All video notes
Video notes · MCP

MCP vs API: Simplifying AI Agent Integration with External Data

Video by IBM Technology on YouTube · 13 min · uploaded May 2025

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

In short

IBM's video explains that MCP and REST APIs are both client-server abstractions, but MCP is built for AI agents: servers list their tools, resources and prompts at runtime, and every server answers the same calls. Its key conclusion, that MCP servers usually wrap existing APIs and so the two are layers rather than rivals, is correct. It undersells existing API description formats and skips the costs of dynamic discovery.

The argument

The video's case is that the right question is not MCP or API but where each sits. APIs are general-purpose contracts between programs, each with its own endpoints, parameters and auth. MCP is purpose-built for LLM applications. A host runs clients that open JSON-RPC sessions with servers, and each server publishes a machine-readable catalog of what it offers, so an agent can learn new capabilities without a code change. Because every MCP server speaks the same protocol, an agent needs one adapter instead of one per service. Underneath, many servers simply translate tool calls into the service's REST API, so MCP is an AI-friendly layer on top of APIs, not a replacement.

Key ideas, with timestamps

  1. 1:44

    Host, clients and servers

    An MCP host, such as a chat app or IDE, runs several MCP clients. Each client opens a JSON-RPC 2.0 session with one external server, for example a database, a code repository or an email server. In the USB-C analogy at 2:36, the laptop is the host, the USB-C link is the protocol and the monitor, drive and power supply are servers.

  2. 3:04

    Two needs: context and tools

    He frames MCP as solving two needs of AI agents. One is retrieving context such as documents, knowledge base entries and database records. The other is taking actions such as a web search, a call to an outside service or a calculation.

  3. 4:04

    Three primitives

    Tools are discrete functions, such as get weather or create event, advertised with a name, description and schema. Resources are read-only items such as files or a database schema that the client fetches on demand. Prompt templates are predefined suggested prompts. At 5:06 he notes many servers expose only tools.

  4. 5:28

    A catalog at runtime

    Every server answers tools/list, resources/list and prompts/list, so an agent can discover new functionality without redeploying code. This catalog is what the rest of the comparison hangs on.

  5. 6:42

    REST in one minute

    APIs are abstraction layers: the client only needs to know how to format a request and read the response. REST, the web's default style, uses HTTP verbs, as in GET /books/123 to fetch a book or POST /loans to borrow one, and usually returns JSON. He notes that commercial LLMs are themselves offered over REST.

  6. 9:19

    Purpose-built versus general-purpose

    MCP bakes in assumptions that suit agents, while APIs were not designed with LLMs in mind. Its main advantage, from 10:03, is dynamic discovery. Asked what it can do, a server returns all its functions and data, whereas a REST client usually needs a developer to update it when endpoints change.

  7. 10:55

    One interface for every server

    Every API has its own endpoints, parameter formats and authentication, so an agent using five REST APIs may need five adapters. Five MCP servers answer the same calls. He sums it up as build once, integrate many.

  8. 11:42

    MCP servers wrap APIs

    The turn of the video: many MCP servers do their work through traditional APIs. The GitHub MCP server exposes high-level tools for repositories and translates each call into GitHub's REST API. So the two are layers in one stack, not competitors.

What it gets right

The video avoids the most common mistake in MCP coverage, which is presenting it as a replacement for APIs. The point at 11:42 that servers usually wrap an existing REST API is the single most useful fact for anyone deciding whether to build one. If your service already has a good API, an MCP server is a thin adapter for agents, not a rewrite.

The host, client and server breakdown at 1:44 is also more precise than most explainers. Naming JSON-RPC 2.0 and the list methods tells developers what is actually on the wire.

The adapter argument at 10:55 is the real benefit in practice. One client implementation reaching many servers is why IDEs and chat apps added MCP support so quickly.

What it leaves out

  • APIs can describe themselves too. The claim that REST APIs lack runtime discovery is only partly true. OpenAPI specs are machine-readable, and agents have long used them to generate tools. MCP's real gain is a single standard that every client already understands, plus descriptions written for a model to read.
  • Discovery has costs. Tools that appear automatically also change automatically. A server update can rename a tool or rewrite its description without anyone reviewing it. And every listed tool's description goes into the model's context, so connecting many servers crowds the window and makes tool choice worse.
  • State and auth. An MCP session is a stateful connection, unlike a stateless REST call, which matters for scaling and hosting. Auth is not removed by MCP; it is moved into the server, which still has to hold the API key or OAuth token.
  • Detail that has moved. The video says servers advertise input and output schemas. Output schemas for tools were added to the spec in 2025, around the time of the May 2025 upload, so older servers may not have them. Check the current spec version.
  • When to skip MCP. If one agent calls one internal API, direct function calling is simpler. MCP pays off when several clients need the same capability.

How this applies to memory across AI tools

The two needs at 3:04, context and tools, map neatly onto memory. A knowledge base you want every agent to use is context, and the cleanest way to share it is one MCP server rather than a separate integration per app. That is the idea behind an AI memory layer: store once, and let any client query it.

The layering lesson holds here too. BrainTube 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. The MCP part is the adapter. The value is in what sits beneath it, which is what the video means when it says MCP and APIs are layers. For the basics, start with what is MCP.

Who should watch it

Watch it if you are a developer, architect or technical manager who knows REST and wants a clear picture of where MCP fits, in about 13 minutes. Engineers who know REST well can skip 5:46 to 8:12, the REST primer, and focus on 1:44 for the architecture and 9:12 to the end for the differences and the wrapper point. Skip it if you want to build a server; there is no code, and you will need a hands-on tutorial afterwards.

Frequently asked

Is MCP a replacement for REST APIs?
No. Most MCP servers call an existing API behind the scenes and translate the result into MCP's standard format. MCP gives AI agents one consistent way to discover and call capabilities, while the API still does the actual work. They are layers in the same stack.
What is dynamic discovery in MCP?
An MCP client can ask a server to list its tools, resources and prompts at runtime, getting back names, descriptions and input schemas. The agent can then use new capabilities without a code change. REST APIs can publish OpenAPI specs, but there is no single standard every AI client reads in the same way.
When should I build an MCP server instead of calling an API directly?
Build one when several AI clients, such as a chat app, an IDE and your own agent, need the same capability, because one server serves them all. If a single agent calls a single API, plain function calling is usually simpler. Either way, the MCP server can wrap your existing API.

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.