skills.ecosyste.ms - agent skill indexing and security service

A dedicated service for agent skills (the SKILL.md-format bundles consumed by Claude Code, Cursor, Copilot, Windsurf and friends), structured like packages.ecosyste.ms: pull skill metadata from multiple registries and GitHub-hosted collections, normalise it, and surface dependency and security data across them.

Skills end up looking very similar to packages: they have names, versions (sometimes), authors, repositories, install URLs, and they depend on packages from other ecosystems (npm/pypi scripts invoked from inside a skill). No one is indexing this surface as a single dataset today.

Registries and sources worth pulling from:

- `skills.sh` - has a public JSON API at `skills.sh/api/v1/` (list, search, curated, detail, plus a per-skill `/audit/` endpoint backed by partner security scans). Source field is `owner/repo` on GitHub.
- `clawhub.ai` (OpenClaw) - documented HTTP API, tracks semver versions, dist-tags, changelogs, download counts, and runs automated security scans on publish.
- `github.com/anthropics/skills` and `anthropics/claude-plugins-official` - official marketplaces shipped as `.claude-plugin/marketplace.json`.
- The various `awesome-*` lists (VoltAgent, travisvn, GetBindu, sickn33) - long tail of GitHub-hosted skills, same shape as how packages.ecosyste.ms picks up the GitHub side of other ecosystems.

What this service could surface that the individual registries do not:

- Cross-registry deduplication. Skills get republished across skills.sh, ClawHub, awesome lists and direct GitHub repos under different slugs.
- Dependency edges into other ecosystems. SKILL.md scripts shell out to npm/pip/etc; parsing those gives a real graph from skills into packages.ecosyste.ms.
- A cross-registry vulnerability view. Both skills.sh and ClawHub do their own scanning in isolation. Advisories.ecosyste.ms could consume a normalised skill identity.
- Repository, commit, license and digest data via the existing services once skills have a stable identifier.

Discovery and security problems specific to this space: skills execute arbitrary code in a developer agent context, the registries are new and inconsistently vetted, versioning is largely absent (the PURL discussion notes zero version fields across 93 sampled skills from Anthropic/Microsoft/OpenClaw), and reuse across registries makes "is this the skill I think it is" hard to answer.

Worth tracking the PURL standardisation effort: https://github.com/package-url/purl-spec/issues/867 proposes `pkg:agentskills` with publisher/owner namespace, SKILL.md `name` as the package name, and qualifiers for `repository_url`, `scope`, `client`, `install_path`. Discussion is open and contested (no canonical registry, weak versions) but SBOM tooling is already treating these as packages, which is the same argument for indexing them here.