We counted every publishing account in the Skillselion catalog on 2026-08-28: 5,690 accounts behind 59,685 agent skills and 162,061,967 registry installs. Authorship is wide, adoption is not. Eight accounts hold 50.05% of all installs, the top 1% hold 83.3%, and the author-install Gini coefficient is 0.982.
Disclosure: we run Skillselion, an independent directory of Claude Code, Codex and Cursor extensions. We are not affiliated with Anthropic, OpenAI or Cursor, and not with any account named below. The full study is The Agent Skill Author Census, published 2026-08-28.
How concentrated is the AI agent skill ecosystem?
Very. We sorted all 5,690 accounts by cumulative installs and added from the top. Eight accounts reach 50.05%. Ten reach 54.6%. The top 1%, which is 57 accounts, reaches 83.3%, and 120 accounts (2.1% of all publishers) cover 90% of every install recorded in the catalog on 2026-08-28.
The Gini coefficient of that distribution is 0.982. For reference, a Gini of 1.0 is one account holding everything.
If you have shipped a skill and watched the install counter sit still, you are in the majority of accounts in this census. That is a fact about the distribution. It is not a diagnosis of your skill.
Which accounts hold half of all installs?
Frozen to the 2026-08-28 census, ranked by share of the 162,061,967 total installs. The first eight rows are the ones that add to 50.05%. Ranks 9 and 10 are shown for context and take the running total to 54.6%.
| Rank | Account | Skills | Install share |
|---|---|---|---|
| 1 | open.feishu.cn | 28 | 10.19% |
| 2 | mattpocock | 45 | 9.46% |
| 3 | microsoft | 186 | 7.67% |
| 4 | prime-skills | 60 | 6.71% |
| 5 | larksuite | 42 | 6.57% |
| 6 | vercel-labs | 117 | 3.8% |
| 7 | heygen-com | 44 | 3.2% |
| 8 | lllllllama | 45 | 2.46% |
| 9 | anthropics | 336 | 2.37% |
| 10 | Leonxlnx | 20 | 2.18% |
Source: Skillselion Agent Skill Author Census, August 2026, Finding 01.
Several of those names are already the ones developers hit first when they browse. Microsoft's Azure work shows up as the microsoft-foundry skill, Vercel Labs ships the agent-browser skill, Anthropic's own account ships the frontend-design skill, and rank 2 belongs to one person: Matt Pocock's grill-me skill sits in the same account that holds 9.46% of the 162,061,967 installs in this census. His public repos are at github.com/mattpocock.
One caveat before anyone reads the table as a quality ranking. An account here is a publishing account, so an organization counts as one author. Case-folding GitHub owner names merged 34 casing variants and brought 5,724 raw spellings down to 5,690 authors.
Do official vendor accounts explain the concentration?
Partly, and less than the table suggests. Each listing carries a flag for whether the publisher is the vendor of the product the skill drives. On 2026-08-28 that was 5,576 skills, 9.3% of the census, carrying 41,384,836 installs, or 25.5%.
So one skill in eleven is official and one install in four lands on one. Vendors do punch above their listing share. But 74.5% of installs still go to community accounts, which means the concentration is not simply big companies winning. Individual accounts are stacked at the top too.
We wrote more about that split in our guide on whether vendors are replacing Claude Code skills.
Does publishing more skills earn more installs?
No. Of the five findings, this is the one we would most want a first-time publisher to read.
The most prolific account in the catalog, affaan-m, published 544 skills. They average 5,392 installs each and sum to 1.81% of the census. mattpocock published 45 skills that average 340,684 installs each and sum to 9.46%. One account out-published the other twelve to one. The other out-earned it five to one on installs.
Publishing volume buys shelf space, and little else.
Source: Skillselion Agent Skill Author Census, August 2026, "Finding 04".
The pattern holds further down the prolific list. sickn33 published 445 skills at 1,303 installs per skill. TheBushidoCollective published 387 at 47 installs per skill. Neither registers in install share.
None of that is a judgement on the work. The census measures installs per publishing account and nothing about quality. See affaan-m's repositories on GitHub.
What should you do if you publish agent skills?
Three things follow directly from the numbers.
- Ship one skill you would use every day rather than a catalog. Dividing the two published averages gives an install-per-skill ratio of 63 to 1 between rank 2 and the most prolific account. That ratio is a derived figure, not a measured one, and it describes those two accounts rather than a rule about yours.
- Read the top of your own category before you write. The AI and agents category and the install leaderboard show what already has traction, which is also what you have to be visibly better than.
- Expect a slow start. The median author in this census has zero installs, and 52.7% of the 5,690 accounts have never recorded one.
How was this counted?
We dumped all 82,157 catalog listings on 2026-08-28 and aggregated the 59,685 skill listings by publishing account. MCP servers, plugins and marketplaces are excluded because they carry no comparable install count. Installs count registry install events, not active users, so read them as distribution rather than daily usage.
The catalog itself is a live catalog, refreshed daily from skills.sh, GitHub and MCP registries, ranked by real installs. The counting rules and their limits are written up in How Skillselion Research is done.
Every number above recomputes from the author census dataset, which has one row per publishing account with skill count, cumulative installs, install share and the official-vendor columns. It is CC BY 4.0. Cite "Skillselion Agent Skill Author Census, August 2026" and you are done.
If you want the market-shape view instead of the author view, that is The Agent Economy Census, and the running totals are on the AI agent statistics page. Reading a table is slower than browsing, and the agent skills catalog is the same data with filters on it.
FAQ
Who builds AI agent skills?
5,690 publishing accounts, counted as case-folded GitHub owners and registry publishers, authored the 59,685 agent skills in the Skillselion catalog as of 2026-08-28. An organization counts as one account, so this measures concentration across publishers rather than across individual people.
How many installs does the average skill author have?
The arithmetic mean is about 28,481 installs per account, which no account near the middle resembles. The median is zero. 2,998 accounts (52.7%) have no recorded install on anything they published, and 3,539 accounts (62.2%) published exactly one skill.
Is 0.982 a normal Gini coefficient for a developer ecosystem?
It is extreme on its face: 0.982 sits very close to the 1.0 ceiling, where a single account would hold everything. We measured it inside one catalog on one date, and we have not run the same computation on a package registry, so treat it as the shape of agent skill adoption on 2026-08-28 rather than a cross-ecosystem comparison.
Can I reuse these numbers?
Yes, under CC BY 4.0. Cite "Skillselion Agent Skill Author Census, August 2026" and link the study. The dataset download is on the same page.
Last updated: 2026-09-10. Census figures frozen to 2026-08-28.
Top comments (0)