MCP Server

What is an MCP server?

An MCP server is a program that makes capabilities available to an AI application through the [Model Context Protocol](/glossary/model-context-protocol/). Depending on what it implements, a server can expose tools the model may invoke, resources the application may read, and prompt templates a user or client may select.

“Server” describes its protocol role, not necessarily a remote machine. An MCP server may run as a local subprocess on the same computer as the AI application or as a network service.

How it works

An MCP host creates an [MCP client](/glossary/mcp-client/) that communicates with the server. The participants exchange protocol information and supported capabilities. The client can then discover the server’s available features and use the relevant MCP operations.

The server owns the implementation behind those features. A filesystem server might read approved files; a CRM server might expose a customer lookup tool; a documentation server might publish reference pages as resources. MCP standardizes the interaction boundary, but it does not decide how the host places returned content into a model’s context or how the model reasons about it.

Simple example

Suppose a support agent needs order information. An MCP server exposes get_order with a defined input schema. The client lists that tool, the model selects it, and the client requests execution through MCP. The server checks the request, queries the order system, and returns a result.

MCP server versus API server

An API server exposes application endpoints to software clients. An MCP server exposes protocol-defined capabilities in forms designed for MCP-compatible AI applications. It may call one or more APIs internally, but it also participates in MCP discovery and capability semantics.

An existing API does not automatically become an MCP server. It needs an MCP-facing layer that describes and handles the supported MCP features. Likewise, MCP does not replace every API; it can provide a standardized AI integration over existing services.

Why it matters

MCP servers separate external capability logic from the AI application. This can make integrations reusable across compatible hosts while keeping data access, validation, and authorization close to the system that owns them.

That boundary is not a security guarantee by itself. Servers still need authentication, authorization, input validation, least-privilege access, and safe error handling. See [Tool Use in AI Agents](/tool-use-in-ai-agents/) for the wider execution pattern.