0:01Everything so far has been theory, what MCP is, and the parts every connection needs.
0:06This is the host, a program that talks to a model, with its MCP client inside.
0:11Today we see it in action. Our host is one file of Python, about fifty lines, so we can read the whole thing.
0:18The server we want is already built and running. A weather company made it public, and we did not write any of it.
0:24Our script gets the address, and its client does the connecting. That is the whole plug-in step.
0:29So the plan is simple. We ask our script something it cannot answer. Then we connect the server and ask it again.
0:36All of this is in one public repo, so you can clone it and run every command you are about to see.
0:42The readme is short. What the two scripts are, how to set up, and how to run.
0:46Setup is three commands. A virtual environment, the packages, and your own env file. Let's look at those last two.
0:55This is the setup it needs. Which provider, which model, and your own key. Copy this, fill in your three values, and it runs.
1:04Three packages, pinned to the exact versions this video was captured with.
1:08The openai package is how our script calls the language model.
1:12It doesn't lock us into OpenAI's models, though. Any provider with an OpenAI compatible endpoint works, and switching is a config change.
1:20And the second one is mcp, the Python SDK for the protocol. The connected version of our script is built on it.
1:29We start from the version with no tools at all.
1:32The system line hands the model today's date, since a model has no clock. That will matter when the answer comes back.
1:39Then it builds an OpenAI client from two of the values in the env file, the base URL and the key.
1:45Then one call to that client, with the model name from the env file and our question, and it prints the answer. That's the whole job.
1:53Now let's open a terminal and run it.
1:54The virtual environment is already made, with the three packages installed.
1:58Here's our question. We're asking the script whether I should carry an umbrella in Paris tomorrow. It goes in as one argument, in plain English.
2:06A question about tomorrow, which is exactly the kind of thing a trained model cannot know.
2:11It says so itself. No access to live weather, go check a weather site.
2:16So that is the problem. Now the fix to it.
2:19This is the same program with the server plugged in. About fifty lines, where the first one had seventeen.
2:26There are four additions, and we will take them in the order the file runs them.
2:30The first addition is the address of the server. The key in it is public, the same one AccuWeather hands to everybody.
2:36It is written once, in plain sight.
2:39Two constants and one function, that is the whole structure.
2:43And inside the function, the order it does things in is the part worth watching.
2:48We open a connection and ask the server what it offers, and we do that at startup, before the model has seen the question.
2:54Each tool arrives with a schema for its arguments, already in the format the model's API expects. So we wrap them and pass them straight along.
3:02So the model can ask for a tool now. Something still has to run it.
3:06The rest of the function is one loop.
3:08Ask the model, and as long as it keeps asking for tools, call them and hand back the results, until it answers instead of asking.
3:16Let's run this one.
3:19Same question as before, word for word. Only the file name changed.
3:23Same model, same machine, the one thing that changed is the server it can reach.
3:28That went by quickly, so let's walk it back from the top.
3:30Before any of the thinking, it connected, and the server reported its own name.
3:35Then it asked what was on offer, and got eighteen tools back, in two matching sets. All of them came from the server, our file only holds its address.
3:45That whole list was gathered before the model saw our question.
3:48The rest of the run is the model at work. It has our question now, and eighteen tools to choose from.
3:53The model picked two of them, in order. One to turn the word Paris into a place the server knows, then one to get that place's forecast.
4:01And now it can answer. The date is right because we handed it today's date, and the rain number came back from the server.
4:08That answered the question we started with. Two more are worth trying.
4:12This next one has nothing to do with weather at all.
4:14This time we're asking the capital of France. Same program, same server, so the same eighteen tools get offered.
4:21And that took one step, where the weather question took three.
4:25The connection still happened, and all eighteen were still put in front of the model.
4:29It used none of them and just answered. Tools are an offer, and the model is the one deciding whether to take it.
4:36Now the other side of that.
4:38Something this server has no tool for.
4:40Now we're asking for Apple's current stock price. That's current and real, and nothing on offer here goes anywhere near it.
4:48It connects, sees the same list of tools, thinks for a moment, and then turns it down.
4:53Everything it was handed is weather, and it says so, which is the correct answer for this program.
4:59We just gave our app the ability to answer weather questions. And there is no weather code in it.
5:05All of that came from the server.
5:07Change that one address, and the same app has a different set of abilities.
5:11And there are many more addresses to choose from. An official registry lists the servers people have published, and apps that support MCP often show it as a catalog you can browse.
5:21Now we have seen how to use an MCP server. Next time we open a server up and look inside one.

Plug In Your First Server

Created by Abhyuday DesaiUpdated 4 Oct, v2Runs 5 min 33 sec

1 view

About this video

In Episode 05 of MCP Explained, the theory goes live. Our host is one file of Python, about fifty lines. We ask it about tomorrow's weather and watch it fail, then plug in a public weather server somebody else built and ask again. Eighteen tools arrive at startup, the model picks two, and the answer comes back with a real forecast. We also see what happens when a question needs no tool, and when the server has no tool for it.

You'll learn how to:

Read a minimal host: one call to a model, made with the OpenAI SDK, and the three values it needs to run
See the four additions that connect a server: the address, the connection at startup, the tool wrapping, and the loop that serves tool calls
Watch discovery happen before the model sees the question, with every tool coming from the server
Follow the model choosing two tools in order, and see why the answer's date and its rain number come from different places
Understand that tools are an offer: the model can answer without them, and turns down a question the server cannot help with
Learn where servers come from: an official registry, surfaced as catalogs in apps that support MCP

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

v204 Oct 2026abhydesai5 min 33 sece72f5eb

One line re-voiced, where the OpenAI client is built from the env file, so it reads more clearly. Nothing else changes.

Same chapters and transcript.

0 views0 likes on this version

v102 Oct 2026abhydesai5 min 32 sec805b43c

First published version.

1 view0 likes on this version

Watch v1