0:01So far, our app has held one connection at a time. This is the stock server we built.
0:05Earlier, we also connected to a weather company's server.
0:08In this video we add a file server that someone else wrote.
0:11Now one app holds all three connections, with one MCP client for each.
0:15The app changes too. It becomes a more capable agent. We hand it a task, and the model decides which tools to call.
0:22Let's read the agent first, then put a real task through it.
0:27This is the agent, one file, sixty one lines. Most of it is the same host app this series has been running since we first plugged in a server. Let's read what changed.
0:37The first change is at the top. Instead of one server address, the app holds a list of servers, one entry for each.
0:43The first entry starts our own stock server as a subprocess.
0:47The second is the address of the weather company's remote server.
0:50The third runs a file server published by the MCP community, we install it and run it without ever reading its code.
0:56This list is what every MCP host asks for. Apps like Claude Desktop or Cursor take these same entries in their settings, a command with its arguments, or an address. They connect the same way.
1:08Next, what the app does with that list.
1:11For every server on the list, the app opens its own MCP client and connects it. Three servers means three clients, one client per connection.
1:20Every tool each server reports goes into one shared pool for the model. And for each tool name, the app records which client owns it.
1:27Below that is the loop that serves the model.
1:30This loop is unchanged from our first host app. Ask the model, serve its tool calls, and stop when it answers.
1:37The one addition is here. Before serving a call, the app looks up which client owns that tool name, so each call reaches the server it belongs to.
1:46And that is the whole file. Everything new is the list, the pool, and the lookup, the rest was already written. Let's give it a task.
1:54Time to run it. The first task is a small weather question, so we can watch the connections come up.
1:59The whole task goes in as one argument, the same way our host app took its questions.
2:04All three servers start and report their names, and together they put thirty five tools in front of the model. Let's read the run from the top.
2:12These names come from the servers themselves, our stock server, the weather company's server, and the community file server.
2:19The file server also prints a notice of its own. It names the one folder it has been given to work in.
2:24Below the names come the model's calls.
2:26Out of thirty five tools, the model called three, all weather. It read the tool descriptions and picked what this question needed.
2:34The forecast came back through the client, and the model turned it into a direct answer.
2:38Next, a task that needs all three servers.
2:41Now the task this agent is really for.
2:44A morning briefing, Apple's price and its news, the San Diego weather, and a copy written into a file.
2:50The connections come up, and the first calls head to the stock server.
2:54Here is the finished run, eight calls, answer at the bottom.
2:57The file it wrote is already in the folder. We will look at it next, but first, let's walk the model calls in the terminal.
3:04The first two went to our stock server, the price, then the headlines.
3:08Then the weather service. The model searched for the city a few times, then asked for today's forecast.
3:14And the last call went to the file server.
3:16The model passed it a file name, and the entire briefing as the content to write.
3:21That is the whole run. Let's close the terminal and open what it wrote.
3:26This is the briefing as written. The price and the news came from our server, the weather from the weather company, and the file server put it on disk.
3:36That is what it means to give your agent MCP. The agent doesn't ship with fixed abilities, it collects them from the servers on its list, every time it starts.
3:45Adding an ability is adding a server to the list. We gained weather and file writing today without writing weather code or file code. Agent frameworks that support MCP do this same discovery for you.
3:56One of today's calls is worth a second look. The model chose to write a file onto our disk, and the file server did it.
4:02Whether a server should be allowed to do that is a real decision. Next time we look at what can go wrong when a model holds the controls, and how to decide what a server may touch.

Give Your Agent MCP

Created by Abhyuday DesaiUpdated 4 Oct, v2Runs 4 min 19 sec

1 view

About this video

In Episode 10 of MCP Explained, one small agent holds three connections at once: our stock server, a weather company's remote server, and a community file server we never read the code of. We read what changed in the host, a list, a pool, and a lookup, then give it two tasks. The second is a morning briefing that needs all three servers and ends with a file written to disk.

You'll learn how to:

Turn a one-server host into a multi-server agent with three additions: a server list, a shared tool pool, and an owner lookup per tool name
See the same list entry every MCP host takes, whether it is your own script, Claude Desktop, Cursor, or VS Code
Watch thirty five tools arrive from three servers, and the model call three of them for a weather question
Follow an eight-call task across all three servers, and read the briefing it wrote
Understand that abilities arrive at run time: adding an ability is adding a server to the list

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 2026abhydesai4 min 19 sec8906869
Scene 2 now builds the picture from what the series has run, and the app card becomes agent.py. Scene 5 gets its click sounds where a file opens after the terminal closes.

Same chapters. Transcript: 5 of 41 lines changed. 6 sec shorter.
Show the differencesHide the differences
1. ~~Here is where the series stands. One app on the left, three servers on the right, and one MCP client inside the app for each connection.~~
  ~~So far our apps have held one connection at a time. The stock server is the one we built, the weather server belongs to a weather company, and the file server is new today.~~
  ~~The app is new too. It is a small agent. We hand it a task, it holds all three connections, and the model decides which tools to call.~~
  
  So far, our app has held one connection at a time. This is the stock server we built.
  Earlier, we also connected to a weather company's server.
  In this video we add a file server that someone else wrote.
  Now one app holds all three connections, with one MCP client for each.
  The app changes too. It becomes a more capable agent. We hand it a task, and the model decides which tools to call.
1 view1 like on this version

v102 Oct 2026abhydesai4 min 25 sec805b43c

First published version.

0 views0 likes on this version

Watch v1