Statistics
What 60 MCP servers returned when we called them
Every MCP directory ranks servers on GitHub stars, commit activity and listing age. None of them connect to what they list. We did, 62 times across 60 public endpoints between 24 July 2026 to 27 September 2026, and this is what came back.
Every figure below is re-derivable from our own connection records. Each endpoint is counted once, using its most recent test. Checks run through the public tester are excluded, because a URL a stranger pasted says nothing about the MCP ecosystem.
Most of them refused an anonymous connection
The single most common result was not a connection and not a failure. It was a server asking who we were. 37 of 60 endpoints (62%) answered a tools/list call with a 401 or 403.
That is a working server, not a broken one, and it is the correction most worth publishing here: a 401 is evidence the server is alive and guarding itself. Anyone reporting it as unreachable is making a claim they did not verify. It is also why we stopped treating a successful connection as a condition of being listed.
| What happened | Endpoints | Share |
|---|---|---|
| Authorization required | 37 | 62% |
| Connected | 14 | 23% |
| Not an MCP endpoint | 3 | 5% |
| Protocol error | 3 | 5% |
| Hostname did not resolve | 2 | 3% |
| Unreachable | 1 | 2% |
The denominator is endpoints, not checks. We ran 62 checks against 60 endpoints, so a share computed over checks gives a different and equally true number. Any figure quoted from this page should say which one it is.
The servers we reached were small
Across the 14 endpoints that connected and listed their tools, we recorded 288 tools in total. The distribution is skewed hard by a single outlier, so the average is the wrong summary and the median is the right one.
- Median tools
- 4.5
- Fewest
- 2
- Most
- 213
- Total recorded
- 288
All 14 of the 14 connected servers carried a description on their tools. That is worth stating plainly because it is a negative result: a quality floor that rejects servers for undescribed tools would have rejected nobody in this sample.
3 protocol revisions answered our handshake
The revision each server negotiated during the handshake, across the ones that connected. A client written against a single current revision would have been assuming something this sample did not do.
| Protocol revision | Servers |
|---|---|
| 2025-11-25 | 9 |
| 2025-03-26 | 3 |
| 2025-06-18 | 2 |
How we measured this
One MCP initialize handshake followed by one tools/list call per endpoint, over Streamable HTTP, anonymously, from our own server. Endpoints came from two curated public catalogs. Nothing was executed: a stdio server has no address to call, and we do not run commands chosen by strangers. You can run the identical check against your own server and get the same shape of record back.
How to read these numbers
Worth being precise about, because a page of numbers invites conclusions the numbers do not support.
- • "Asked for credentials" is a measurement, not a rating. We recorded that a server wanted a token. Auditing its code, its data handling and its permissions is a different job, and requiring authorization is not the same as being well built. Any directory presenting a check like this as a safety rating is overselling it.
- • Timings stay out of it. We store them and publish none. The checks ran back to back from one connection, so the spread measures our own network at least as much as anyone's server. A latency chart from that data would be a fact about our afternoon sold as a fact about the ecosystem.
- • Every figure has a date on it. Each endpoint was checked once, on the date recorded, and never again. The numbers above describe 24 July 2026 to 27 September 2026.
- • This is a sample of the reachable ones. 60 endpoints we chose, all of them remote. Servers distributed as an npx or PyPI package have no address to call. The registry census counts how much of the catalog that actually is.
Check your own server
The MCP connection tester runs exactly the check described above and hands back a dated record you can share. If you want to know what each field in a real tools/list response means, the field-level reference is written against the same code that produced this page, and why a 401 is not a failure has the long version of the finding at the top of this page.
Built an MCP server? Launch it on the weekly board. Give us a URL and we connect to it and publish what it answered, the same way as everything above.
62 checks across 60 endpoints, 24 July 2026 to 27 September 2026. Figures on this page update when the corpus does.