Skillselion Research · Edition of 2026-10-30
The Agent Skill Dependency Census
What do AI agent skills need before they run? On 2026-10-30 we read the prerequisite list the Skillselion catalog stores for each skill. 16,798 of 64,756 skills (25.9%) have one, and those skills hold 65.7% of all skill installs. Among them, 18.7% name an account, login or API key, 12.8% name a command-line tool, 9.8% name Python and 7.9% name a JavaScript runtime. Separately, 1,034 skills name another skill that has to run first.
We dumped all 87,443 catalog listings, kept the 64,756 skills with their 236,629,160 installs, and sorted the 27,561 prerequisite entries on 16,798 skills into nine keyword classes. A model wrote those entries while reading each skill's SKILL.md or an excerpt of it, so a skill without a list has not been shown to need nothing. Every number except the 87,443 listing total recomputes from the dataset, which ships with the next site release; the rules and their limits are in the methodology.
Finding 01
How many AI agent skills have a stored prerequisite list?
16,798 of the 64,756 skills in the catalog (25.9%) carry a prerequisite list, and they hold 155,552,276 installs, 65.7% of all skill installs. The lists are short by design: the tagging pass that writes them asks the model for one to three entries, so 8,250 skills list one prerequisite, 6,333 list two and 2,215 list three. A skill at the cap may need more. The 27,561 entries are free text and mostly distinct: 22,944 distinct entries remain after case-folding (83.2%).
Coverage leans toward popular skills. The median skill with a list has 761.5 installs and the median skill without one has 48, so the shares on this page describe the more-installed part of the catalog. A suite installed as a bundle records one install per skill, so install shares here and in Finding 02 can overstate distinct installs. The other 47,958 skills have no entry stored. 13,694 of them carry a tagging timestamp, meaning a tagging pass ran and stored none; an empty list is not evidence either way.
Finding 02
How many agent skills name a third-party account or login as setup?
3,134 of the 16,798 skills that list prerequisites (18.7%) name an account, API key, access token, credential, login, subscription or secret. They come from 600 publishing accounts and hold 16.9% of installs on skills that list prerequisites; a suite installed as a bundle records one install per skill, so this share can overstate distinct installs. Entries include "Deepgram API key" and "Auth0 account and application". Entries like these point to a third-party sign-up or key that the skill needs; the methodology gives an example of an entry the keyword rule catches by mistake.
The share varies by category. Among the 14 categories with at least 500 skills that list prerequisites, it is highest in Automation & Workflows (395 of 955, 41.4%) and lowest in Documentation (17 of 595, 2.9%). This class is wider than the bring-your-own-key label in the Pricing Census: it reads a different field and also counts accounts and logins, so the two shares measure different things. That census also covers which API keys skills ask for. The Security Census counted API credentials in a separate permissions field, which is also a different measure.
Finding 03
Do agent skills need Python or Node.js?
More skills name Python, but skills that name a JavaScript runtime have more installs. 1,650 skills that list prerequisites (9.8%) name Python or a Python tool such as pip, uv or pandas, and they hold 6,341,435 installs (4.1% of installs on skills that list prerequisites). 1,322 (7.9%) name Node.js, npm, npx, pnpm, Yarn, Bun or Deno, and they hold 16,770,856 installs (10.8%); vercel-labs skills hold 32.5% of those installs, and a suite installed as a bundle records one install per skill. Only 66 skills name both.
Git or GitHub appears in 826 lists (4.9%), a container tool such as Docker or Kubernetes in 279 (1.7%), a cloud platform (AWS, Google Cloud, Azure, Cloudflare, Vercel, Supabase or Firebase) in 529 (3.1%), and FFmpeg, ImageMagick or yt-dlp in 119 (0.7%). 8,980 lists (53.5%) match none of the nine classes. Entries in that group include "Existing PostgreSQL or MySQL database schema" and "Spring Boot JPA project".
- Account, key or login (26,344,527 installs)3,134 skills
- Command-line tool (44,728,646 installs)2,142 skills
- Python (6,341,435 installs)1,650 skills
- JavaScript runtime (16,770,856 installs)1,322 skills
- Git or GitHub (4,306,662 installs)826 skills
- MCP server (5,347,915 installs)790 skills
- Cloud platform (16,877,125 installs)529 skills
- Container tool (1,271,148 installs)279 skills
- Media tool (1,453,926 installs)119 skills
Finding 04
Which command-line tools do agent skills depend on?
2,142 skills that list prerequisites (12.8%) name a command-line tool, and they hold 28.8% of installs on skills that list prerequisites. One vendor CLI carries about half of that weight. Lark CLI appears in 46 prerequisite lists that hold 23,659,918 installs, 15.2% of installs on skills that list prerequisites and 52.9% of installs on the skills that name any command-line tool (45.0% when catalog-marked copies are removed from both); 43 of those skills, with 23,659,902 installs, come from two accounts with Lark or Feishu names, larksuite and open.feishu.cn, and carry 29 distinct skill names, because 14 names appear under both accounts. The catalog marks 15 of the 28 larksuite listings that name Lark CLI as copies of open.feishu.cn listings; without the 16 catalog-marked copies among the 46, skills that name Lark CLI hold 17,083,071 installs, 11.6% of installs on skills that list prerequisites once copies are removed from both. Skills in suites like these can be installed as a bundle, which records one install per skill, so their install share can overstate how many developers installed them. The gws CLI appears in 107 lists, all from the googleworkspace account (for example gws-gmail), with 3,189,373 installs.
Other tools spread across many publishers. The GitHub CLI (gh) appears in 208 lists from 106 accounts, Claude Code in 381 from 147, Playwright in 123 from 82, Docker in 156 from 66 and FFmpeg in 99 from 52.
Finding 05
Do agent skills depend on other skills or MCP servers?
Some do. 1,034 skills (1.6% of all skills) name another skill that has to run first: 876 name one and 158 name two, 1,192 links in all, from 274 publishing accounts. They hold 15.0% of all skill installs, and three accounts carry 54.1% of that: larksuite, mattpocock and heygen-com. A suite installed as a bundle records one install per skill, so this share can overstate distinct installs. The skill with the most installs in this group is hyperframes-cli from heygen-com (1,311,338 installs), which points to the hyperframes skill in the same repository. The most-named required skills are belt-sh-cli (27 skills, all from inference-sh), gws-gmail (21, googleworkspace) and longbridge (20, longbridge). Each of these three is named only by skills from a single publishing account. 8 of the 27 belt-sh-cli listings are catalog-marked copies; without copies, gws-gmail (21) ranks first, then longbridge (20) and belt-sh-cli (19).
MCP appears in 790 prerequisite lists (4.7% of skills that list prerequisites) from 232 accounts. Some name an MCP server to connect before the skill runs, such as "Chrome DevTools MCP server connected". Others offer MCP as one option, such as "Longbridge CLI or MCP access".
Method
Methodology
The census is a complete dump of the Skillselion production catalog taken on 2026-10-30: 87,443 listings in total, of which the 64,756 skill listings form this census, with 236,629,160 cumulative installs aggregated from public registry data, chiefly the skills.sh registry. Installs count registry install events, not active users. We pulled the full record of every skill with parallel paged requests and matched the pulled ids one for one against the catalog index. Between the catalog index (the installs_index column) and the detail pull, install counts differed on 2,039 skills (median difference 4 installs, largest 3,597), a net 64,301 installs or 0.03% of the total; the dataset keeps the detail-pull reading. The census counts every skill listing, including the 4,136 listings (6.4%) that the catalog marks as copies of another listing (the duplicate_of column), so counts on this page are counts of listings.
Two catalog fields carry the data. prerequisites holds up to three short setup entries, and requiresSkills holds the slugs of skills that must run first, either as a bare name or as a long form skill-<author>-<repo>-<name>; when ranking required skills we count a long form that matches a catalog skill as that skill's name (230 of 368 long forms match). A slug equal to the skill's own name or long form is a self-reference and is not counted; 4 skills list only themselves and are left out. A language model writes both during catalog tagging, from the skill's SKILL.md text (in some tagging passes, an excerpt of it). They are a model's reading of the documentation, not a parsed package manifest, so an entry can be incomplete or worded loosely, and an empty field means no entry was written. 7,701 skills carry a list but no tagging timestamp, so the timestamp undercounts tagging passes. The skill tagging prompt gives example entries (git repo, spec doc, node installed), which can raise the Git and JavaScript classes. Every share in Findings 02 to 04 is a share of the 16,798 skills that list prerequisites unless it says otherwise.
Classes are keyword matches, case-insensitive, on each entry. A skill joins a class if any entry matches, so one skill can join several. Account, key or login: api key, *_KEY, api/access/auth/bearer/bot/OAuth/GitHub/personal access token, *_TOKEN, credential, account, subscription, OAuth, login, logged in, authenticated, secret, access key. Python: python, pip, uv, conda, poetry, pandas, numpy, pytorch, jupyter. JavaScript runtime: node, node.js, npm, npx, pnpm, yarn, bun, deno. Git or GitHub: git (not after a slash, so api/git is skipped), github, gh, gitlab. Container tool: docker, podman, kubernetes, kubectl, k8s, container, helm. MCP server: mcp, model context protocol. Command-line tool: cli, command line. Cloud platform: aws, amazon web services, gcp, google cloud, azure, cloudflare, vercel, supabase, firebase. Media tool: ffmpeg, imagemagick, yt-dlp. Named tools in Finding 04 use their own patterns (for example lark-cli or lark cli; gws not followed by a hyphen; gh or github cli; claude code with a space). Keyword matching makes mistakes in both directions: "Account size and sizing constraints" joins the account class, entries such as "gh-pages-deploy skill" count as the GitHub CLI, and a key named in other words is missed. The dataset carries every entry so anyone can re-sort them. A publishing account is the catalog's author field for the listing, case-folded. Install counts are registry counts, and a suite installed as a bundle can record one install per skill. Skillselion is not affiliated with any publisher named on this page.
Data
The dataset
The dataset has 64,756 rows, one per skill listing, with the publishing account, category, installs, the stored prerequisite entries, a flag for each of the nine classes, the required-skill slugs, the catalog's copy marker (duplicate_of), the tagging flag (tagged), the vendor flag (official_vendor) and the install count read from the catalog index (installs_index). It will be free to use with attribution under CC BY 4.0: cite "Skillselion Agent Skill Dependency Census, October 2026" and link this page.
The file ships with the next Skillselion site release, and this section will carry the download link once the file is served.
FAQ
Questions journalists ask
What do AI agent skills need to run?
Of the 64,756 skills in the Skillselion catalog on 2026-10-30, 16,798 (25.9%) carry a prerequisite list. Among those, 18.7% name an account, API key, token, login or subscription, 12.8% name a command-line tool, 9.8% name Python and 7.9% name a JavaScript runtime such as Node.js. The other 47,958 skills have no entry stored, which is not evidence either way.
How many agent skills list a third-party account or login as setup?
3,134 of the 16,798 skills with a prerequisite list (18.7%) name an account, API key, access token, credential, login, subscription or secret, on 2026-10-30. The share is highest in Automation & Workflows (41.4%) and lowest in Documentation (2.9%) among categories with at least 500 skills that list prerequisites.
Do agent skills depend on other skills?
Some do. 1,034 of 64,756 skills (1.6%) name another skill that has to run first, and they hold 15.0% of all skill installs. The most-named required skills are belt-sh-cli (named by 27 inference-sh listings, 8 of them catalog-marked copies) and gws-gmail (named by 21 googleworkspace listings).
Can I reuse these numbers or the dataset?
Yes. Cite "Skillselion Agent Skill Dependency Census, October 2026" and link this page. The raw dataset (64,756 rows, one per skill listing, with every stored prerequisite entry and its keyword classes) will be released under CC BY 4.0 with the next Skillselion site release, together with a script that recomputes every figure from it.