Model Context Protocol (MCP) interview questions

Before MCP, every AI app needed its own bespoke connector to every tool — with N apps and M tools that's N×M one-off integrations to build and maintain, and a connector for one app was useless to the next. MCP (Model Context Protocol) defines one common protocol both sides speak, so each tool exposes itself once as a server and each app speaks MCP once as a client. The N×M tangle collapses to a tidy N+M, and anything new instantly works with everything that already exists.

There are two roles. The host app — Claude Desktop, an IDE, a coding agent — runs an MCP client; the tool or data source runs an MCP server. They exchange small structured messages using JSON-RPC, either over standard input/output for a local server or over HTTP for a remote one. When the client connects it first does a handshake to learn what the server offers, so neither side has to be hard-coded with knowledge of the other.

Three kinds of things, all discovered automatically on connect. Tools are functions the model can call, like send_email or query_database — the same idea as ordinary tool calling, standardized. Resources are read-only data the app can pull in as context, like a file's contents or a database row. And prompts are reusable templates the server suggests, such as a ready-made "summarize this ticket" instruction.