What is an MCP server?

MCP server, explained#

The server is the half of MCP that knows about a system: a database, a file tree, a SaaS product, an internal API. It describes what it can do in a machine-readable way and does the work when asked. The AI application never needs to know how the system works underneath, only what the server says it offers.

When a client connects, the two exchange an initialize request in which each declares its capabilities. The client then calls methods such as tools/list to see the available tools with their names, descriptions and JSON Schema inputs, tools/call to run one, resources/list and resources/read for context, and prompts/list and prompts/get for templates. A server need not offer all three; many offer only tools.

Servers come in two shapes. A local server runs on the same machine as the host, launched as a subprocess and spoken to over stdio. It typically reads its credentials from environment variables. A remote server runs on the internet behind an HTTPS endpoint and is shared by many users, so it has to authenticate each request and keep each user's data apart.

Because a server can run code and change data, the specification treats tool calls as something the host must get the user's consent for, and it tells hosts to treat tool descriptions and annotations from untrusted servers as untrusted. A well-built server limits each connection to what the user has approved and records what was done.

For example, a team writes a small MCP server around its billing system with two tools, find_invoice and mark_paid. Connected to an assistant, a person can ask, "Has Acme paid March?" and the assistant calls find_invoice, reads the status and answers. Marking the invoice paid needs a second, explicit approval.