Definition
The Model Context Protocol is an open standard for connecting language-model applications to external capabilities. A server exposes tools, resources, and prompts over the protocol; a client — typically the application hosting the model — discovers what is available and invokes it. The point is to make integrations reusable: one server can serve any compliant client instead of each application writing its own adapter for every system it touches.
Architecturally it is the familiar move of replacing an N-times-M integration mesh with a common interface, which is why it spread quickly once agent applications started needing the same handful of connections — file systems, databases, issue trackers, internal APIs. The security considerations are the interesting part in a design discussion: a server defines what an agent is able to do, so its scoping, authentication, and the human-approval boundary around destructive actions are load-bearing rather than incidental.
How it comes up in interviews
Naming MCP where it belongs signals that you have followed how agent tooling actually evolved, and it converts a vague "the agent calls our systems" hand-wave into a concrete interface with a discovery story. Expect follow-ups on authorization and blast radius — what stops the agent from taking an irreversible action, and who approves it.