MCP in numbers · official MCP registry · counted 2026-10-28
MCP in numbers
Every figure we count on the official MCP registry, each with its date and source. They are counted on all entries, not a sample, and they are counts of public records, not verdicts about a server. Each server's own page says when its facts were checked.
homepage · every server · changes, day by day · how we measured · tool-list sizes in tokens
The registry list, pull by pull
38,977 on 2026-10-04 → 51,834 on 2026-10-28 (+12,857 in 24 days)
The numbers
| pull | servers | new or listed again · no longer listed |
|---|---|---|
| 2026-10-28 | 51,834 | 671 · 5 (since 2026-10-27) |
| 2026-10-27 | 51,168 | 724 · 9 (since 2026-10-26) |
| 2026-10-26 | 50,453 | 1,014 · 6 (since 2026-10-24) |
| 2026-10-24 | 49,445 | 928 · 34 (since 2026-10-23) |
| 2026-10-23 | 48,551 | 9,803 · 229 (since 2026-10-04) |
| 2026-10-04 | 38,977 | first full pull |
Findings · today’s records · registry list of 2026-10-28 · facts checked 2026-10-22 to 2026-10-28
Where listings and reality disagree, today
The five checks of our first report, with the same definitions, counted on the current records of all 51,834 entries and 32,168 publishers. Each count is the number of public records that show it, on the dates above; it is not a verdict about a server, and each server page says when its own facts were checked. An entry can appear in more than one row; “any of the below” counts it once.
| Mismatch | Registry check | Share of entries | % of all 51,834 entries | % of all 32,168 publishers |
|---|---|---|---|---|
| Any of the below 6,025 entries · 3,560 publishers | 11.6% | 11.1% | ||
| Repository link can’t be opened publicly 4,718 of 35,047 with a GitHub repository link · 13.5% | never checked | 9.1% | 9.0% | |
| Listed version is withdrawn upstream 738 of 17,096 shipping an npm or PyPI package · 4.3% | never checked | 1.4% | 0.8% | |
| Active, but declares no way to run it 638 of 51,834 in the registry · 1.2% | allowed by schema | 1.2% | 1.1% | |
| Linked repository is archived 333 of 30,145 with a reachable GitHub repository · 1.1% | never checked | 0.6% | 0.5% | |
| Listed version can’t be installed 231 of 17,096 shipping an npm or PyPI package · 1.4% | verified once | 0.4% | 0.5% |
sources api.github.com · registry.npmjs.org · pypi.org · registry.modelcontextprotocol.io · facts checked 2026-10-22 to 2026-10-28 · counted 2026-10-28 · definitions and caveats
Tool lists and security policies · data as of 2026-10-28
What the servers say about themselves
We also read each server's own tool list, still without calling a tool, and the public security policy of its GitHub repository. An AI model read every tool description (two differently worded runs had to agree, otherwise no label); the policy text is the publisher's own words. Nobody has checked the AI labels by hand. 27,912 servers have no tool list we could read yet.
- 23,922
- servers whose tool list we read (392,119 tools)
- 9,620
- of them with a tool whose description says it changes something
- 8,460
- servers whose repository publishes a security policy (of 28,478 repositories read)
- 314
- of those with a statement about security research, quoted on the page
Servers by what their tools doSearch MCP tools by name
sources: the official MCP registry, our read-only sessions, npm audit and OSV.dev, and each repository's public SECURITY.md · what’s in the registry · how we measured
Tool list size · counted with Anthropic's token counter · figures as loaded on 2026-10-28
How many tokens a server's tool list adds
A client that loads a server's tools sends that list to the model with every request. For the servers whose tool list we read (no tool called), we counted how many tokens it adds. Each server page gives its own count; this is how the counts spread. A count, not a rating.
Counted servers by tool-list size, in token ranges
23,424 servers counted · median 2,725 tokens · 90th percentile 13,920
The numbers
| tokens | servers | of counted |
|---|---|---|
| under 1K | 3,081 | 13.2% |
| 1K–2K | 5,492 | 23.4% |
| 2K–5K | 6,576 | 28.1% |
| 5K–10K | 2,726 | 11.6% |
| 10K–20K | 3,766 | 16.1% |
| 20K–50K | 1,567 | 6.7% |
| 50K or more | 216 | 0.9% |
Transparency · why a server has no result
Why 26,368 servers are not security-tested
Of all 51,834 registry entries, we opened a live session with 25,466. The other 26,368 have no exercised checks, and each one says why on its own page. Here is the split by whose limit it is — our limit, the server’s own choice, a defect in the package, or a cause the record does not pin down. A check we did not run is never counted as passed.
opened in a live, read-only session · no tool called
25,466 49.1%
not opened: the server's own choice · required sign-in or payment, or declares no way to run
15,732 30.4%
not opened: a defect in the package · the published package does not start
1,794 3.5%
not opened: our limit · our sandbox, client or time limit — ours to fix
2,396 4.6%
not opened: cause not determined · no session, and the record does not say which side
6,446 12.4%
Run one of these servers? You can test it yourself, from your own machine, with our free command: a server it starts there, or your own server at its URL after a one-time proof that you control it, with your own credential if it needs one.
The full breakdownTest your own server
counted from the current server records and our read-only sessions · data as of 2026-10-28 · every reason code and what it means
First report · registry snapshot 2026-10-04 · checked 2026-10-06 · not updated
Where listings and reality disagreed in our first report
13.9% of all 38,977 registry entries had at least one problem the registry doesn’t catch (registry snapshot 2026-10-04; 23,016 publisher namespaces; sources npm · PyPI · GitHub; full count, not sampled; last checked 2026-10-06)
This is the published report, kept exactly as it was (it is cited elsewhere). Every percentage uses the same base — all 38,977 entries, or all 23,016 publishers — so no row can exceed the total. Each row’s own subset is shown beneath it. The registry check column is read from the registry’s own validator code.
| Mismatch | Registry check | Share of entries | % of all 38,977 entries | % of all 23,016 publishers |
|---|---|---|---|---|
| Any of the below 5,435 entries · 3,169 publishers | 13.9% | 13.8% | ||
| Repository link can’t be opened publicly 4,298 of 28,743 with a GitHub repository link · 15.0% | never checked | 11.0% | 11.3% | |
| Listed version is withdrawn upstream 663 of 14,431 shipping an npm or PyPI package · 4.6% | never checked | 1.7% | 1.0% | |
| Active, but declares no way to run it 459 of 38,977 in the registry · 1.2% | allowed by schema | 1.2% | 1.3% | |
| Linked repository is archived 309 of 24,256 with a reachable GitHub repository · 1.3% | never checked | 0.8% | 0.6% | |
| Listed version can’t be installed 157 of 14,431 shipping an npm or PyPI package · 1.1% | verified once | 0.4% | 0.5% |
sources api.github.com · registry.npmjs.org · pypi.org · registry.modelcontextprotocol.io · checked 2026-10-06 · full report, definitions and caveats