What is an MCP Server?

A server is the provider side of MCP: it wraps some capability, such as a database, a SaaS API, or a local filesystem, and describes it in the protocol's standard shape so any compliant client can discover and use it without custom integration code. On connection, the server advertises its capabilities, meaning which tools, resources, and prompts it supports, and the client's host application decides how to surface those to the model and user. Servers are typically narrow and single-purpose by design, such as one server per integration, which keeps them simple to build, audit, and swap independently. They can run locally as a subprocess talking over stdio, or remotely as a networked service reached over Streamable HTTP, depending on whether the resource they wrap is local or cloud-hosted. Because a server's tool descriptions and returned data become part of what the model reasons over, a well-built server is careful about what it exposes and returns, treating its own output as something the host and user should be able to trust, not just functionally correct code.