MCP server
An MCP server is a program that exposes tools, resources, or prompts to AI applications over the Model Context Protocol. It answers an initialize request with a protocol revision, then answers tools/list with the capabilities it offers. It can run as a local process or as a URL on the public internet.
What is an MCP server?
An MCP server is the side of the connection that owns capabilities. It waits for a client, agrees a protocol revision, and then describes what it can do in a machine-readable form the client can act on.
The server decides what a tool is called, what arguments it takes, and what it does. The protocol only decides how that description travels.
What does a server have to answer?
Two calls, at minimum: initialize, which agrees a protocol revision and reports the server's name, and tools/list, which returns the tools it exposes.
Everything else is optional. A server can offer resources and prompts as well as tools, can require authorization, and can send an instructions string telling the client how it expects to be used. On 5 August 2026, 8 of the 14 servers we connected to sent a non-empty instructions string; the other 6 left it out.
How big is a real MCP server?
Smaller than the discussion suggests. Across the same 14 servers, the median exposed 4.5 tools and 11 of the 14 exposed 10 or fewer.
The total was 288 tools, but a single game server accounted for 213 of them, which is 74% of the corpus from one address. An average across those servers would describe none of them. Most useful MCP servers are a handful of well-described functions, not a sprawling API surface.
Example
https://mcp.context7.com/mcp answered our initialize request as "Context7" on revision 2025-11-25, then returned exactly 2 tools, both with descriptions. https://game.spacemolt.com/mcp answered on the older 2025-03-26 revision and returned 213. Both are MCP servers by every definition the specification offers, and no single number describes them both.