Skip to main content
Skills And MCP Basics

What Is An MCP Server?

Understand what MCP servers are, what transports they use, and why agents rely on them for live tools.

Short answer

An MCP server is a program that exposes tools, prompts, or resources to an AI client over the Model Context Protocol. Instead of packaging instructions for the agent to read once, it stays available as a live integration the client can call during a session.

That makes MCP servers the right abstraction for databases, browsers, file systems, SaaS APIs, issue trackers, and other capabilities that need runtime interaction.

How MCP works in practice

MCP follows a client-server design. The AI host is the client. The MCP server advertises a set of tools and optional resources. The host discovers those capabilities, lets the model select or invoke them, and sends requests over a supported transport such as standard I/O, Server-Sent Events, or HTTP.

That architecture is what allows the same server to work across multiple clients without every vendor inventing a brand new plugin protocol.

Why MCP is different from ordinary package installs

An npm package is just a package. An MCP server is a runtime integration surface. Once configured, it can receive user data, agent-generated plans, file paths, arguments, and auth material. It can also expose read-only or write-capable tools.

That is why installation guidance alone is not enough. You also need to know the permissions and trust assumptions around the server you are enabling.

Typical examples

  • A browser automation server such as Playwright.
  • A design integration such as Figma.
  • A GitHub or Linear integration that lets the model read and update issues.
  • An internal documentation or database bridge that provides search and fetch tools.

Sources and further reading