On 2026-08-14 we dumped every MCP server in the Skillselion catalog and counted what the rankings you read are made of. The result became the MCP Server Census: 8,433 servers from 5,392 publishers, and not one of them carries an install count.
Disclosure: we run Skillselion, the directory this data comes from.
No install number exists. Not a low one, none at all.
Agent skills report installs through public registries, which is how our Agent Economy Census can count 158M installs across 79,827 listings. MCP has no per-server install event. All 8,433 servers in the census carry no install figure. Partial signals exist (npm, PyPI and Docker download counters for servers shipped as packages), but nothing that covers the ecosystem.
So every "top MCP servers" table you have seen ranks on something else. Usually stars.
Stars belong to the repository, not the server
79.8% of servers (6,731) map to a GitHub repository, and 175 repositories host more than one server, covering 969 servers. Each of those servers inherits the star count of whatever it ships inside.
Concretely: the Metabase MCP server shows 48,750 stars, which belong to the business-intelligence platform. The PostHog MCP server shows 37,662, earned by the analytics product. ByteDance's UI-TARS browser MCP server inherits 38,581 stars from the desktop app it lives with. The top of a star-ranked MCP table is largely a list of whole products, with a server attached.
The inequality is close to the theoretical maximum
Run a Gini coefficient across the star counts of all 8,433 servers and it comes out at 0.986. The top 1% (84 servers) hold 89.1% of every star in the ecosystem. The top ten alone hold 43.6%. Meanwhile 54.9% of servers have never been starred once, and the median starred server holds 5.
Publish an MCP server today and the most likely measurable outcome is silence.
The code is alive anyway
Half the catalog having zero stars sounds like abandonware. The push dates disagree: of the 5,926 servers with a last-push date, the median pushed about 50 days ago, and 43.4% pushed within the last 30 days. Only 6.2% of dated servers have gone six months untouched. Stars measure visibility. They do not measure whether anyone is working on the thing. We ran the same test on agent skills in the Maintenance Census and found the same disconnect.
87.3% of publishers ship exactly one server
5,392 publishers account for the census, and 4,707 of them published a single server. The ten largest publishers together hold only 11.27% of all servers, nothing like the suite concentration in the skill economy. The largest publisher by volume, CSOAI-ORG, ships 192 servers that hold 15 stars between them. Volume is cheap here. Being noticed is not.
What to check instead of stars
- Recency: last push under 60 days beats a four-digit star count from a repo the server shares. The MCP server directory exposes recency per listing, and the ranked MCP servers for the build stage apply it.
- Repo scope: check whether the stars are for the server or for a product it ships inside.
- Audit trail: our Security Census found most of the ecosystem has never been audited; the Clone Census measures how much of it is copies.
- Publisher shape: a one-server publisher and a 192-server publisher carry different maintenance expectations.
Every number above is frozen to 2026-08-14 so you can check it; the dataset and limits are in the research methodology. Our catalog is not the whole ecosystem (PulseMCP reported 22,070 servers the same day), which is exactly why none of these numbers claims to measure usage.
Top comments (1)
"What gets measured gets managed" — or in this case, completely distorted.
As long as the only public metric is stars from parent repos, these lists are useless. Connecting an MCP to a production CRM or DB is a security nightmare without proper governance and auditing, but stars will never show you that. Great work bringing actual data to the table.