All video notes
Video notes · MCP

Model Context Protocol (MCP) Explained for Beginners: AI Flight Booking Demo!

Video by KodeKloud on YouTube · 24 min · uploaded July 2025

Notes and analysis by BrainTube, written from the video. Not affiliated with KodeKloud. Watch on YouTube

In short

The video walks a complete beginner from chat apps to agents, APIs, MCP and Google's agent-to-agent protocol, then books a simulated flight through an MCP server in VS Code. The API section and the lab are clear and honest. Its core MCP pitch is off: MCP does not remove adapter code, it moves that code into a server someone still has to write.

The argument

The argument builds in steps. An LLM can only produce text, so a request to book a flight returns instructions, not a booking. An agent closes that gap by looping between the model, tools and memory until the job is done. Tools reach outside services through APIs, but every airline names its endpoints and fields differently, so an app talking to hundreds of them needs hundreds of adapters. MCP is presented as the fix: each service is described to the agent as a set of capabilities with input and output shapes, and the agent connects to MCP servers through a client. When one agent gets too broad, such as flights plus hotels, a second agent takes the new job and the two talk through the agent-to-agent protocol. A free lab then shows the difference between an agent with and without an MCP server.

Key ideas, with timestamps

  1. 1:34

    The FlyGPT problem

    A pretend app called FlyGPT gets a request to book a flight and replies with booking steps. That is the whole limit of a bare model: it can write text, images or video, but it cannot act. Everything after this point is about giving the model a way to act.

  2. 3:10

    What makes something an agent

    An agent talks to third-party tools, keeps memory of past conversations and preferences, and calls the LLM repeatedly until it has enough to decide and act. At 4:13 the example is an IDE agent finding which git commit removed a button by reading front end, back end and the git log through the terminal.

  3. 6:48

    From scraping to APIs

    Before APIs, sites like booking aggregators scraped airline pages and parsed the HTML for flight details. Airlines then offered structured endpoints for listing and booking flights. The section is aimed at people who have never met an API, and it says so, inviting others to skip ahead.

  4. 8:56

    Every API is different

    Three airlines expose flight search at three different paths and return three different shapes: one uses origin and destination, another from and to, a third start and finish. This is the real problem the video wants MCP to solve, and it is well illustrated.

  5. 9:59

    MCP as a capability description

    The MCP server tells the agent that an airline offers search flights and book flight, with defined inputs and outputs. The video says MCP came from Anthropic, was open sourced, and is now the default standard for building agents, and points to the official servers repository.

  6. 10:30

    Where the config file lives

    Each agent reads an MCP config that names the server, the command to start it and its arguments. The example runs a MongoDB server with a local connection string, so the agent can read and change data. Cursor keeps this file in a folder in your home directory, and Windsurf and Claude have their own paths.

  7. 12:35

    Agent-to-agent for a second job

    Adding hotel servers to the flight agent would bloat it, with a second set of preferences to remember. Instead a hotel agent is built and the flight agent calls it. Google's agent-to-agent protocol covers how agents discover each other's capabilities, hand off tasks, check status and return results.

  8. 18:17

    The lab: with and without MCP

    In a VS Code agent extension set up with a GPT-4.1 key from the lab, asking for flights from SFO to JFK makes the agent try to open a browser. After pasting a flight simulator server into the extension's MCP settings file at 19:18, the same request returns flights. Booking asks for approval and fails twice on phone number validation before it succeeds at 22:23.

What it gets right

The progression is well chosen for a true beginner. Showing the bare model failing to book, then the agent trying a browser, then the MCP server succeeding, makes each layer's job visible. The lab is the best part because the failure cases are left in. The phone number validation errors show that tool input schemas are enforced and that the model has to fix its own arguments.

The approval prompts before each call are also worth noticing. An agent that can book or write to a database should ask before it acts, and the demo treats that as normal.

The point about splitting agents at 12:03 is sound design advice. One agent with one job and one set of preferences is easier to test than a single agent that does everything.

What it leaves out or gets wrong

  • MCP does not remove adapter code. At 9:28 the pitch is that you no longer write code for each airline. In fact each airline still needs an MCP server that maps its own endpoints and fields to tools. The gain is that this adapter is written once, ideally by the provider, and any MCP client can reuse it. The flight simulator in the lab is exactly such an adapter.
  • "Default standard used by everyone" was an overstatement in July 2025. Adoption was fast, but many agent frameworks still used their own tool formats alongside MCP.
  • Tools, resources and prompts. The lab code has all three folders, but the video only explains tools and defers the rest to a later part.
  • Transports and security. The MongoDB example hands an agent a connection string with write access. Nothing is said about read-only users, scoped credentials, or the fact that tool descriptions from a third-party server go straight into the model's context.
  • Memory is asserted, not shown. Memory is one of the three parts of the agent definition, and flight preferences are the motivating example. The lab never stores or reads a preference.

How this applies to memory across AI tools

The video's agent definition puts memory inside each agent, and its hotel example gives each agent its own preferences. That works for one app. It breaks down when the same person uses several assistants, because each one starts with no knowledge of what the others learned.

The config file at 10:30 is the hint at a better pattern. If your notes and saved sources sit behind an MCP server, any client that reads that config can search them. That is the idea behind an MCP memory server. BrainTube is one example: it saves YouTube videos with their transcripts, articles, PDFs and podcasts into one compiled knowledge base and exposes it as an MCP server, so Claude, ChatGPT, Cursor and other MCP clients can search it. For the protocol basics behind this, see what is MCP.

Who should watch it

Watch it if you are new to the whole topic and want one video that connects LLMs, agents, APIs, MCP and agent-to-agent without assuming code. If you already know what an API is, skip 5:46 to 8:25 as the video suggests. Developers who already use agents can jump to 10:30 for config files and 15:43 for the lab, which is the most useful stretch. Skip it if you want to build an MCP server; that is promised for a later part, and this one never explains resources, prompts or transports.

Frequently asked

What is the difference between an AI agent and MCP?
An agent is a program that loops between a model, tools and memory until a task is done. MCP is a protocol that standardises how an agent discovers and calls tools from outside services. An agent can work without MCP, but MCP lets it reuse servers written by others instead of custom integrations.
Does MCP replace APIs?
No. An MCP server usually wraps an existing API. The server describes what the service can do and how to call it, then makes the normal API request when the agent picks that tool. Someone still writes that wrapper once per service.
What is the agent-to-agent (A2A) protocol?
A2A is a protocol started by Google for agents to work with other agents. It covers how one agent learns what another can do, assigns it a task, checks the task's status and gets results back. MCP connects an agent to tools and data; A2A connects agents to each other.

Keep videos like this one searchable

BrainTube saves YouTube videos with their transcripts into one compiled knowledge base that Claude, ChatGPT, Cursor and other MCP clients can search.

Start free

Go deeper

More video notes

Are you the creator and want this page changed or removed? Contact us and we will act within 48 hours.