Model Context Protocol (MCP)

What it means

A protocol is a set of rules that two systems agree to use when they talk to each other. The Model Context Protocol (MCP) is one of these for AI. It gives an AI assistant or [agent](/definitions/ai-agent) a shared way to connect to the files, databases, and software an organization already uses. The project's official site calls it "an open-source standard for connecting AI applications to external systems" and offers a picture from everyday life: "Think of MCP like a USB-C port for AI applications."¹

The site also gives an everyday example: "Agents can access your Google Calendar and Notion, acting as a more personalized AI assistant."¹ Without a shared standard, every AI product needs its own custom link to every calendar. With MCP, the calendar side publishes one connector, which MCP calls a "server." Any AI product that speaks MCP can then use that connector.

Where the word comes from

Anthropic, the AI company, introduced MCP on November 25, 2024. It described the problem MCP was meant to solve this way: "Every new data source requires its own custom implementation, making truly connected systems difficult to scale."² It also wrote that "Early adopters like Block and Apollo have integrated MCP into their systems."² The design borrows from an older idea. The specification says "MCP takes some inspiration from the Language Server Protocol," which had standardized how programming tools work together.³

Since then, the meaning has widened from reading to doing. In 2024, Anthropic described a standard "for connecting AI systems with data sources."² Today's official description adds tools and workflows, so that AI applications can "access key information and perform tasks."¹

Who controls MCP has also changed. On December 9, 2025, Anthropic gave MCP to the Agentic AI Foundation, which it described as "a directed fund under the Linux Foundation co-founded by Anthropic, Block and OpenAI."⁴ At that time, Anthropic reported that "MCP has been adopted by ChatGPT, Cursor, Gemini, Microsoft Copilot, Visual Studio Code, and other popular AI products" and that "There are now more than 10,000 active public MCP servers."⁴ These counts come from Anthropic itself. We have not found an independent count.

How others define it

• Anthropic (2024): "The Model Context Protocol is an open standard that enables developers to build secure, two-way connections" between their data and AI tools.²
• The official MCP site (2026): "MCP (Model Context Protocol) is an open-source standard for connecting AI applications to external systems."¹
• The MCP specification (June 2025 version): "an open protocol that enables seamless integration between LLM applications and external data sources and tools."³ An LLM is a [large language model](/definitions/large-language-model), the kind of AI behind ChatGPT and Claude.
• IBM: MCP "serves as a standardization layer for AI applications."⁵

Where they agree and differ

All four sources agree on the core idea: MCP is one common plug that replaces many one-off connections. They differ on two things. The first is scope. The 2024 announcement talks mostly about data, while later definitions add tools and actions. The second is audience. Anthropic's first definition speaks to developers, while the official site's examples speak to everyday users.

The strongest opposing view is that the USB-C comparison makes MCP sound safer than it is. MCP's own specification supports that concern. It warns that "The Model Context Protocol enables powerful capabilities through arbitrary data access and code execution paths."³ It also admits that "MCP itself cannot enforce these security principles at the protocol level."³ A charging cable has no say in what a device does once it is connected, and MCP has no say in what an agent does. We think this view is correct: MCP is a plug, not a policy. The standard still has value, but the safety rules have to come from the people who use it.

In the news

On October 27, 2026, Apple announced that Keynote, Pages, and Numbers will get MCP support. Apple says the support "enables external AI agents and tools to read content, perform edits, and execute automated actions directly in Keynote, Pages, and Numbers."⁶ Apple says the feature "requires macOS 27 or later."⁶ It describes the new AI features in these apps as a beta and says the updates will reach subscribers in November.⁶ We are confident about what Apple announced, because the source is Apple's own press release. We cannot yet say how the feature works in practice, because it has not shipped. The announcement does show that agents can now act inside everyday office documents, not just read them.

What it means for your organization

MCP makes it cheaper and faster to connect agents to your systems. That changes the question leaders need to answer. It is no longer "Can we connect this agent?" It is "Should this agent be allowed to act here, and who decides?" The specification sets that expectation for the people who build MCP products: "Hosts must obtain explicit user consent before invoking any tool," and "Users must retain control over what data is shared and what actions are taken."³ Those rules fit the principle that [a person stays in the middle](/definitions/a-person-in-the-middle) of every chain of command. However, the specification also says the protocol cannot enforce them.³ Your organization has to make sure each connection is approved and that someone keeps watching what the agent does with it.

One question to ask your team: Which of our systems can an AI agent reach through MCP today, who approved each connection, and who reviews what the agent does with it?