What is an MCP server?
Definition
An MCP server is a program that exposes a set of tools to AI assistants using the Model Context Protocol. Each tool has a name, a description, an input schema, and hints about whether it reads or writes. Assistants connect to the server, list its tools, and call them during a conversation, with the user signing in first when the server requires it.
Tools and annotations
A server's tool list is its whole interface. Good servers keep tools narrow (one job each), give every tool a title, and mark read-only tools apart from ones that create, change, or delete data. Assistants use those hints to decide when to ask the user before acting.
Remote and local servers
A local MCP server runs on the user's machine and is started by the assistant. A remote MCP server runs on the internet at an https address and is shared by everyone who connects; that is the kind products publish, because it needs no installation and can sign users in with OAuth.
Authentication
Remote servers typically use OAuth: the assistant registers itself as a client, sends the user to the product to sign in and approve, and receives a token. The token is scoped, so a server can offer read-only access separately from write access.
Example
The Emboss MCP server exposes eleven tools for PDF forms, from listing a user's forms to filling one for every row of a spreadsheet, behind OAuth sign-in with an Emboss account.