MCP server
An MCP server is a program that publishes a list of actions an AI client can call, each with a name, a description and a JSON schema. Remote servers are reached by URL and OAuth; local ones are packages you install and run yourself.
The server is the half of MCP that does the work. When a client connects, it returns a list of tools — each with a plain-English description the model reads like a menu — and optionally resources (documents addressable by URI) and prompts (reusable templates). In practice almost every marketing-relevant server is a list of tools and nothing else.
The distinction that decides whether a server is usable at all is remote versus local. A remote server runs on the vendor’s infrastructure: you paste an HTTPS URL, sign in with OAuth, and it works from any device with the grant revocable from your assistant’s settings. A local server is a package installed on one machine, usually with a long-lived API key in a plain JSON config file, and it works only there.
The scale is larger than most write-ups suggest and less useful than it sounds. Counted from the official registry on September 14, 2026, there were 31,793 distinct servers; 19,332 published a remote endpoint and 12,461 were local-only packages. A registry entry is a published manifest, not a maintained product.
Almost every MCP server is a wrapper around an API that already existed. What the wrapper adds is self-description, so a model that has never seen the system can decide to use it — which is why the useful question is never “MCP or API” but “does something need to choose at runtime”.