Blog · November 13, 2026

How MCP connects AI apps to tools

MCP is an open standard for connecting an AI app to tools and data. This note names the parts, who started it, and who keeps it now.

What it is

On November 25, 2024, Anthropic open sourced the Model Context Protocol, called MCP. It is an open standard for connecting an AI assistant to the systems where data lives. Those systems include content repositories, business tools, and development environments.

The MCP documentation describes the same idea in current terms. An AI application such as Claude or ChatGPT can connect to data sources, tools, and workflows.

The problem it solves

Anthropic said each new data source had needed its own custom connection. That made a fully connected system hard to grow. MCP replaces those separate connections with one protocol.

In that announcement, a developer can publish data through an MCP server. A developer can also build an AI application, called an MCP client there, that connects to those servers.

The launch included the specification, software kits, local MCP support in the Claude desktop apps, and a public set of servers. The first servers covered Google Drive, Slack, GitHub, Git, Postgres, and Puppeteer. Anthropic named Block and Apollo as early adopters that had integrated MCP.

Hosts, clients, and servers

The architecture page for specification revision July 28, 2026 names three participants.

The host is the AI application. It coordinates one or more clients. The page gives Claude Code, Claude Desktop, and Visual Studio Code as examples.

The client keeps one connection to one server and brings context back to the host. The host creates a new client for each server it connects to.

The server is a program that provides context to clients. It can run on the same machine or on another machine. The page uses a filesystem server as a local example and the Sentry server as a remote example.

Tools, resources, and prompts

The server concepts page says a server offers three building blocks.

Tools are functions the model can call. The model decides when to use them. The page gives search, sending a message, and creating a calendar event as examples. A tool can ask for a person's approval before it runs.

Resources are passive data sources. They give read only access to information such as a file, a database schema, or API documentation. The application decides how to use them.

Prompts are instruction templates. They tell the model how to work with specific tools and resources. The page gives planning a vacation, summarizing meetings, and drafting an email as examples. A person chooses a prompt. The model does not start one on its own.

Who has adopted it

On May 20, 2025, Google said the Gemini API and its software kit would support MCP, so developers could use a wide range of open source tools.

On June 25, 2025, Google said Gemini CLI, then in preview, could be extended through built in support for MCP.

Visual Studio Code's documentation says MCP connects AI to tools and data sources. A person can add an MCP server when Copilot in VS Code needs a tool or a data source that is not already available. That page does not show a publication date.

OpenAI's developer guide says a remote MCP server can connect models to new data sources and capabilities. The guide shows how to use that server as a plugin in ChatGPT and Codex, and through the API. That page does not show a publication date.

On December 9, 2025, Anthropic said ChatGPT, Cursor, Gemini, Microsoft Copilot, and Visual Studio Code had adopted MCP. The Linux Foundation said the same day that Claude, Cursor, Microsoft Copilot, Gemini, VS Code, and ChatGPT had adopted it. The MCP documentation site also names Claude, ChatGPT, Visual Studio Code, and Cursor among the products that support MCP.

Who keeps it now

On December 9, 2025, Anthropic gave MCP to the Agentic AI Foundation, a directed fund under the Linux Foundation. Anthropic, Block, and OpenAI founded it together. The Linux Foundation lists the platinum members as Amazon Web Services, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft, and OpenAI.

MCP sits there with goose from Block and AGENTS.md from OpenAI. Anthropic said the governance model stays the same. Maintainers still prioritize community input, and decisions stay open to view.

In that same post, Anthropic said there were more than 10,000 active public MCP servers. Claude's directory had more than 75 connectors. The Python and TypeScript kits had more than 97 million downloads a month.

What the July 28, 2026 specification changed

The specification revision dated July 28, 2026 lists its changes since the November 25, 2025 revision.

Each request now stands on its own. It carries its protocol version and what the client can do. A server does not keep that from an earlier request.

Every server must offer a discovery call, so a client can learn the server's versions, what it can do, and its identity before it sends other requests.

Roots, sampling, and logging are deprecated. They remain in the specification during a deprecation window of at least twelve months. New implementations should not add them.

Sources

Opened and checked on November 13, 2026. Each date below is the date on the source, unless the line says the page does not show one.