and read pages.
- A database MCP server lets Claude run read-only queries against your data.
Claude is the thing plugging in. The server is the adapter on the other end. Once they’re connected, the server’s abilities just appear as new tools Claude can use — right alongside reading files and running commands.
Why this is the aha
Up to now, Claude’s world ended at your project folder. That’s the wall from Lesson 11.1.
MCP means Claude doesn’t have to live alone on your filesystem. The rest of your stack — tickets, messages, docs, data — becomes things Claude can read and act on directly.
And because it’s a standard, you’re not waiting on Anthropic to build each connection. Anyone can write an MCP server for anything, and it works with Claude the day it ships. There are already hundreds — for Linear, Sentry, Stripe, Google Drive, Figma, Playwright, and more.
What it is not
MCP is not magic, and it’s not a brain. The server only does what it was built to do — a read-only database server cannot delete your data, because nobody gave it that ability. Claude can only use the doors each server opens.
That’s worth holding onto, because it’s also the safety story: a tool Claude can reach is a tool someone deliberately plugged in. We’ll come back to that in the next lesson, because plugging one in is exactly what you’re about to do.
What’s next
Enough theory. Let’s plug one in — you’ll add a real MCP server and watch new tools show up in Claude, then talk about how to trust the thing you just connected.