All video notes
Video notes · MCP

Model Context Protocol (MCP), clearly explained (why it matters)

Video by Greg Isenberg on YouTube · 20 min · uploaded March 2025

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

In short

In this episode Ross Mike explains MCP as the third stage after bare LLMs and LLMs with hand-wired tools: a shared standard that sits between a model and outside services. The framing is clear and good for non-technical viewers, and his point that service providers now write the servers is the key insight. The translator analogy oversells what MCP does, and several of his open questions, such as whether OpenAI would adopt it, were answered within weeks of the March 2025 upload.

The argument

Ross Mike's argument is that language models on their own can only predict text, so anything useful (sending email, querying a database, searching the web) depends on tools. Wiring each tool to a model by hand works, but stacking many tools into one coherent assistant is brittle, and every API change can break the chain. MCP, in his telling, is a standard layer that lets any client talk to any service through one consistent protocol, and its clever move is that the service provider, not the AI company, builds the server. He closes by telling non-technical founders to watch rather than bet, because the standard may still change or lose.

Key ideas, with timestamps

  1. 1:31

    Engineering runs on standards

    He opens with REST APIs as the familiar example of a standard that lets one company's software talk to another's. The setup matters because his whole case is that MCP is a standard, not a new capability. It is a useful way to lower the hype before explaining anything.

  2. 2:31

    A bare LLM only predicts the next word

    Ask an early ChatGPT to send an email and it says it cannot. At 3:30 he uses the classic example: given the start of a film title, the model predicts the last word because that is the likeliest continuation. The point is that action has to come from outside the model.

  3. 4:00

    Stage two: models plus tools

    Perplexity searching the web is his example of a model given a tool, with Brave search as a service behind such features. At 4:30 he adds automation platforms like Zapier and n8n, for instance logging every new email as a spreadsheet row. Each connection makes the model more useful, but each one is custom glue.

  4. 5:30

    Why there is still no Jarvis

    Combining one tool with a model is manageable. Stacking search, email and summaries into one dependable assistant is, in his words, a nightmare, because you also have to stop the model doing something wrong with each tool. At 10:00 he gives the failure case: Slack or a texting service changes its API and a multi-step chain breaks.

  5. 7:31

    The languages analogy

    Every service speaks its own language: tool one English, tool two Spanish, tool three Japanese. MCP is pictured as a layer that translates all of them into one language the model understands. At 9:01 he gives a concrete case: telling the model to create a database entry in a backend like Supabase or Convex, and it knowing how.

  6. 11:02

    Client, protocol, server, service

    He draws the ecosystem as four boxes. Clients are model-facing apps; he names Tempo (where he works), Windsurf and Cursor. The protocol is the two-way connection, and the server exposes an external service's capabilities to the client. This is the most reusable diagram in the episode.

  7. 11:32

    Providers build the servers

    He calls Anthropic's design "3D chess": the server is now the service provider's job. A database company that wants AI clients to use its product has to ship an MCP server, which explains why so many companies started publishing them. The work of integration shifts from every app builder to each service, once.

  8. 13:31

    Setup pain and an app store idea

    He admits that installing servers in early 2025 meant downloading, moving and copying local files. At 16:31 he gives away an idea, an MCP app store where you browse servers, click deploy and get a URL to paste into a client. For non-technical founders, his advice from 17:31 is to watch which standard wins before building on it.

What it gets right

The three-stage story is a good on-ramp. Bare model, model plus hand-wired tools, model plus a standard protocol is an accurate history, and it explains why MCP got attention without claiming it makes models smarter. His line that it is just a standard is the right corrective to the hype.

Where the work moves is the real insight. Before MCP, every AI app that wanted Slack, GitHub or a database had to write its own integration. With MCP, the service ships one server and every compatible client can use it. That is why adoption snowballed: the incentive sits with the companies that want to be reachable from AI tools.

The setup complaint was accurate. In March 2025 most servers ran locally and were configured by editing JSON files. Saying so plainly was more honest than most launch-week coverage.

What it gets wrong or leaves out

MCP does not translate APIs. The languages analogy implies a layer that automatically understands every service. In practice someone still writes an MCP server for each service, calling that service's own API. If Slack changes its API, the Slack server must be updated, exactly the breakage he describes at 10:00. What MCP standardises is the client side: how a model discovers tools, calls them and gets results, so the fix happens in one server instead of in every app.

Tools are only part of the protocol. The episode treats MCP as tools only. The specification also defines resources (data a client can read) and prompts (reusable templates), which matter for things like exposing a knowledge base rather than an action.

Several open questions were settled soon after. He wonders whether OpenAI might ship a rival standard. Instead, OpenAI announced MCP support in its Agents SDK later in March 2025 and later added it to ChatGPT. Remote servers over HTTP, which remove most of the local-file setup he complains about, became common, and directories such as Smithery and an official MCP registry now cover much of his app store idea. If you watch this today, treat the last five minutes as a time capsule.

Security is absent. Giving a model access to your email or database through a third-party server raises permission and prompt-injection questions. A clear explainer for non-technical viewers should at least say to check what a server can do before installing it.

How this applies to memory across AI tools

The client-server split is what makes shared memory practical. If your knowledge lives behind one MCP server, every client that speaks the protocol can read it, so you are not rebuilding context separately inside each chat app. That is the idea behind an MCP memory server.

BrainTube is one example of the pattern: 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 the same material. For the basics of the protocol itself, start with what MCP is.

Who should watch it

Watch it if you are non-technical and want a calm, jargon-light picture of why MCP exists; the core explanation runs from 1:31 to 15:30, and the client-server diagram at 11:02 is the part to remember. Skip the first minute of intro and the membership ad at 6:31 to 7:31. Developers will learn nothing new, and anyone deciding what to build today should know the business advice from 15:31 onward reflects March 2025, before OpenAI adopted MCP and before remote servers and registries were common.

Frequently asked

What is MCP in simple terms?
MCP, the Model Context Protocol, is an open standard for connecting AI apps to outside tools and data. An app that supports MCP can use any MCP server, such as one for a database, a code host or a notes store, without custom integration code for each. Anthropic introduced it in November 2024.
Who builds MCP servers?
Usually the company or developer behind the service being connected, though anyone can write one. A service that wants to be reachable from AI assistants publishes a server that describes its tools and data. Any MCP client can then connect to it.
Does MCP replace APIs?
No. An MCP server normally calls the service's existing API behind the scenes. MCP standardises how AI clients discover and call those capabilities, so each service needs one server instead of a separate integration for every AI app.

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.