AI glossary · Basic · Agents
MCP
MCP (Model Context Protocol) is an open protocol that lets AI apps connect to external resources such as files, databases and work tools in the same way. A connector built once for MCP can be used as-is in many AI apps that support MCP.
About this term
- Level
- Basic · Technical terms you often meet in AI news
- Category
- Perception & action
- Fields
- Agents
- Added
- 2026-10-08
How it works
MCP has three roles: the AI app the user uses (the host), the MCP client inside that app that handles connections, and the MCP server that actually provides tools or data. An MCP server announces, in a defined format, a list saying "these are the tools and resources I can provide," and the AI model picks one of them and calls it when needed. When the call result returns to the model, the model continues building its answer based on that content. A server can run on your own computer or be a remote server across the internet.
Why it matters
Before MCP, connection code had to be written separately for each AI app and each tool. As the number of apps and tools grew, the number of connections to build multiplied. With a common protocol, the tool side only needs to build one MCP server, and the app side only needs an MCP client. As a result, the number of tools available to AI agents grew quickly, and many AI companies and developer tools came to support MCP.
Things to know
MCP servers hand real permissions to the AI, such as editing files or sending messages. So you need to be careful about installing servers from unclear sources and about prompt injection, where hidden instructions are mixed into retrieved documents or tool descriptions. If too many tools are connected, the tool descriptions take up a lot of the context window and the model may pick the wrong tool. It is safer to grant only the permissions that are truly needed and to have a person confirm actions that are hard to undo.
Example
If you connect an MCP server for GitHub to a coding agent or desktop AI app, then when the user says "summarize the open issues," the AI can fetch the issue list itself and summarize it. Because the same server can be connected to other MCP-supporting apps, a company may build just one MCP server for searching internal documents and share it across several AI tools.