On November 11, 2026, Agenstry's public statistics listed 222,744 MCP server records. Of those, 76,932 had no endpoint a directory could probe remotely, 51,571 had been probed, and 15,060 answered with a live response. These numbers describe one index at one moment, not the size of the ecosystem.
Agenstry's mid-2026 directory survey makes the same point: many records are local MCP projects that start on a user's machine and never had a remote address for a directory to probe. Dividing the remote probe results by every record and calling the rest dead would be a mistake.
That is the first thing to sort out when reading a tool directory: records, projects, endpoints and tools can each be counted in their own way.
One Project Can Have Several Installation Methods
GitHub's official MCP server documents both a local setup and an official remote service. The local setup starts a program in the user's environment; the remote one connects to a hosted service. A directory that lists both gets two records, which does not necessarily mean two independent sets of capabilities.
Other choices change the total as well: whether old versions are kept as separate entries, whether community wrappers are included, whether MCP servers and A2A agents are counted together, and whether records are merged by project identity. Checking these rules is more useful than comparing totals directly.
| What is counted | Roughly what it means | What to ask next |
|---|---|---|
| Records | Which declarations or entries the directory keeps | Are versions, aliases and deployment methods counted separately? |
| Projects | Which software provides the tools | Is the same repository or package listed more than once? |
| Remote endpoints | Which addresses run a live service | When and how were they probed, and under what authentication? |
| Tools | Which callable operations a server exposes | Do those operations fit the task? |
Each number has its use. Finding candidates needs listing data, connecting to a remote service needs endpoint data, and planning a specific task needs the operations the tools provide.
The Same Server, Reported by Several Sources
Our own index shows how much overlap there is. agent-tools.cloud collects MCP listings from the official MCP Registry, Smithery, PulseMCP, Agenstry and other directories, then merges records that point to the same server. On November 11, 2026 it held 61,233 MCP listings built from 111,856 source records. 36,148 listings came from a single source; the rest were reported by two or more sources, and 1,436 by five or more.
The same index also gives different answers depending on the question. 60,610 listings had a remote endpoint we could probe. 52,568 of those were reachable, a figure that includes servers asking for credentials. 34,213 completed both the initialize handshake and tools/list, which is the closest of the three to "an anonymous client can see the tools right now". None of these numbers is the "real" size of MCP; each answers a different question.
Listings, Scores and Runtime Evidence Each Answer Part of the Question
The official MCP Registry is an entry point for publishing and discovering servers. Its publishing flow verifies namespace ownership, for example through a GitHub identity or a domain check, to confirm that the publisher may use a name. That evidence tells you who stands behind a declaration, which is a different question from whether a task will succeed.
Glama's Tool Definition Quality Score looks at tool definitions: purpose, conditions of use, parameter meaning and context. Clear definitions help a caller understand how to use a tool. They do not prove that a particular deployment can be reached now, or that a given output is correct.
Endpoint probing adds another layer: at a given time, under stated conditions, did the address respond, and did it return the expected protocol content? Read such results together with the method and the time window.
A service that requires authentication and rejects an anonymous request has not gone offline. In the other direction, an address that returns HTTP 200 is not necessarily a working MCP interface: a home page, an error page and a protocol response can all come from the same server.
Scores built from these signals measure different things, too. On agent-tools.cloud each listing now shows a quality grade, which sets search order and weighs availability, speed and how complete the description is, and a separate readiness score built on the nine criteria Agenstry publishes, which checks protocol version, declarations and a 30-day probe record. The two often disagree: a fast server on an older revision with thin declarations can rank high in search and still score modestly on readiness.
Using Directory Information for Your Own Task
If you want to connect a documentation tool to your assistant, the project description tells you which material it supports, the tool definitions explain query parameters and return values, and the setup guide covers client and authentication requirements. This information narrows the candidates.
Then look at evidence that matches your environment: the versions and connection methods the tool supports, when the service or project was last updated, and exactly which step a public verification record covers. Avoid merging maintenance scores, uptime records and task quality into one unexplained "reliable" label.
In the end, fitness still comes back to an approved, specific task. Give the tool a known document and a question, and see whether it retrieves the relevant content, keeps the source and completes the delivery in your own client. One success shows the result under those conditions at that time; it is not a long-term reliability record.
Totals Help Discovery; Methods Decide Comparison
A directory with more listings may help you find more candidates. Classification, deduplication, tool definitions and verification data decide how easy those candidates are to filter. Both kinds of value can exist at once.
When two totals differ, check the protocol scope, the local and remote split, the version rules, the deduplication method and the date. Once those match, the numbers can be compared directly. When they don't, treat them as different views and use each for what it measures.
A directory that serves its users well makes clear what an entry describes, where the information came from, which conclusions have evidence behind them, and what you still need to verify yourself.