How to Use the Model Context Protocol Today

The Model Context Protocol (MCP) is an open standard, announced by Anthropic November 2024, that lets any AI application talk to any tool through one shared interface. It replaces the old pattern where ten apps connecting to ten tools required a hundred hand-built integrations; with MCP, you build each connector once and every compliant client can use it.

What the Model Context Protocol Solves

The Model Context Protocol (MCP) is an open standard, announced by Anthropic in November 2024, that gives AI applications one shared way to connect to external tools and data. Before it, every AI app needed a custom connection to every tool: ten apps and ten tools meant a hundred integrations to build and maintain. With MCP, each connector is built once and any compliant client can use it, so the count becomes twenty.

Engineers call the old pattern the N-by-M problem. Each application speaks its own API dialect, and each tool demands its own authentication and rules, so nothing is shared between integrations. MCP replaces that with a single protocol layer, which the video compares to what USB-C did for hardware connectors: one universal adapter instead of a drawer of cables. The protocol is now maintained under the Linux Foundation, so no single company owns it.

Host, Client, Server: The Three-Part Architecture

MCP's architecture has three parts, and the relationship between them is fixed. The host is the AI application you use, such as Claude Desktop or an editor like Cursor. Inside the host, one or more clients each manage a connection to exactly one server. The server is a separate program, local or remote, that exposes tools and data.

All three communicate using JSON-RPC 2.0 messages. One host can run many clients in parallel, each connected to its own server, which is how a single assistant can reach a file system, a database, and GitHub at the same time. The official SDKs and reference servers live in the modelcontextprotocol GitHub organization.

How One Request Flows End to End

A single MCP request follows five steps: discover, request, approve, execute, respond. The video traces a plain question, "what's in my project's README file?", through the whole chain.

1. The client asks the server to list its available tools.
2. The server replies with, for example, a read-file tool and the inputs it needs.
3. The model decides to use that tool, and the host asks you for permission first.
4. You approve; the client calls the tool and the server reads the file.
5. The model turns the returned content into a plain answer.

The approval step matters more than it looks. A human stays in the loop before any action runs, which is the protocol's main safety mechanism in day-to-day use.

The Six MCP Primitives

An MCP session has a vocabulary of exactly six primitives: three the server offers and three the client offers. The most common mistake is confusing tools with resources, so keep the rule simple: a tool takes an action, a resource gives information.

Primitive	Offered by	What it does
Tools	Server	Actions the model can take: search, send, query, create a file
Resources	Server	Read-only context identified by a URI: a file, record, or document
Prompts	Server	Reusable templates that guide the model
Sampling	Client	Lets a server ask the host to run the model on its behalf
Roots	Client	Tells the server where it is allowed to operate; a safety boundary
Elicitation	Client	Lets a server pause mid-task and ask the user for more information

Server-side primitives describe what the model can do. Client-side primitives give servers controlled access back: the ability to request model inference, know its operating boundaries, and ask the human a clarifying question.

Transports: stdio and Streamable HTTP

Clients and servers connect over two standard transports. stdio is the default and is used when the server runs locally on your machine. Streamable HTTP is used when the server lives on a remote service. An older method, HTTP plus Server-Sent Events (SSE), is deprecated in the current specification, so new builds should not depend on it.

Before any work happens, the two sides run an initialization handshake. The client declares its capabilities, and the server lists its tools, resources, and prompts. That exchange is why any compliant client can talk to any compliant server without custom Claude Code.

MCP Security: OAuth 2.1, TLS, and Consent

MCP security rests on three layers. For remote servers, the protocol uses OAuth 2.1 authorization protected by TLS encryption. The specification's recent revisions have focused heavily on hardening authorization flows for remote connections.

The most important guardrail is human consent: a well-built host asks before running any tool. But consent at the host does not make a malicious server safe. Because servers can read files and take real actions, the guidance is to connect only servers you trust and to review what a server can reach before enabling it.

Using MCP Today and the 2026 State of the Protocol

Claude Desktop was the first MCP client and remains the easiest starting point. Support has since spread to Claude Code, Cursor, VS Code, and other applications. You do not need to build servers from scratch: ready-made servers exist for GitHub, databases, file systems, Slack, and many other services, so setup is usually install a server, point your client at it, and approve.

Adoption moved quickly. The video reports that by mid-2026 there were nearly sixteen thousand public MCP servers on GitHub and that the official SDKs had crossed a billion downloads; treat those as speaker-reported figures rather than independently verified counts. Adoption by Anthropic, OpenAI, Google, Microsoft, and GitHub is documented in the specification's history and vendor announcements. In July 2026 the specification received its biggest change since launch, making the protocol core stateless and adding new capabilities such as MCP Apps, which is where current development is concentrated.

FAQ

• What is MCP in one sentence? MCP, the Model Context Protocol, is an open standard announced by Anthropic November 2024 that lets any AI application use any tool server through one shared protocol, turning N-by-M integrations into N plus M.
• What is the difference between an MCP tool and a resource? A tool is an action the model can take, such as sending a message or creating a file. A resource is read-only context, like a file or database record, identified by a URI.
• **Is MCP owned by Anthropic. Anthropic announced it in November 2024, but the protocol is now maintained under the Linux Foundation, with adoption across OpenAI, Google, Microsoft, and GitHub.
• Which transport should an MCP server use? Use stdio for a server running locally on the same machine as the client, and Streamable HTTP for remote servers. The older HTTP plus SSE transport is deprecated and should not be used for new work.
• Is MCP safe to use? The protocol builds in OAuth 2.1, TLS, and human consent before tool execution, but a malicious server remains a real risk. Only connect servers you trust and review what they can access.

Turn Your Own Explainer Into an Article

MCP works because one clear explanation of a shared standard saves every developer from rebuilding the same mental model. If your own knowledge lives in YouTube videos, whether explainers, interviews, or lessons, it deserves the same treatment.

That is exactly what Skala Blog does: paste a YouTube URL, get a transcription, and turn it into a structured written article, the way this piece was drafted from a twelve-minute video with the help of contributors like Gustavo dev doido. Visit skalablog.com to convert a video you already have into something searchable.

Source video