0:01You've seen what MCP is for, and you've seen an app pick up new abilities by connecting to a server.
0:06Now we write one of our own.
0:08It is a small one, a single tool, in about ten lines of Python.
0:12Then we point our own app at it, so the two are talking.
0:15And then the test. We ask it something the model has no way of knowing by itself, and our server gives it a way to answer.
0:23Here is our app with no tools at all, the same one from the last video.
0:28One call to a model, and it prints whatever comes back.
0:31We're going to ask it something from today, a stock price, which changes every few seconds while the market is open.
0:37Let's open the terminal.
0:39The libraries this project uses are already installed.
0:42Here's our question to the model. We are asking the latest stock price for Nvidia.
0:47Let's run it.
0:48Obviously the model cannot answer that. Let's see how it says so.
0:52No access to real-time data, in its own words.
0:55It even names where to look, and Yahoo Finance is one of them.
0:59That is the gap we have to close. Because getting that number is not the hard part.
1:04There is a package that fetches it, and this file uses it.
1:07yfinance reads Yahoo Finance data. Unofficial, and the quotes can lag a little, but it works and it is one install away.
1:14It is one call, hand it a ticker, print what comes back.
1:18Let's bring the terminal back up.
1:21This is a plain API call. No model involved this time, and no MCP either.
1:26We get our stock price back.
1:28So the data is one call away. But currently our model has no way to reach it.
1:34Now let's build the MCP server for this.
1:37This is the server script. About ten lines of code. It's not that complicated.
1:41There are four parts to it, and the same map carries over to any server you read.
1:45The first is the server itself, with a name on it. Clients see that name when they connect.
1:50The second is a tool. One decorator above an ordinary function, and that function is now something a model can ask for.
1:57It takes one argument, a ticker symbol, and the type says it is text.
2:02Which the caller has to provide.
2:03And the sentence underneath says what it does, in plain English.
2:07The third part is the work, and there is barely any. Hand the ticker to yfinance, take the price, round it off.
2:14Nothing in there is about MCP. It is the same call the plain script made.
2:19And the fourth part starts it. From then on it reads requests on standard input and answers them.
2:26This line also picks how the two sides talk. There are two standard ways.
2:30Over standard input and output, where the host runs the server as a subprocess.
2:34Or over HTTP, where the server waits at a web address.
2:38We are keeping it local here.
2:40Later in the series we take this same server remote, and this is the line that changes.
2:45A name, a tool, the work, and a way to run. That is a server.
2:50Okay, our server is built. Now let's look at the client that connects to it.
2:55The imports come first, and one is the MCP SDK again, this time the client half of it.
3:01Right under them, the only line that says which server we mean. A command, and a file to run.
3:06For a remote server this would be a web address, ours is a program on this machine.
3:11The work itself is further down.
3:13It opens a connection, asks what the server offers, and hands those tools to the model.
3:18All of that is ordinary. About fifty lines, none of them about stocks, and you would write them once.
3:23So our app starts the server itself, as a second process, and the two talk over standard input and output. You never run server dot py yourself.
3:33That's both halves. Time to put them together.
3:36The same question we started with, word for word.
3:39The server started, it reported one tool, the model asked for it, and the price came back.
3:43That middle line is our function running, with the argument the model chose.
3:48That question handed the model the ticker symbol.
3:51The next question skips the ticker.
3:53Just the company name, the way anybody would ask it.
3:56The tool still got its ticker, and the price came back. Let's look at the middle line.
4:00The ticker came from the model. It already knew Apple's symbol, read what the tool takes, and put the two together.
4:07One more, to show what the loop does.
4:09This time the answer needs two prices, not one.
4:11And our server can only ever give one at a time.
4:15So the model made two calls on its own, one after the other, before it answered.
4:19It made one call for each ticker, with the same tool. That is the loop doing the work.
4:24And the comparing itself happened in the model, with the two numbers it got back.
4:30How did the model know what this tool takes? The decorator above the function had already worked that out, and the server sent it when our app asked what was on offer.
4:39Three things went across, all of them built from the function itself. Its name, the sentence written under it, and a schema for its one argument.
4:47So that one sentence deserved more care than any other line.
4:51To test a server before any app is involved, there is MCP Inspector, the standard tool for this. It connects to one server, lists what it offers, and calls a tool by hand.
5:02Next time we give this same server a resource and a prompt.

Build the Smallest Possible MCP Server

Created by Abhyuday DesaiPublished 2 Oct, v1Runs 5 min 11 sec

2 views

About this video

In Episode 06 of MCP Explained, we write a server of our own: ten lines of Python that wrap a stock-price lookup into one tool. We start from an app with no tools, watch the model fail on a live question, then build the server, connect our host to it over standard input and output, and ask again. Three questions show the model supplying a ticker symbol by itself and making two calls when the answer needs two prices.

You'll learn how to:

Build a server from four parts: a name, a tool, the work, and a way to run
Understand how one decorator above an ordinary function turns it into something a model can ask for
Know the two standard transports, stdio and Streamable HTTP, and which one this server uses and why
Connect a host to a local server with one line, a command and a file to run, launched as a subprocess
Watch the model fill in an argument it was never given, and loop through two calls for a comparison
Understand what actually crosses to the model: the tool's name, its description, and a schema, never your code
Know how to test a server before any app is involved, with MCP Inspector

This video is for developers comfortable with Python and calling a model API. Every command is shown on screen and runs from the companion repo.

What changed

v102 Oct 2026abhydesai5 min 11 sec3da3401

First published version.

2 views0 likes on this version