What Is an MCP Server? Explained the Way I’d Explain It to a Ten-Year-Old

My nephew asked me a question last week that stopped me for a second: “You always talk about AI — can it do literally anything?”

I thought about it and said, “AI is very smart, but when it starts out, it can’t actually use anything — it doesn’t even know how to press the buttons on a calculator.”

He didn’t believe me. This post is the longer version of how I explained it to him — and along the way, what an “MCP server” actually is.

Step 1: AI is smart, but its hands don’t listen

Picture the AI model as a brilliant scholar — one who’s read every book in the library, can recite classical poetry from memory, knows the history of every dynasty, but has never once left the house to do a chore. Send him to the corner shop to buy a bottle of water, and he won’t know how the till works. Ask him for tomorrow’s weather, and he has no window to look out of.

Everything in the scholar’s head, he got by reading. But buying water, checking the weather, reading someone else’s code — those are things you get by doing. They require hands. They require being plugged into the world outside your own head.

That’s the AI’s starting condition: it knows an enormous amount, and it can do almost nothing. It can’t hand you a real weather forecast, because it isn’t connected to any weather service. It can’t reliably do complicated arithmetic either — it’s doing the math in its head, same as you or me, and just as capable of getting it wrong.

Step 2: The old way — teach it every tool, one at a time, from scratch

So how do you fix that? The clumsy way is: engineers wire up external tools to the AI one by one — a calculator today, a weather API tomorrow, a calendar the day after.

The problem is that a calculator “speaks” one way, a weather service speaks another, and a calendar app speaks a third. It’s a bit like China before the Qin dynasty unified it: seven separate states, each with its own dialect, its own units of measurement, its own cart-axle width. A merchant traveling from one state to another had to relearn the local rules every single time — everyone doing things their own way, and nothing lining up between them.

Every time you bolted on a new tool, someone had to write a fresh batch of “translation” code: turn what the AI says into something the tool understands, then turn the tool’s answer back into something the AI understands. Add enough tools, and the translating never stops.

Step 3: Enter the MCP server — a shared language for everyone

MCP stands for Model Context Protocol. Stripped of the jargon, it’s simply an agreed-upon way for AI models and tools to talk to each other.

One of the first things the First Emperor of Qin did after unifying the six warring states was standardize the width between cart wheels and the written script — carts could suddenly travel any road in the empire, and a letter written in one province could be read in any other, no translation required. MCP does exactly that, except what it’s standardizing isn’t cart axles or characters — it’s the rules for how an AI and a tool talk to each other.

Once that shared rulebook exists, any tool can wrap itself in a thin layer that follows it, and that layer turns it into an MCP server. That layer does two things:

It publishes a list of everything it can do: “Here’s the menu — order whatever you like.”
When the AI “orders” something (sends a request), the server actually does the work, then translates the result back into something the AI can understand.

On the AI’s side, it only has to learn one set of ordering rules, and from then on it can read the menu at any MCP server on earth — whether that “restaurant” serves weather forecasts, stock prices, or the files sitting on your own computer. Different roads, same destination: the tools can be wildly different underneath, but the door they use to reach the AI looks identical every time.

Step 4: A tool I actually use

While writing code, I run an MCP server whose whole job is reading codebases. It scans an entire project up front and builds a map of it — which file calls which function, which piece of code belongs to which module, all recorded ahead of time.

Without it, if the AI needs to understand a project with dozens of files, it has to open and read them one at a time — slow, easy to miss something. With the server running, I can just ask “what does this project’s architecture look like,” and it hands over the map it already built. The AI glances at it and understands most of the shape immediately, no page-by-page reading required.

It’s the difference between cramming through an entire textbook the night before an exam, versus someone having already highlighted the important parts for you and handing you the summary. The tool does the tedious organizing work up front, and the AI just gets to use the finished result — twice the outcome for half the effort.

Step 5: A 2,300-year-old line that fits this exactly

There’s a line from the Xunzi, a Chinese philosophical text from the third century BCE, that I keep coming back to because it fits so precisely:

假舆马者，非利足也，而致千里；假舟楫者，非能水也，而绝江河。君子生非异也，善假于物也。

“Someone who borrows a carriage and horses isn’t unusually fast on foot, yet they can travel a thousand miles. Someone who borrows a boat and oars can’t swim, yet they can cross a great river. A person of wisdom isn’t born different from anyone else — they’re simply good at making use of things outside themselves.”

That was written about people, over two thousand years ago, but it describes an AI model almost exactly. On its own, an AI’s “legs” aren’t especially fast — it can still get arithmetic wrong, and it can’t see the real world. But the moment it borrows the “boat” that is an MCP server, it can check a real weather feed, read a real codebase, query a real database. It doesn’t get faster at what it already did; it gets access to an entirely new class of things it can do.

Confucius put a more practical spin on the same idea: “A craftsman who wants to do good work must first sharpen his tools.” An MCP server is that sharpened tool, sitting in the AI’s hand.

The end of the conversation

My nephew thought about all this and asked one last question: “So can anything, in theory, turn into an MCP server?”

“In theory, yes,” I told him. “If a piece of software wants an AI to be able to use it, it just wraps itself in that same layer of rules.”

He paused, then said: “So no matter how much it changes on the outside, it’s all built on the same root underneath.”

That’s a phrase he’d only just learned in school, and he’d landed it more precisely than I had. There will be more and more of these tools — weather ones, calendar ones, code-reading ones, probably homework ones eventually — but as long as they all speak this one shared language, the AI only has to learn it once, and it can go anywhere.