All video notes
Video notes · MCP

The Model Context Protocol (MCP)

Video by Anthropic on YouTube · 20 min · uploaded June 2025

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

In short

In this 20-minute conversation, Anthropic's MCP team explains the protocol as a standard way to hand context to an LLM app: tools, resources and prompts. They cover its origin in copy-paste frustration, why it was open-sourced, the move to remote servers, and a roadmap of a registry, long-running tasks and elicitation. It is a good primary source on intent, but light on how the protocol works and on its risks.

The argument

The guests argue that models do not talk to APIs directly; they consume prompts, tools and whatever context they are given, so the useful thing to standardise is how outside data reaches the model. MCP does that once, on both sides: a server built once works in every client that speaks the protocol, and a client that supports MCP gets every server for free. Keeping it open was deliberate, so integration builders would not have to bet on which AI app survives, and so competition would move to model quality and workflows. As models handle longer tasks, they expect less-used parts of the spec, such as sampling and stateful sessions, to matter more.

Key ideas, with timestamps

  1. 1:02

    Models do not call APIs

    Asked how MCP differs from calling an API, the answer is that the model only ever sees prompts, tools and ingested text. MCP standardises the step from a data source, whether an API or an internal system, to what the model receives. This is the cleanest one-line case for the protocol in the video.

  2. 2:03

    Tools, resources and prompts

    A server exposes three things. Tools are actions the model can take. Resources are raw data, such as files, that an app could load or feed into a RAG pipeline. Prompts are templates the user triggers, usually shown as slash commands, and can edit before sending.

  3. 3:07

    Born from copy and paste

    David, one of the co-creators, says it started with his frustration copying things between Claude Desktop and his IDE. He took the idea to Justin, the other co-creator, and they built it into Claude Desktop. At an internal hack week around September 2024, almost every project became an MCP server, from Slack integrations to a 3D printer.

  4. 6:16

    A slow launch

    MCP launched around Thanksgiving, November 2024, and the reception was slow, with people asking what it was and confusing it with "MPC". The turning point was client adoption: IDEs first, then model providers, which gave server builders a reason to build.

  5. 8:20

    Why it is open

    With a closed integration ecosystem, builders cannot tell which AI app will last, so they hesitate to invest. An open standard removes that risk. The team's view is that an AI app's value should come from model intelligence and workflow, not from owning integrations.

  6. 9:52

    From local to remote

    They cite more than 10,000 server builders and describe a shift from local servers running on a developer's machine to hosted remote MCP servers. Claude.ai integrations, which let you connect a website's MCP server to everyday chat, are named as the first big example.

  7. 17:40

    How many servers is too many

    Asked how many servers Claude can handle, the answer is that it depends on overlap. Three issue-tracker servers side by side will confuse the model; unrelated servers can coexist in large numbers. Tool wording and overlap matter more than the count.

  8. 18:13

    What was next in mid-2025

    The roadmap had three items for agents: a registry API so models can search for and add servers on demand, support for long-running tasks, and elicitation, which lets a server ask the user for more information mid-task. They also mention more examples, better docs and security primitives.

What it gets right

The framing at 1:02 is the most useful thing here. Comparing MCP with simply calling an API misses that the model never calls anything itself. Something has to describe the capability to the model and carry the result back. MCP standardises that layer, and once you see it that way the many-to-many argument (build one server, reach every client) follows.

The point about overlapping servers is also practical advice. In practice the same rule holds: tool choice degrades when two servers offer similar tools with vague descriptions, not when you simply add more servers. Name and describe tools so they cannot be confused.

The explanation of prompts as user-triggered templates is clearer than in most third-party explainers, which often describe all three capabilities as if the model picks them.

What it leaves out

  • How it works. There is no mention of the client and server handshake, JSON-RPC, or transports (stdio for local servers, HTTP for remote ones). For a protocol talk, it is almost entirely about why, not how.
  • Security. "Key security primitives" gets one sentence. Remote servers mean auth, and tool descriptions are injected into the model's context, so an untrusted server can steer the model. Anyone connecting third-party servers needs to treat them as dependencies with permissions.
  • Control over each capability. The talk explains that prompts are user-triggered but does not spell out the full design: tools are model-controlled, resources are application-controlled. That split explains why many clients support tools well and resources less.
  • Hard numbers. The 10,000-plus figure is for server builders and is not sourced. Treat it as a rough signal of adoption in June 2025.

Much of the roadmap has moved since the video's June 2025 upload. As far as we know, elicitation entered the spec shortly after, a public MCP registry followed later in 2025, and the protocol's governance moved to a foundation outside Anthropic. Check the current spec before relying on any detail here.

How this applies to memory across AI tools

The open-standard argument at 8:20 applies to personal knowledge as much as to integrations. If your notes and saved sources sit behind one AI app, you are betting on that app. If they sit behind an MCP server, any client that speaks the protocol can reach them, which is the core idea of 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. The overlap lesson from 17:40 carries over too: one well-described memory server beats three that all claim to search your notes. For a starting point, see what MCP is.

Who should watch it

Watch it if you want MCP's intent from the people who designed it: why it exists, why it is open, and what they planned next. It suits product people and developers new to MCP. Jump to 1:02 for the difference from an API, 2:03 for the three capabilities, 8:20 for the open-source reasoning and 18:13 for the roadmap. The getting-started tips at 13:29 (build a one-tool hello world server first, then try a prompt and a resource) are short but sound. Skip it if you need implementation detail, security guidance or the current state of the spec, since it is a mid-2025 conversation with no code.

Frequently asked

Who created the Model Context Protocol?
MCP was created at Anthropic. In this video David, one of the co-creators, says it began from his frustration copying context between Claude Desktop and his IDE, and that he built it out with Justin, the other co-creator. It launched publicly in November 2024.
Why did Anthropic make MCP open source?
The team says a closed integration ecosystem makes builders unsure which AI app to invest in. An open standard lowers that risk, so servers get built once and work everywhere. They want competition between AI apps to be about model quality and workflows, not about who owns integrations.
What are MCP prompts used for?
MCP prompts are reusable templates a server provides that the user chooses to insert, usually through a slash command in the client. The user can edit the text before sending it. They differ from tools, which the model calls, and resources, which provide data.

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.