Model Context Protocol
The Model Context Protocol (MCP) is an open standard, introduced by Anthropic on Nov 25, 2024, for connecting AI applications such as Claude to outside tools and data through MCP servers. OpenAI, Google and Microsoft support it, and Anthropic donated it to the Linux Foundation's Agentic AI Foundation on Dec 9, 2025.
The Model Context Protocol (MCP) is an open standard for connecting AI applications to the systems where data and tools live: a service publishes an MCP server once, and any AI app that speaks MCP (an MCP host, such as Claude, ChatGPT or VS Code) can use its tools and data without a custom integration. Anthropic introduced it on Nov 25, 2024, calling it a new standard for linking AI assistants to the systems where data lives. It was created at Anthropic by David Soria Parra and Justin Spahr-Summers. On Dec 9, 2025, Anthropic donated MCP to the Agentic AI Foundation, a fund under the Linux Foundation, so no single company owns it.
MCP's own documentation likens it to a USB-C port for AI applications: one plug shape instead of a different cable for every device (OpenAI Agents SDK docs, quoting the MCP docs).
Quick facts
| Introduced | Nov 25, 2024, by Anthropic |
| Steward | Agentic AI Foundation (Linux Foundation) since Dec 9, 2025 |
| Current spec | 2026-07-28 (as of Nov 6, 2026) |
| License | Apache 2.0 for code and specs, CC BY 4.0 for docs |
| Transports | stdio (local) and Streamable HTTP (remote) |
| Public servers | "More than 10,000 active", per Anthropic, Dec 9, 2025 |
| Home | modelcontextprotocol.io |
What does MCP standardize?
MCP fixes how three parts talk to each other (Architecture overview):
- Host: the AI application, such as Claude Desktop, Claude Code or VS Code.
- Client: a connection the host opens for each server.
- Server: a program offering context and actions. A local server (for example a filesystem server) runs on your machine over stdio; a remote one (for example Sentry's) runs as a web service over Streamable HTTP.
Messages are JSON-RPC 2.0. A server can offer three kinds of things, called primitives:
- Tools: functions the model can call, such as running a database query or creating a ticket.
- Resources: data the app can read, such as file contents or a database schema.
- Prompts: reusable templates, such as a prepared instruction with examples.
Clients can in turn let a server ask the user for input (elicitation). Apps find tools with tools/list and run them with tools/call.
How has the spec changed?
Versions are dates marking the last breaking change: 2024-11-05, 2025-03-26, 2025-06-18, 2025-11-25 and the current 2026-07-28.
The 2026-07-28 version is the biggest change so far: MCP became stateless. It dropped the initialize handshake and the Mcp-Session-Id header, so every request carries its own version and capabilities; moved tasks into an official extension; and deprecated three client features, Roots, Sampling and Logging (Key Changes 2026-07-28). Deprecated features stay at least twelve months before removal, and older tutorials describe the earlier, handshake-based protocol.
Who supports MCP?
At launch, Anthropic named Block and Apollo as adopters and shipped servers for Google Drive, Slack, GitHub, Git, Postgres and Puppeteer. By MCP's first anniversary on Nov 25, 2025, the large AI companies had joined:
- OpenAI said MCP is now "a key part of how we build at OpenAI", across ChatGPT and its developer platform; its Agents SDK supports local and remote MCP servers.
- Google said it supports MCP "across Gemini", including Gemini CLI, and published servers for Google Maps and Google Cloud databases.
- Microsoft said Foundry grew from a small set of tools to thousands because of MCP; GitHub runs a GitHub MCP Server.
- AWS said it started and keeps contributing to MCP.
Anthropic's Dec 9, 2025 count: MCP adopted by ChatGPT, Cursor, Gemini, Microsoft Copilot and Visual Studio Code, with 97 million monthly SDK downloads across Python and TypeScript. These are the companies' own statements; exact first-adoption dates were not confirmed at a primary source.
The Agentic AI Foundation
The Agentic AI Foundation (AAIF) was co-founded by Anthropic, Block and OpenAI, with support from Google, Microsoft, AWS, Cloudflare and Bloomberg (Anthropic, Dec 9, 2025). MCP was a founding project beside Block's goose and OpenAI's AGENTS.md, and Anthropic said MCP's governance would stay unchanged. Its maintainers decide changes through Specification Enhancement Proposals, with no seats reserved for any company (Governance). Anthropic later gave the AAIF its Agent Skills standard too, on Oct 22, 2026.
MCP Apps and the MCP Registry
MCP Apps is an official extension that lets a server send an interactive interface (a chart, a form, a dashboard) that the host shows inside the chat in a sandboxed frame (MCP Apps). Claude, Claude Desktop, VS Code GitHub Copilot, Microsoft 365 Copilot and Goose support it. In the Claude apps it powers the interactive Connectors launched on Jan 26, 2026 with Asana, Figma, Slack and others.
The MCP Registry is the official public catalog of MCP servers, backed by Anthropic, GitHub, PulseMCP and Microsoft. It is still in preview, and its page warns of possible breaking changes or data resets before general availability (The MCP Registry).
How Claude uses MCP
- Claude apps: custom connectors connect to remote MCP servers from Anthropic's cloud, and Free users get one custom connector (About connectors). Claude Desktop also runs local servers packaged as desktop extensions. See Connectors and How to Connect Claude to MCP Servers.
- Claude Code:
claude mcp add --transport http <name> <url>adds a remote server, and servers can be shared with a team in.mcp.json(Connect Claude Code to tools via MCP); see Claude Code MCP Servers. - Claude API: the MCP connector (beta) lets a Messages API request use remote MCP servers directly, without your own MCP client; local stdio servers cannot connect this way (MCP connector). Managed Agents also take MCP servers.
Is MCP safe?
MCP moves risk rather than removing it: an MCP server can read data and take actions with whatever access you give it. The specification's own security guide lists the main attacks (Security Best Practices):
- Local server compromise: a one-click install can run any command with your privileges, so clients must show the exact command and ask first.
- Confused deputy and token passthrough: proxy servers that forward or accept the wrong OAuth tokens can give an attacker a user's access; servers must only accept tokens issued to them.
- Server-side request forgery: a malicious server can point a client at internal addresses such as cloud metadata endpoints.
- Over-broad scopes: asking for every permission at once makes a stolen token far more damaging.
Two further cautions come from Anthropic's tool docs: a tool's results can contain instructions that Claude may follow (prompt injection), and selection accuracy drops once a model has more than about 30 to 50 tools loaded, which is why many-server setups use tool search. Install servers only from publishers you trust and give each one the least access it needs.
Frequently asked questions
What is MCP in Claude? MCP (Model Context Protocol) is the open standard Claude uses to connect to outside tools and data: Claude's connectors, Claude Code's MCP servers and the Claude API's MCP connector all speak it.
Who created the Model Context Protocol? Anthropic created the Model Context Protocol and released it on Nov 25, 2024; David Soria Parra and Justin Spahr-Summers built it, and since Dec 9, 2025 it belongs to the Linux Foundation's Agentic AI Foundation.
Does ChatGPT support MCP like Claude does? Yes. OpenAI said in Nov 2025 that MCP is integrated across ChatGPT and its developer platform, so the same MCP server can work with both Claude and ChatGPT. Learn more: ChatGPT on ai.vi.ki.
What is the difference between an MCP server and an MCP client? An MCP server offers tools, data or prompts (for example a GitHub or database server), while an MCP client is the connection an AI app such as Claude opens to one server.
Is MCP free to use? Yes. The Model Context Protocol specification and SDKs are open source under Apache 2.0, though the services an MCP server connects to may charge for their own use.
Sources
- Why we built-and donated-the Model Context Protocol (MCP), Anthropic, Dec 11, 2025 (video)
- Introducing the Model Context Protocol, Anthropic, Nov 25, 2024
- Donating the Model Context Protocol and establishing the Agentic AI Foundation, Anthropic, Dec 9, 2025
- Donating Agent Skills to the Agentic AI Foundation, Anthropic, Oct 22, 2026
- One Year of MCP, MCP Blog, Nov 25, 2025
- Architecture overview, modelcontextprotocol.io, read Nov 6, 2026
- Versioning, modelcontextprotocol.io, read Nov 6, 2026
- Key Changes 2026-07-28, modelcontextprotocol.io, read Nov 6, 2026
- Governance and Stewardship, modelcontextprotocol.io, read Nov 6, 2026
- MCP Apps, modelcontextprotocol.io, read Nov 6, 2026
- The MCP Registry, modelcontextprotocol.io, read Nov 6, 2026
- Security Best Practices, modelcontextprotocol.io, read Nov 6, 2026
- Model context protocol (MCP), OpenAI, read Nov 6, 2026
- Your favorite work tools are now interactive connectors inside Claude, Claude, Jan 26, 2026
- About connectors, Claude Help Center, read Nov 6, 2026
- Connect Claude Code to tools via MCP, Anthropic, read Nov 6, 2026
- MCP connector, Anthropic docs, read Nov 6, 2026
- Tool search tool, Anthropic docs, read Nov 6, 2026
Moments
Why we built-and donated-the Model Context Protocol (MCP)
Anthropic
- The Model Context Protocol (MCP) is an open standard for connecting AI applications to the systems where data and…
- MCP's own documentation likens it to a USB-C port for AI applications: one plug shape instead of a different cable for…
- At launch, Anthropic named Block and Apollo as adopters and shipped servers for Google Drive, Slack, GitHub, Git…
- The Agentic AI Foundation (AAIF) was co-founded by Anthropic, Block and OpenAI, with support from Google, Microsoft…
- MCP Apps is an official extension that lets a server send an interactive interface (a chart, a form, a dashboard) that…
- MCP moves risk rather than removing it: an MCP server can read data and take actions with whatever access you give it…
See also
-
Claude CodeClaude Code is Anthropic's agentic coding tool. It reads a codebase, edits files, runs commands and opens pull requests from the terminal, VS Code, JetBrains IDEs, the Claude desktop app, the browser and Slack, and it comes with every paid Claude plan or a pay-as-you-go Console account. -
SkillsA Claude skill is a folder with a SKILL.md file of instructions, plus optional scripts and files, that Claude loads when a task needs it. Skills are on every Claude plan, Free included, and the same skill runs in the Claude app, Claude Code and the API. -
ConnectorsClaude connectors link Claude to apps such as Gmail, Google Drive, Slack, Notion and Asana so it can read and act in them with your permissions. They are on every Claude plan, Free included; Free accounts may add one custom connector. -
How to Connect Claude to MCP ServersConnect Claude to a remote MCP server in Customize > Connectors > Add custom connector, to a local one in Claude Desktop through a desktop extension or claude_desktop_config.json, and in Claude Code with claude mcp add.
-
Claude Code MCP ServersClaude Code connects to MCP servers with claude mcp add, for example claude mcp add --transport http notion https://mcp.notion.com/mcp, stored per user, per project in a shared .mcp.json file, or locally for one project.
-
Messages APIThe Messages API is the Claude API's main endpoint, POST /v1/messages: you send a model, max_tokens and a list of user and assistant messages, and Claude returns a message with content blocks, a stop_reason and token usage, whole or streamed.