MCP Server
An MCP server is a program that exposes capabilities to AI agents over the Model Context Protocol. It publishes MCP tools the agent can invoke, MCP Resources the agent can read and MCP Prompts that shape how the agent works with it, so any compatible client can discover and use it without integration code written for that one app.
Why It Matters
A vendor ships an MCP server for its own product, the way Atlassian does for Jira and Confluence. A community author publishes one to the public registry. An internal team builds one over a private system such as a data warehouse. Which of those three a company is running decides who it can ask when the server breaks.
Before an MCP server does anything, it describes itself. When a client connects, the server lists the MCP tools it offers, each with a name, a description and a schema for its arguments. The language model chooses from that list, so the wording of a tool description shapes its behavior as much as the code behind it. A vague description produces a tool the agent calls at the wrong moment.
The three things an MCP server exposes differ in who decides to use them. The language model decides when to call an MCP tool. The application decides which MCP Resources to pull in as context. The person invokes an MCP Prompt deliberately, which is why prompts show up as slash commands rather than as something the model reaches for.
An MCP server that a company's agents use is a system somebody else operates. It reads the arguments of every call, returns whatever it chooses and can change what it exposes between one call and the next. That is why an employee adding a public server to their MCP host is making a decision with more reach than installing a browser extension, and why the practice has its own name, shadow MCP.