How we curate
Methodology
Every signal on AgenticSkills is either verifiable against a primary source (GitHub) or labeled as editorial judgment. We do not fabricate metrics. This page is the long-form version of that promise.
What we show on every skill
Every skill page surfaces four kinds of signal. Each one has a primary source you can verify yourself.
Per-path last commit
Fetched from the GitHub commits API filtered to the skill's specific subdirectory. Solves the monorepo problem: anthropics/skills contains dozens of skills with different commit histories, and we show each one's actual freshness.
Source: GitHub commits API · path filter
Recent contributors
Unique author logins from the 100 most recent commits to the skill's path. Optimized for 'who is currently maintaining this' rather than lifetime contribution count.
Source: GitHub commits API · deduped authors
Stars · source repo
Stars on the GitHub repository the skill lives in. For monorepo skills this number is shared across all skills in that repo — we label it 'source repo' in the UI to make the scope clear.
Source: GitHub repo API · stargazers_count
Verified org badge
A binary indicator showing the GitHub org has been cross-checked against the company's official identity. Applies to the org, not to individual skill quality.
Source: Editorial — see Verified orgs section below
What we deliberately don't show
Install counts
Install counts cannot be independently measured by anyone outside the agent platform itself. Anthropic could tell you how many people installed a Claude skill; we can't. The skill ecosystem at large — GitHub, the major install CLIs (npx skills, openskills) — does not surface this number either, because nobody can.
We previously displayed install counts seeded from editorial judgment. We removed them entirely in May 2026 because the first question anyone reasonable would ask is "how do you know?" — and we did not have a credible answer. Stars, commits, and contributors are imperfect proxies for usage, but they are verifiable, and we'd rather show a smaller set of real numbers than a larger set of made-up ones.
If Anthropic or another platform ever publishes per-skill install telemetry through an official API, we'll integrate it and label it accordingly.
Editorial quality ranks
Quality ranks are editorial. They reflect our judgment about a skill's value to developers, not a popularity score derived from data. We label them as editorial in the UI so the distinction is clear.
Featured pick
Maintained by a verified org, exemplifies SKILL.md best practices, and (where applicable) appears in at least one workflow stack.
High quality
Clear documentation, recent maintenance, fits cleanly into common workflows. Solid choice in most contexts.
Functional
Works as advertised but with lower editorial confidence — narrow scope, less maintenance signal, or partial documentation.
Verified orgs program
A verified org has had its GitHub identity cross-checked against the company's official website, public communications, or staff GitHub profiles. The badge means "we believe this is really {Vercel Labs, Anthropic, Trail of Bits, etc.}" — not that every file in their repo is high quality.
Currently 118 skills come from verified orgs across the directory. See the full list at /authors.
If you maintain an org and want it verified, tell us through our contact form using an email address on a domain that matches your GitHub org's public identity.
For how this approach compares with the aggregators that index the same ecosystem, see AI agent skill directories.
How skills enter the catalog
- Discovery. Candidates come from GitHub topics (
agent-skills,claude-skills), platform marketplaces, and community signals (Reddit, X). Anyone can also submit at /submit. - Inclusion criteria. To be listed, a skill needs a published SKILL.md, a reachable source repo, a discernible scope, and an author identity that resolves to a real account. Candidates that fail these are not listed. There is no listing fee, and inclusion is never for sale — see commercial relationships below.
- Classification. Each entry carries a category, the platforms it targets, and a quality rank. Category and platform describe the skill; the rank is editorial judgment, and is labeled as such wherever it appears.
- Continuous refresh. Per-path commit data, contributor count, and repo stars are refreshed by an automated script (
scripts/refresh-skill-stats.ts) against the GitHub API. The freshness date in the navbar reflects the most recent refresh. - Delisting. A skill is removed if its source repo disappears, the author requests removal, or a credible security finding goes unaddressed. See /incidents for the public incident log.
What listing is not
Listing means an entry met the criteria above and was deliberately added — the catalog is selected rather than crawled. It does not mean every line of its SKILL.md has been read and verified. We are not claiming a line-by-line audit of 193 skills, because you would have no way to check that claim.
A skill is instructions an agent executes, so treat it like any other third-party code: read the SKILL.md before you install it. Where automated checks do run, they are described and dated — the security audit on each /mcp server page is the current example, and the skill-side equivalent is still being wired in.
What we don't list
We don't list tools whose main purpose is getting around a safeguard, or that let an agent move money or act on accounts without a limit a human set. In practice:
- Not listed: CAPTCHA solvers, and tools built to evade bot detection, spoof browser fingerprints, or remove AI watermarks.
- Not listed: tools that let an agent place trades, move funds, or log in to a bank account.
- Listed only on conditions: bulk messaging, if it sends through the platform's official API to people who opted in; tools an agent pays for per call in crypto (such as x402), if the price is published, you set up the wallet and a spending limit yourself, and the page says so; financial data, if it is read-only through a regulated, consent-based API.
The same rule applies to every submission, however popular the project is.
Commercial relationships
Sponsorship is sold at /advertise — featured partner placement, sponsored workflows, and launch campaigns. Everything paid is labeled “Sponsored” and looks different from editorial content. None of it touches the catalog. Every commercial relationship the site has, affiliate programs included, is listed at /disclosures.
What sponsorship does not buy:
- A spot in the directory. Listings are selected on the same criteria whether or not anyone paid.
- A better rank or trust signal. Stars, commits, license and author come from GitHub, not from us.
- A search ranking. A sponsored workflow is indexed like every other page here — whether Google ranks it is not ours to sell, and we would rather say so than imply otherwise.
- Unlabeled placement. Everything paid says “Sponsored” and looks different from editorial content.
- Our opinion. A sponsored workflow that doesn’t actually work gets fixed or pulled, not published.
- A page forever. A sponsored workflow runs for six months, not indefinitely; at the end of the term it is renewed or it comes down. It also comes down early if your product is sunset, or if the skills it depends on disappear and cannot be swapped — logged at /incidents, same as any other delisting.
Semantic search
The ⌘K command palette returns both keyword matches and semantic suggestions. Semantic suggestions surface skills whose meaning is close to your query even when the words don't overlap — type "deploy a Next.js app" and you'll get deployment skills back even if none of them say "Next.js" in the description.
Implementation: we embed every skill's document (name + description + category + tags + author) at build time using Voyage AI's voyage-3-lite model, store the vectors in lib/skill-embeddings.ts, and rank by cosine similarity against your query embedded at request time.
Privacy: when you type a query of 3+ characters into the palette, the query text is sent to Voyage AI to generate the request-time vector. Voyage doesn't see your IP or any account identifier from us. See /privacy for the full disclosure.
Watchers & auditors
We continuously monitor curated skills and MCP servers for security findings, source-repo disappearance, and license changes. The MCP audit framework currently surfaces results on every /mcp server page; the skill audit equivalent is being wired in.
Our model is "continuous monitoring of a curated set" rather than "one-time scan of every public skill." A directory that has scanned every skill once tells you about historical state. We tell you about current state on the skills that matter.
Found a problem?
If any number on the site looks wrong, or you suspect a skill should be delisted, tell us through our contact form or use the "Report this skill" button on the skill page. Public incident log: /incidents.