MCP (Model Context Protocol)
An open standard that lets AI assistants connect to external tools and data, with key terms and safety basics.
MCP (Model Context Protocol): an open standard that lets AI assistants connect to external tools and data, such as email, files or code repositories, in a consistent way. An app that offers this connection is called an MCP server.
In plain English
Before MCP, every AI app needed a custom connection to every service. MCP is like a common plug: if an app offers an MCP server and your assistant supports MCP, they can work together. Anthropic introduced it in 2024 and many assistants and editors now support it.
What it lets an assistant do
- Search and read your files, email or notes.
- Create drafts, tasks or documents.
- Work with code repositories and issue trackers.
It can only do what the connection and your permissions allow, and well-behaved assistants should ask before actions that change data.
Key terms
| Term | Meaning |
|---|---|
| MCP server | The app-side connector that exposes tools and data |
| Client | The AI app or editor that connects to servers, such as Claude or Cursor |
| Remote server | Hosted by the provider at a web address, usually with sign-in through OAuth |
| Local server | Runs on your own computer |
Safety
- Connect only official or trusted servers.
- Grant the least access that works, and prefer read-only.
- Treat content the assistant reads, such as emails and web pages, as untrusted.
- Never paste tokens or keys into chat.
Common confusion
- MCP is not an AI model. It is a way to connect models to tools.
- It is not only for one company’s assistant. It is an open standard, though support and features vary.
Related terms
Learn more
Read How to Connect an MCP Server to Your AI Assistant and browse examples such as the GitHub MCP server.