MCP Server

An MCP server is a lightweight program that publishes a set of tools, data resources, and prompt templates for AI agents to use. It follows the Model Context Protocol, so any compatible client can connect, see what the server offers, and call those capabilities without custom integration code.

What a server exposes

Servers advertise three kinds of things. Tools are actions the model can invoke — create a ticket, query a database, send a draft for review. Resources are readable data the model can pull in, like a file or an API response. Prompts are reusable templates the server suggests for common jobs. A client discovers all three at connection time, which is why an agent can start using a brand-new server the moment it is plugged in.

Transports and a real example

Servers run in two common ways: locally over standard input/output, handy on a developer's own machine, or remotely over HTTP for a shared, hosted service. Picture a support team that builds a server exposing "search_tickets", "create_ticket", and "add_internal_note". An agent connected to it can be told, "Find open tickets about failed refunds and draft a reply to the oldest one." It calls "search_tickets", reads the results, then drafts a note — all through one server, with no bespoke wiring per assistant.

Server versus client

• The server holds the capabilities; the MCP client inside the AI app consumes them.
• One server can serve many clients at once.
• Good servers scope permissions tightly, since anything they expose an agent can attempt.

For the standard both sides follow, see the Model Context Protocol.