— THE NUMBERS

The MCP ecosystem, measured.

We track 95,574 MCP servers, agent skills and plugins. 54,313 carry a score, 9,162 have been scanned for advisories at the version you would install today, and 72% of those ship with no build provenance.

Every figure on this page is a count of rows, recomputed nightly and derived from public evidence. Nothing here is behind the paywall — a fact you have to pay for is a fact nobody repeats. How we measure · What changed recently

What is measured

95,574

capabilities tracked Every MCP server, agent skill and plugin we have discovered, measured or not.

54,313

carry a tashan score Enough public evidence to rank on upkeep, freshness and real adoption.

9,162

scanned for advisories Queried against OSV.dev at the version you would install today, so a finding means the current release is affected.

3,969

graded on how well they document themselves Read against a published, conjunctive rubric rather than a feeling.

What the ecosystem looks like

2,872

are deprecated, archived or abandoned Their publisher, npm or the registry says to stop using them.

22,980

have a single primary maintainer One person away from unmaintained. Not a fault, but a fact worth knowing before you depend on it.

516

run a script on your machine at install time A postinstall or preinstall hook — code that runs before you have agreed to anything.

94%

of those install scripts run code we CANNOT read The hook runs a file inside the tarball — postinstall.js, install.js — and a static check cannot see what is in it without downloading and running the package, which no auditor should do. This is the honest limit of every scanner, including ours.

72%

of scanned packages ship with NO build provenance npm signs every tarball it hosts, so a signature proves nothing about who built it. Only a build attestation ties the artifact to its source.

6

are in OSV's malicious-packages database Not a vulnerability — the package IS the attack. They are refused a place on every ranking and kept reachable only so a tool can warn someone already running one.

Coverage, weighted by demand

The aggregate ratio falls every time discovery succeeds — finding a thousand new capabilities makes the fraction we have measured smaller while nothing has got worse. So it is reported and never targeted. What we commit to is the demand curve, because what costs a reader is an absence on the thing they looked up, not a gap in the tail.

100%

of the top 100 by adoption are scanned for advisories

98%

carry a score

30%

are graded on documentation

The record behind it

2,186,635 measurements across 45 days. That series cannot be backfilled by anyone, including us — you cannot know what a score was in June unless you measured it in June — which is why what changed cannot be rebuilt from today’s data — only from having recorded it every day since July.

Cite this

Published under CC BY 4.0. Quote any number here, with attribution. If you are writing about which MCP server to use, these are the figures behind the rankings — and llms.txt is the machine version.