What is MCP (Model Context Protocol)?

In plain English

An AI assistant on its own can only chat. To be really useful it needs to reach your files, calendar, company systems and other apps. Before MCP, every AI app had to build a separate connection to every tool, and every tool had to support every AI app separately.

MCP, the Model Context Protocol, is a shared standard for those connections. A tool builds one “MCP server” that describes what it can do (search these documents, create this ticket, read this calendar), and any AI app that speaks MCP can use it. The project’s own documentation compares it to a USB-C port for AI: one standard plug instead of a drawer full of different cables.

Why it matters

MCP is a big part of how AI agents get things done in the real world. Anthropic, the company behind Claude, introduced it as an open standard in November 2024. Within a year it was supported by ChatGPT, Gemini, Microsoft Copilot and coding tools such as Cursor, and in December 2025 Anthropic donated it to the Agentic AI Foundation, part of the non-profit Linux Foundation, so no single company controls it.

An example

You ask your AI assistant, “Find the latest sales figures and put them in a chart for tomorrow’s meeting.” Through MCP it connects to your company’s spreadsheet tool to fetch the numbers, to a charting tool to draw the chart, and to your slide software to add it to the deck, all using the same standard instead of three custom integrations.

What to watch out for

Connecting AI to tools also connects it to risk. A badly built or malicious MCP server can leak data or give the AI harmful instructions, and documents the AI reads through MCP can contain hidden prompt injection attacks. Only install MCP servers from sources you trust, give them the smallest access they need, and keep a human approving anything that can’t be undone, such as sending, paying or deleting.

See how agents use tools in AI agents, explained simply.