An MCP server is a small program that sits in front of one system and offers its capabilities to an AI agent in a standard shape, so any agent that speaks the Model Context Protocol can call that system without custom integration code being written for it first.
Before MCP, connecting an agent to a tool meant writing glue: read the vendor's API docs, wrap the endpoints you care about in functions, describe those functions to the model, handle the auth, and then do the whole thing again for the next tool and again for the next agent framework. The work scaled with tools multiplied by agents.
An MCP server collapses that. The server is written once, by whoever knows the system best, and it advertises three kinds of thing to any client that connects: tools, which are actions the agent can invoke; resources, which are pieces of context the agent can read; and prompts, which are reusable instruction templates the user can trigger. The agent discovers what is available at connection time rather than being told in advance, which is why the same agent can pick up a server it has never seen and use it correctly.
Two things an MCP server is not. It is not a hosted service you sign up for, necessarily: many run as a local process on your own machine and talk to the client over standard input and output. And it is not a security boundary you get for free. The server runs with whatever credentials you hand it, and it can only be as careful as the person who wrote it.
This directory checked 293 go to market tools and found 144 with an official server, the vendor's own, plus 21 where somebody outside the vendor built one. That is 56% of the entries reachable by an agent through MCP at all. 117 had no server found on the date they were checked, and 7 could not be settled either way.
The gap between those two numbers is the whole reason this directory exists. A category can be sold entirely on the language of AI and still have almost nothing an agent can call.
Every number on this page is generated from directory.json at build time and carries the date it was baked: 2026-08-25. Nothing is typed by hand, nothing is rounded, and nothing is estimated. The underlying data is published in full. Where the honest answer is a zero, the zero is printed.