All video notes
Video notes · MCP

What is MCP? Integrate AI Agents with Databases & APIs

Video by IBM Technology on YouTube · 4 min · uploaded February 2025

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

In short

IBM's short video explains MCP through its three parts, host, client and server, and walks through one tool call from question to answer. The request flow is correct and is the useful part. It wrongly calls the protocol itself a transport layer, blurs the host and client, and at under four minutes leaves out resources, prompts, transports and security.

The argument

MCP is an open standard that lets an AI agent reach data sources such as databases, APIs and local files through a common interface. An application, the host, contains one or more MCP clients and can connect to any number of MCP servers. Each server sits in front of a data source and exposes tools. When a user asks a question, the host gets the list of tools from its servers, sends the question and tool list to the LLM, calls whichever tools the model picks, and passes the results back to the model for a final answer. The video's message to agent builders is that this one standard lets any data source reach any agent.

Key ideas, with timestamps

  1. 0:13

    Host, client, server

    The host is the application, such as a chat app or a code assistant in an IDE. It contains an MCP client, or several. It can connect to as many MCP servers as you like, and the number of servers does not change how the host works.

  2. 1:04

    Servers sit in front of data

    When the host needs a tool, it calls an MCP server, and the server talks to the data source. The video stresses that the source type does not matter: a relational or NoSQL database, an API of any style, local files or code. Local file and code access is what makes MCP useful for IDE assistants.

  3. 2:03

    Step one: discover tools

    In a chat app, a user asks something like what the weather is in a place or how many customers they have. Before the model sees the question, the host asks its MCP servers which tools are available, and each server replies with its list.

  4. 2:23

    Step two: the model chooses

    The host sends the question plus the available tools to the LLM. If all goes well, the model replies with which tools to use. The host then knows which servers to call, since each tool belongs to one server.

  5. 2:43

    Step three: execute and answer

    The chosen server executes the work against its database, API or local code and returns a result, possibly after several calls. The host sends the results back to the LLM, which writes the final answer shown in the chat.

What it gets right

The request flow between 2:03 and 3:07 is a clear, short explanation of an MCP tool call. It makes one point many people miss: the model never talks to the MCP server. The host discovers the tools, the model only chooses among them, and the host makes the call. That division explains both what MCP standardises (how hosts find and call tools) and what it does not (how well the model picks).

It also gets the many-to-many shape right. One host can use many servers, and a server written once works with any host. For a team deciding whether to wrap an internal database, that is the argument in one diagram.

What it gets wrong or leaves out

  • MCP is not a transport layer. At 0:57 the video calls the protocol a transport layer in the middle. MCP is a message format and set of rules (built on JSON-RPC) that runs over a transport. Local servers usually run over stdio as a subprocess; remote servers run over HTTP. The difference matters when you deploy a server, because it decides how auth and sharing work.
  • Host and client are blurred. The video often says host and client together. The distinction is that the host is the app and runs one client per server connection. That design keeps servers isolated from each other.
  • Only tools. Servers can also expose resources (data the app can load) and prompts (templates the user picks, often as slash commands). A video about connecting agents to databases would have been the place to mention resources.
  • "If all is well". The model choosing the right tool depends heavily on tool names and descriptions. With many overlapping servers, wrong choices become common.
  • Security. A server that can run queries against a customer database is a privileged component. The video gives no advice on permissions, read-only access or reviewing third-party servers.

The video dates from February 2025, early in MCP's adoption. The spec has changed since, notably around remote transports and auth, so check the current version before building.

How this applies to memory across AI tools

The same flow works for personal knowledge. Put your notes and saved sources behind an MCP server, and any host, whether a chat app or an IDE assistant, discovers a search tool, the model decides when to call it, and the results come back as context. That is what an MCP memory server means.

BrainTube applies this: 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 a fuller introduction to the protocol, see what is MCP.

Who should watch it

Watch it if you have heard of MCP and want the basic architecture in under four minutes, for example before a meeting where it will come up. The useful part is the worked example from 1:36 to 3:07. Skip it if you already know what an MCP host and server are, or if you need anything you can build from: there is no code, no named servers and no setup.

Frequently asked

What is the difference between an MCP host, client and server?
The host is the AI application, such as a chat app or IDE assistant. Inside it, an MCP client manages the connection to one MCP server. The server sits in front of a data source or service, such as a database or API, and exposes tools the model can use.
Does the LLM call the MCP server directly?
No. The host asks its MCP servers for their tools, sends the user's question and the tool list to the LLM, and the model replies with which tool to use. The host then calls the server and returns the result to the model, which writes the answer.
Can MCP connect AI agents to a database?
Yes. An MCP server can wrap a relational or NoSQL database and expose tools such as running a query or counting records. Any MCP host can then use it. Limit the server's database permissions, ideally to read-only, because the model decides when to call it.

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.