SkillMD is an open skills registry at skillmd.com that indexes public SKILL.md files at scale and differentiates on published verification: per-skill lint verdicts, capability flags, and third-party scanner results (NVIDIA SkillSpector and Cisco AI Defense Skill Scanner) shown on each skill page, installable through its own MIT CLI, MCP server, Claude Code plugin marketplace, or GitHub Action.
The verification-first registry now competes with skills.sh at scale, and its verified-core evidence just went opaque: the reviewed count it displayed through 2026-11-17 (last reading 2,233) is gone, and the hero now labels the whole 1,335,850-skill index safety-reviewed as of 2026-11-19.
What it is #
The registry (skillmd.com) crawls and accepts SKILL.md files, and its hero stat has stopped separating the counts: through 2026-11-17 it reported 1,242,721 skills with a separate reviewed count that moved four times (most recently 2,233, labeled “verified publishers” on 2026-10-20, “safety-reviewed” on 2026-10-27, “verified publishers” again on 2026-11-07, “safety-reviewed” since 2026-11-12); as of 2026-11-19 the hero reads “1,335,850 safety-reviewed skills” and displays neither the reviewed count nor the author count.
The toolchain (github.com/skillmds/SkillMD, MIT, the repo’s display name since 2026-11-10) is what runs on your machine: a skillmd CLI whose canonical npm package is now get-skill (since v1.3.7 on 2026-11-08, ten versions there; the homepage install lines and the MCP command npx -y get-skill now use it, while the legacy skillmds package publishes in lockstep at v1.5.0, the 44th version, 3,418 downloads in the month to 2026-11-15), a Claude Code plugin marketplace, and a GitHub Action, plus a JSON search API and OAuth-protected endpoints aimed at agents.
The repo shipped its first release outside the CLI train on 2026-10-26, core-v1.0.1 (@skillmds/core on npm, two versions), landing a slugify fix (names can no longer end in a hyphen after the 60-character cut) and a marketplace refresh (24 plugins), then crossed back to the CLI train with v1.3.4 on 2026-10-28 (a companion gone upstream no longer fails the install) and v1.3.5 on 2026-11-03 (the agent picker no longer claims Claude Code already has the plugin).
Six more shipped in the three days to 2026-11-10: v1.3.6 through v1.3.11, the batch that renamed the package to get-skill (v1.3.7), made every registry skill install from its bundle (v1.3.8), rebranded the prompts and README wordmark (v1.3.9), renamed find-skills to find-a-skill (v1.3.10), and pointed repository links at the renamed repo (v1.3.11).
v1.3.12 shipped 2026-11-15, teaching --ref and the registry fallbacks to read a skill’s actual source repository, because a registry slug is not a GitHub repo (the antfu/vite skill lives in the antfu/skills repository, and the old lookup failed with a 404).
v1.3.13 shipped 2026-11-16, moving the login into the browser (skillmd login prints a short code you approve at skillmd.com) and teaching logout to revoke the token on the server before removing it locally.
v1.4.0 shipped 2026-11-17, making plugin installs two requests instead of dozens (manifest plus archive, sha256 checks, six-at-a-time companions), retrying busy-registry requests on 429/502/503/504 with Retry-After, refusing to install a pack as its SKILL.md alone when members fail to fetch, and summarizing multi-skill output.
v1.5.0 shipped 2026-11-18 and made dependency-following the install contract: add asks the registry’s /deps endpoint (three levels deep, all four edge kinds), installs skill_tool and invoke dependencies with related: lines and --with-related, carries --no-deps for CI, and gives the stdio MCP tool the same resolver, replacing the v1.4.0 pointer detector with the registry’s own extractSkillRefs.
Skills land in each detected agent’s own directory (.claude/skills/, .cursor/skills/, .agents/skills/), and every listed skill is pinned to a commit.
The maintainer also publishes the ecosystem’s most self-critical data: their directory-landscape post estimates 83.4 percent of indexed skills declare no license and roughly 47 percent of public SKILL.md files are byte-identical copies of another file.
Status #
Active and growing, with a thin independent footprint.
The toolchain repo was created 2026-08-25 and shows 200 stars as of 2026-10-13 with 8 open issues, pushed 2026-11-18; the npm CLI has shipped 44 versions on the legacy skillmds name (plus ten on get-skill) since 2026-06-29.
The site’s own claimed index grew from 860,000 skills (its September 19 directory post) to 1,242,721 as of 2026-11-17, and that post itself warns that directory counts are claims, not measurements, skills.sh included. No Hacker News thread surfaced under its name in my searches as of 2026-10-05; the largest near-match is the 48-point “Skill.md: An open standard” story about the format, not this registry. The business model is unstated; installing needs no account and the site quotes no price.
Strengths #
- It is the only directory of its size that publishes what it checked and how, per skill, instead of a claimed security stance.
- Install paths that go beyond a copy button: CLI, MCP server, plugin marketplace, and CI action, all writing into the agents’ native skills directories.
- Agent-facing surface: JSON API, OpenAPI spec,
.well-knownagent-skills index, and an MCP server card, so an agent can search and install without a human. - The self-published ecosystem data (license gaps, duplication rates) is decision-useful even for people who never install from it.
Cautions #
- The verified-core evidence is now invisible: through 2026-11-17 the same page that claimed a pre-publication safety review also displayed 2,233 reviewed out of 1,242,721 listed, and since 2026-11-19 the hero labels all 1,335,850 indexed skills safety-reviewed with no reviewed count shown, so a reader can no longer see how much of the catalog is actually checked.
- The toolchain repo’s 200 stars and the near-zero HN footprint mean no independent security eyes on the verification pipeline itself.
- A verdict computed from the files alone cannot see runtime-fetched payloads: PromptArmor’s backdoored-skill research (fetched 2026-10-07, re-verified live 2026-10-11) demonstrated skills that passed Anthropic’s own scanner by retrieving malicious payloads at runtime, a class-level caution for the whole scanner-verdict layer this registry’s badge is built on.
- Its blog states Agent Skills is stewarded through the Agentic AI Foundation, but the AAIF’s own project list carries AGENTS.md, MCP, goose, and agentgateway, not Agent Skills, so treat the blog’s ecosystem claims as unverified.
- Index counts are self-reported and the site says so itself; the 860k-to-1.13M jump in two weeks is crawl churn as much as ecosystem growth.
- A registry this young is one team’s roadmap; skills.sh at least has Vercel behind it.
Pricing #
Free. No account is needed to search or install, the toolchain is MIT, and the site quotes no paid tier as of 2026-10-05.
Compared to #
- skills.sh: the telemetry-ranked incumbent; SkillMD trades usage signal for published verification, and both index the same public GitHub repositories.
- skillregistry.io: the hosted upload registry positioning itself as definitive; tiny, no published method, and orders of magnitude smaller than either.
- Skilleton: the lockfile-based, telemetry-free manager; SkillMD is the hosted-registry version of the same verification-first instinct.
Bottom line #
Recommended as a second opinion before installing an unfamiliar skill: search it there, read the lint and scanner verdicts, then pin the commit. Not as a replacement for reading the SKILL.md yourself, and not because its catalog is bigger or safer at scale, because neither is demonstrated. My disagreeable claim: SkillMD’s published verdicts (however many skills they cover, a number the site stopped showing on 2026-11-19) matter more than either registry’s million-skill index, and the winner of this layer is whoever verifies the long tail, not whoever crawls it.
Changes #
- 2026-10-05 - Created in the daily refresh’s skills entrant scan.
- 2026-10-07 - Added the skillmds/skillmd star history chart to the Status section.
- 2026-10-07 - Added PromptArmor’s backdoored-skill scanner-bypass research as a caution: verdicts computed from the files alone cannot see runtime-fetched payloads; the index count refreshed to 1,132,704 with 1,948 reviewed unchanged.
- 2026-10-10 - The toolchain repo moved from 1 to 58 stars in three days (pushed 2026-10-09), still orders of magnitude below any peer, while the index inched to 1,132,705 skills from 30,613 authors with the 1,948 reviewed count unchanged.
- 2026-10-11 - The toolchain repo climbed again to 89 stars, still orders of magnitude below any peer, while the index inched to 1,132,706 skills from 30,612 authors with the 1,948 reviewed count unchanged and the PromptArmor research re-verified live.
- 2026-10-13 - The toolchain repo doubled again to 200 stars in two days, still orders of magnitude below any peer, while the index inched to 1,132,712 skills from 30,616 authors with the 1,948 reviewed count unchanged.
- 2026-10-22 - The index inched to 1,132,721 skills from 30,620 authors with the 1,948 reviewed count unchanged; the hero label reads “safety-reviewed” again after the 2026-10-20 “verified publishers” reading, and the npm CLI re-verified at 1.3.3 with 2,017 downloads in the month to 2026-10-18.
- 2026-10-27 - The reviewed count moved for the first time: 1,947 safety-reviewed (down one from the 1,948 held since at least 2026-10-13) against an index of 1,132,729; the repo shipped core-v1.0.1 (2026-10-26, @skillmds/core on npm with a slugify fix and a 24-plugin marketplace refresh), and the toolchain repo moved to 9 open issues, pushed 2026-10-26.
- 2026-10-31 - The CLI train crossed back: v1.3.4 shipped 2026-10-28 (a companion gone upstream no longer fails the install), the npm package’s 34th version; the index inched to 1,132,737 skills with the reviewed count holding at 1,947 and the author count down to 30,538 from 30,620.
- 2026-11-07 - Volatile refresh: v1.3.5 shipped 2026-11-03 (the agent picker no longer claims Claude Code already has the plugin), the npm package’s 35th version with 1,951 downloads in the month window 2026-10-08 to 11-06; the index inched to 1,132,752 skills with the reviewed count holding at 1,947, the hero label reads “verified publishers” again (its third state), and the author count is no longer displayed in the hero stats; the repo moved to 8 open issues, pushed 2026-11-03.
- 2026-11-12 - Volatile refresh: six CLI releases in three days (v1.3.6 through v1.3.11, 2026-11-08 to 11-10) made
get-skillthe canonical npm package (the homepage install lines and the MCP command now use it, the legacyskillmdspackage publishing in lockstep at 1.3.11) and renamed the repo to skillmds/SkillMD; the reviewed count moved for the second time (1,947 to 2,204) against an index of 1,146,032, the hero label back on “safety-reviewed” with the author count displayed again (30,604), and the repo pushed 2026-11-10. - 2026-11-16 - Volatile refresh: v1.3.12 shipped 2026-11-15 (
--refnow reads a registry skill’s actual source repository, since a registry slug is not a GitHub repo;get-skillat seven versions,skillmdsin lockstep at its 41st), the reviewed count moved for the third time to 2,229 against an index of 1,217,421 from 32,805 authors, and the repo pushed 2026-11-15 with 9 open issues. - 2026-11-17 - Volatile refresh: v1.3.13 shipped 2026-11-16 (browser-based
skillmd loginwith a short code approved at skillmd.com, andlogoutrevoking the token server-side;get-skillat eight versions,skillmdsin lockstep at its 42nd), the reviewed count moved for the fourth time to 2,233 against an index of 1,242,721 from 34,309 authors, and the repo pushed 2026-11-17 with 8 open issues. - 2026-11-19 - Volatile refresh: v1.4.0 shipped 2026-11-17 (two-request plugin installs with sha256 checks, Retry-After backoff, all-or-nothing member installs) and v1.5.0 shipped 2026-11-18 (registry-backed dependency installation, three levels of
skill_toolandinvokeedges,related:lines with--with-relatedand--no-deps, MCP tool parity;get-skillat ten versions,skillmdsin lockstep at its 44th), the repo pushed 2026-11-18 with 8 open issues, and the registry’s hero claim changed: the reviewed count (last displayed at 2,233) and the author count are gone, and the hero now reads “1,335,850 safety-reviewed skills”, so the verified-subset size is no longer publicly checkable.
See also #
- skills.sh - the telemetry-ranked incumbent registry
- Agent Skills open standard - the format every directory indexes
- Anthropic Agent Skills - the vendor pack and partner directory
- Agentic Coding Tools Landscape - the map this distribution layer sits over
References #
https://skillmd.com/ - the registry surface: the hero now reads “1,335,850 safety-reviewed skills” with the reviewed count (last displayed at 2,233 of 1,242,721 on 2026-11-17) and the author count no longer displayed; agents 60+, install lines naming the
get-skillpackage (fetched 2026-11-16 and re-fetched 2026-11-17 under the old display; re-fetched 2026-11-19 under the new hero)https://github.com/skillmds/skillmd - the toolchain repo: MIT, 200 stars as of 2026-10-13, 8 open issues, created 2026-08-25, pushed 2026-11-18, display name now SkillMD with repository links updated in v1.3.11 (GitHub API, as of 2026-11-19)
https://api.github.com/repos/skillmds/skillmd/releases - the release train: v1.5.0 published 2026-11-18 (registry-backed dependency installation,
--with-related, MCP tool parity), following v1.4.0 published 2026-11-17 (two-request plugin installs, Retry-After backoff, all-or-nothing member installs), v1.3.13 published 2026-11-16 (browser-based login and the server-side logout revoke), v1.3.12 published 2026-11-15 (--refreading the skill’s actual source repository), v1.3.11 published 2026-11-10 (repository links pointing at the renamed repo), v1.3.10 (find-a-skill rename) and v1.3.9 (11-09), v1.3.8 (bundle installs, 11-09), v1.3.7 (the package is get-skill, 11-08), v1.3.6 (11-08), v1.3.5 published 2026-11-03, v1.3.4 published 2026-10-28, core-v1.0.1 published 2026-10-26 (@skillmds/[email protected]) (fetched 2026-11-16; re-fetched 2026-11-17 and 2026-11-19)https://registry.npmjs.org/get-skill - the canonical CLI package since v1.3.7: latest 1.5.0, ten versions, created 2026-09-21 (fetched 2026-11-16; re-fetched 2026-11-19)
https://registry.npmjs.org/@skillmds/core - the new core package: 1.0.1 latest, two versions (fetched 2026-10-27)
https://raw.githubusercontent.com/skillmds/skillmd/main/README.md - CLI, MCP server, plugin marketplace, GitHub Action, per-agent install directories
https://skillmd.com/blog/agent-skills-directories-compared - the self-critical directory landscape post: 47 percent byte-identical duplication, 83.4 percent license gaps, counts-as-claims disclaimer (September 19, 2026, fetched 2026-10-05)
https://registry.npmjs.org/skillmds - the legacy CLI package, publishing in lockstep: latest 1.5.0, 44 versions, created 2026-06-29 (fetched 2026-11-16; re-fetched 2026-11-19)
https://api.npmjs.org/downloads/point/last-month/skillmds - 3,418 downloads, window 2026-10-17 to 2026-11-15 (re-fetched 2026-11-17, the window not yet rolled forward and the figure unchanged); get-skill read 1,633 in the same window (fetched 2026-11-16)
https://hn.algolia.com/api/v1/search?query=skillmd&tags=story - the footprint scan behind the missing-footprint statement (fetched 2026-10-05)
https://www.promptarmor.com/resources/anthropics-skill-scanner-defeated-by-backdoored-skills - the backdoored-skill bypass behind the scanner-verdict caution: skills that passed Anthropic’s scanner by fetching payloads at runtime, and exfiltrated files in production (fetched 2026-10-07, re-verified live 2026-10-11)