The MCP Server Census

Finding 01

8,433 servers, and not one install count

Every MCP server in this census carries a zero install count. Not a low number: no number at all, for all 8,433 of them. Agent skills report installs through public registries, which is what makes the agent economy census possible. MCP has no equivalent per-server install event.

Partial signals do exist, and they are worth being precise about. Servers shipped as packages inherit a download counter from npm, PyPI or Docker, and PulseMCP surfaces that as a package_download_count field, populated for a minority of the servers it lists. We are not putting a precise share on that, and the reason is worth stating: repeated samples of its API on 2026-08-14 disagreed by several points, because coverage is heavily clustered along the catalog’s ordering, with individual pages running from 0% to 89%. A sample of a few thousand rows therefore carries a confidence interval too wide to quote a share from. What held across every sample we took is that the field is the exception rather than the rule. Even where a download number exists it counts package pulls, including from CI and mirrors, rather than use of the server, and it is missing entirely for anything distributed as a remote endpoint or a bare repository.

What does not exist is a per-server usage figure covering the population, which is why this report ranks nothing by adoption. The rest of the census measures how the available substitute, GitHub stars, behaves when you lean on it.

Finding 02

Stars belong to the repository, not to the server

79.8% of servers (6,731) map to a GitHub repository, and those repositories are frequently much larger things than an MCP server. 175 repositories host more than one server, covering 969 servers (14.4% of the repo-backed population), and each of those servers inherits the full star count of the repository it lives in.

So the top of any star-ranked MCP table is largely populated by whole products. Metabase’s 48,750 stars are for the business-intelligence platform. PostHog’s 37,662 are for the analytics product. ByteDance’s 38,581 belong to the UI-TARS desktop app and are counted four times, once per MCP server in the repository. Adding stars per server instead of per repository inflates the ecosystem total from 1,101,230 to 1,270,936, a factor of 1.15x.

1. 1modelcontextprotocol/servers - reference server monorepo89,547
2. 2D4Vinci/Scrapling - scraping library73,854
3. 3upstash/context7 - docs product60,701
4. 4metabase/metabase - the whole BI product48,750
5. 5bytedance/UI-TARS-desktop - desktop app, 4 servers inside38,581
6. 6PostHog/posthog - the whole analytics product37,662

These six are a selection chosen to show the effect, not the top six by stars: they sit at positions 1, 3, 4, 6, 8 and 9 in the repository star ranking. On a per-server ranking PostHog falls to 12, because the four ByteDance servers each occupy a slot with the same repository count, which is the very effect this finding is about. The numbered badges above order the selection, not the ecosystem. A further 20.2% of servers (1,702) carry no repository at all, so they are excluded from every repository-level figure here.

Finding 03

Star inequality reaches a Gini of 0.986

Run the Gini coefficient across the star counts of all 8,433 servers and it comes out at 0.986. The top 1% of servers, 84 of them, hold 89.1% of every star in the ecosystem. The top ten alone hold 43.6%.

At the other end, 54.9% of servers (4,633) have never been starred once, and among the 45.1% that have, the median server holds 5 stars. Publishing an MCP server, for most people, produces no measurable signal of any kind.

1. 1Top 10 servers43.6%
2. 2Top 1% (84 servers)89.1%
3. 3Top 100 servers90.5%

The August skill census reported a Gini of 0.96 on installs. That figure and this one describe different populations measured on different metrics, installs against stars, so the two are not directly comparable and should not be read as a like-for-like ranking of inequality.

Finding 04

The code is alive, whatever the stars say

The obvious reading of a catalog where half the entries have no stars is that it is full of abandoned projects. The push dates say otherwise. Of the 5,926 servers carrying a last-push date (70.3% of the census), the median server pushed code about 50 days ago, and 43.4% of them pushed within the last 30 days. Measured against the whole census rather than the dated subset, that is 30.5%.

Only 6.2% of dated servers (369, or 4.4% of the census) have gone six months untouched. A star count near zero is measuring visibility, not whether anyone is still working on the thing.

The dated subset is not a random sample. The 2,507 servers without a push date are 97.8% zero-star and hold 299 of the ecosystem’s 1,270,936 stars, so they are systematically the thinnest entries and their exclusion flatters this finding. Both denominators are given above for that reason.

1. 1Pushed within 30 days2,570
2. 2Not pushed in 180 days369

Finding 05

87.3% of publishers ship exactly one server

5,392 distinct publishers account for the census, and 4,707 of them (87.3%) have published exactly one server. The average publisher ships 1.56. Unlike the skill economy, where a handful of suite publishers dominate the install table, the ten largest MCP publishers together account for only 11.27% of all servers.

Server count and attention also come apart completely. The largest publisher by volume, CSOAI-ORG, ships 192 servers holding 15 stars between them. Volume is cheap here; being noticed is not.

1. 1CSOAI-ORG192
2. 2Ansvar-Systems121
3. 3cyanheads115
4. 4Br0ski777101
5. 5io.github.Evozim98
6. 6mcparmory76
7. 7clauxel72
8. 8nirholas63

Finding +

Bonus: what the servers actually do

An MCP server is defined by the tools it exposes, and most of them do not say what those are. Only 1,115 servers (13.2%) declare at least one tool; the remaining 86.8% publish no tool manifest we can read. Among those that do declare, the median server exposes 6 tools, and the declared total across the census is 14,349.

By category the census is concentrated: developer tooling and AI plumbing together account for just over half of every server in it.

1. 1Developer Tools (31.5%)2,654
2. 2AI & LLM Tools (18.8%)1,588
3. 3Data Analytics (11.4%)962
4. 4Finance (10.8%)911
5. 5Security & Pentesting (7.7%)649
6. 6Communication (5.9%)501

Practical

How to pick an MCP server when nothing ranks them

If installs do not exist and stars measure the host repository, the useful move is to rank on the fields that are actually populated. In order of how much they told us in this census:

1. Last push date. The strongest populated signal. The median dated server pushed about 50 days ago and 43.4% pushed inside 30 days, so a server untouched for six months is in the bottom 6.2% and worth a second look.
2. Declared tool count. Only 13.2% declare one, but where present it tells you the surface area you are adopting; the median is 6 tools.
3. Whether the repository is the server. Before reading a star count, check the repository is not a whole product that happens to contain an MCP server, which is the trap in Finding 02.
4. Publisher shape. 87.3% of publishers ship exactly one server, so a suite publisher and a one-off carry different maintenance expectations.

You can apply all four on the MCP server directory, which exposes recency, category and publisher per listing, and cross-check the wider catalog on the leaderboard.

Method

Methodology

The census is a complete dump of every type=mcp listing in the Skillselion production catalog on 2026-08-14, pulled through the public listings API using keyset pagination: 8,433 rows, matching the catalog’s own reported MCP total on the same date. Star counts are GitHub repository stars as recorded in the catalog, spot-verified against the GitHub API on the snapshot date; the six repositories named in Finding 02 all matched live GitHub to within about fifty stars.

This is our catalog, not the ecosystem. Other public registries count more. On the same day we took this snapshot, PulseMCP’s API reported 22,070 servers against our 8,433, and the official MCP registry paginates without publishing a total. We tried to put a number on how many of PulseMCP’s servers carry a download count and could not: three sampling draws taken the same day disagreed across a five-point range, because that field clusters strongly along the catalog’s ordering and no sample of a few thousand rows pins it down. Finding 01 therefore reports only that the field covers a minority. Registries differ on what they admit, how they dedupe, and whether they keep unreachable entries, so the honest reading of 8,433 is a floor for what our ingestion has found and verified. The findings below are properties of that population. We would expect the shape of them, an absent install metric and a very long tail, to hold more widely, but this census does not prove that.

Four further limits shape what can be said. First, install counts are absent across the entire population, which is Finding 01 rather than a gap we worked around, and it is why no ranking here claims to measure usage. Second, stars are a property of the repository: 14.4% of repo-backed servers share a repository with another server and inherit its full count, and 20.2% of servers have no repository at all and sit outside every repository-level figure. Where a total is affected we give both the naive and the deduplicated number. Third, declared tool counts exist for only 13.2% of servers, so every tool figure states that denominator rather than projecting onto the whole census. Fourth, the 70.3% of servers carrying a push date are not a random sample of the rest, as Finding 04 sets out, so maintenance shares are quoted against both denominators. Day counts use whole-day flooring against 2026-08-14.

Skillselion is an independent project and is not affiliated with Anthropic, OpenAI, Cursor, Claude, Claude Code, Codex, or the Model Context Protocol project.

Data

Download the dataset

The full census runs 8,433 rows, one per MCP server, with publisher, repository, category, stars, declared tool count, last push date, quality tier, deployment model and pricing fields. It is free to use with attribution under CC BY 4.0: cite “Skillselion MCP Server Census, August 2026” and link this page.

Download the census CSV · 8,433 rows · 270 KB gzipped