The Model Context Protocol, introduced by Anthropic as an open standard in November 2024, splits an AI integration into two sides. The client lives inside the AI application and speaks to the model. The server owns a domain, such as a database, a code host, a design tool or a work system, and describes what it can do in a form the client understands. Because both sides follow one protocol, one server works with many clients.
A server offers three kinds of capability. Tools are actions the model can call, with a name, a description and a typed input schema. Resources are data the client can read, addressed by URI. Prompts are reusable templates a user can pick. Servers run either locally, started by the client as a process and spoken to over standard input and output, or remotely, reached over HTTP. Remote servers usually sign users in with OAuth, so the server knows whose data each call touches.
For the people using an AI client, a good MCP server feels like a built-in ability. For the people building one, the hard part is not the protocol but the design: which actions to expose, how to describe them so the model uses them well, and how to keep authority with the signed-in user rather than with whatever the model types.