Disclosure: we run Skillselion, the directory this census was pulled from. Every number below is reproducible from the free CSV.
On 2026-08-09 we dumped the entire Skillselion production catalog and counted it: 79,827 public listings covering agent skills, MCP servers, plugins and marketplaces, holding 157,833,061 cumulative installs. Then we ran the statistic economists use for income inequality across the 58,751 of those listings that are agent skills.
The Gini coefficient came out at 0.96.
Gini measures how unevenly a quantity is spread across a population: 0 is perfectly even, 1 is one holder taking everything. Applied to install counts across 58,751 AI agent skills in August 2026, it is 0.96.
South Africa, which reports the highest national income Gini the World Bank publishes, sits near 0.63. Brazil is around 0.52. A 0.96 is not a country. It is closer to a raffle.
One honest caveat before you quote that. Income and software installs are different measurement domains, so read the comparison as a sense of scale and not a like-for-like ranking. A catalog where a tenth of entries have zero installs can reach a Gini that no national income distribution ever will. Long-tail software distributions run hot on this measure in general.
What 0.96 looks like from the inside
- Half of all 158 million installs go to 198 skills. That is 0.34% of the catalog.
- The median skill has 56 installs.
- 59.3% have fewer than a hundred installs. 9.9% have none at all.
- The top 1% of skills hold 77.4% of every install in the census.
The median skill sits at 56 installs, and these are cumulative all-time totals, so anything published this week starts from zero against a field that has been accumulating for months. For most people who publish a skill, release produces no measurable adoption. That is worth knowing before you read your own numbers as a verdict on the work.
The economy sitting on top of it has almost no money in it
Of the 79,827 listings, 55 carry a price. That is 0.07%. Another 1,404 ask you to bring your own API key and 184 run a freemium tier. Everything else is free.
Adoption on this scale normally arrives with a paid tier, a store cut or a marketplace fee. Almost none of that has been built here yet. If you are weighing whether to charge for a skill, that is the competitive context: you would be one of 55.
Publisher rank stops matching company size
Combining the two publisher tables from the census, where ByteDance is broken out separately for the reason below:
| Rank | Publisher | Installs |
|---|---|---|
| 1 | ByteDance Lark/Feishu (70 listings, bundled, see caveat) | 24,261,658 |
| 2 | Microsoft | 11,153,893 |
| 3 | Matt Pocock (independent educator) | 10,584,924 |
| 4 | Vercel Labs | 5,576,605 |
| 5 | Anthropic | 3,484,550 |
| 6 | 730,896 | |
| 7 | AWS | 171,446 |
| 8 | OpenAI | 150,665 |
One caveat we owe you on row 1. ByteDance's Lark/Feishu suite holds 15.4% of every install in the census, but Lark skills frequently install as bundles: one user action installs several skills at once, and the counts record each bundled skill separately. Those per-skill totals are therefore not comparable to a hand-picked install, and row 1 should not be read as 24 million deliberate choices. The share is real, and the mechanism producing it is documented in the methodology.
Now look at row 3. One person, publishing TypeScript education material, sits above Vercel Labs, Anthropic, Google and AWS, and roughly 70 times ahead of OpenAI on installs. Reach here tracks what developers choose to install, which is a different contest from benchmarks or funding.
The fastest-growing thing in the census is a skill that argues with you
One whole category exists to give the developer a hard time. Across 88 skills built for interrogation the cluster reaches 2,425,925 installs, about 1.5% of the census. What makes it notable is the slope: it is the platform's #1 rising search at +160% over the census window, and nothing else in the census grew faster.
| Skill | Installs | Repo |
|---|---|---|
| grill-me | 787,155 | mattpocock/skills |
| grill-with-docs | 669,499 | mattpocock/skills |
| triage | 534,688 | mattpocock/skills |
| grilling | 380,693 | mattpocock/skills |
Install counts above are frozen to the census date of 2026-08-09. The linked pages show live totals, which have moved since.
A second cluster sits on the meeting layer. Skills that write standup updates, summarize meetings, triage inboxes, tailor resumes and draft performance reviews reach 2,101,631 installs between them. Standups, meeting notes and status reports account for 1,889,988 of that across 111 skills. Every cluster here automates the reporting of work.
Method, and the raw file
The census is a complete dump of our production catalog taken on 2026-08-09: 79,827 listings, of which 58,751 are agent skills and the rest are MCP servers, plugins and marketplaces. Install counts are cumulative all-time totals. They aggregate public registry data, chiefly the skills.sh registry, alongside GitHub and other public sources, and they count install events rather than active users.
The category clusters were built by keyword matching over names and descriptions, then curated by hand to drop false positives. A "brutalist" UI-design skill is not an interrogation skill.
All five findings, the charts and the full methodology: The Agent Economy Census
The raw dataset, all 79,827 rows as a CSV, is downloadable on that page. Take it, check our arithmetic, disagree with us in public. Cite it as "Skillselion Agent Economy Census, August 2026".
We run Skillselion, an independent directory of Claude Code, Codex and Cursor extensions ranked by real installs. Not affiliated with Anthropic, OpenAI, Cursor, Claude, Claude Code or Codex.
Top comments (0)