Agent skills need a package registry, and nobody runs one yet

The idea that nobody runs a skills package registry yet doesn't hold up. Public registries already exist: skills.sh, SkillsMP and TanStack Intent. Sentry dotagents handles version pinning. JFrog, Chainguard and localskills.sh sell private, governed registries, and JFrog explicitly signs and scans skills on publish. Snyk, Mondoo, NVIDIA and several open-source projects already cover scanning, so the signed private registry idea runs straight into JFrog and Chainguard, which enterprise platform teams already buy from. The weakest spot in the market is enforcement and lineage, not storage. I found no product that reliably blocks Claude Code, Codex or Cursor from running unsigned skills at runtime, across every agent, on developer machines. Building that would mean hooking into each agent's skill loading, or using MDM and endpoint policy, and a registry by itself doesn't do it. Lineage is the other gap. StepSecurity already inventories the skills in repos and on laptops, but I saw nobody fingerprinting copied and modified skills across repos, tracing them back to the upstream version, and opening fix PRs on every copy. That is roughly Dependabot for vendored skills. So the provenance graph idea has a narrow opening if it is built around matching copies whose content has drifted from the original, which matters because skills are mostly copied and pasted rather than installed. The risk is that Snyk, StepSecurity, JFrog or GitHub add this as a feature. Trail of Bits has reportedly shown bypasses against current scanners, so better detection quality could also set a newcomer apart, but scanning alone is already a commodity.

• JFrog Agent Skills Registry and Skill Scanner

  Central registry for governing skills across agent tools. Publishing includes scanning and signing, and a beta Skill Scanner checks skill packages for malicious or suspicious behavior when they are uploaded. This overlaps directly with the signed private registry idea and comes from an incumbent that enterprises already use for artifacts.

• Chainguard Skills Registry

  Chainguard's registry installs skills into agent directories such as .claude/skills. Chainguard already sells trusted, provenance-focused supply chain products to enterprises.

• localskills.sh

  A hosted private skill registry with organizations, roles, SSO and audit logs. It also covers self-hosted and git-based options.

• Snyk Agent Scan

  Scans skills and MCP servers on developer machines and can report to a central Snyk instance for company-wide monitoring. Snyk also published research on how common vulnerable skills are.

• StepSecurity skill inventory

  Inventories the agent skills found in GitHub repositories and on developer machines. This is the closest existing product to the provenance graph idea's discovery layer.

• Vercel skills CLI and skills.sh

  A public registry and npx installer that handles installing, discovering and updating skills across 40+ coding agents.

• Sentry dotagents

  Beta tool that declares skill dependencies in agents.toml and locks versions so installs are reproducible. It is effectively a package manager with pinning.

• TanStack Intent registry

  Indexes SKILL.md files that ship inside npm packages, so it uses npm for versioning and distribution.

• SkillsMP

  A directory that calls itself the npm registry for agent skills. It claims 400k+ SKILL.md packages, a figure I have not verified.

• Mondoo skillcheck

  Searches across tens of thousands of skills and scans them for hidden prompt injection.

• SkillScan

  Open-source CLI that combines static YARA rules, an LLM dry run and a Docker sandbox with canary files. It outputs SARIF for GitHub.

• Superagent skill-security

  Open-source scanner that uses regex, AST and taint analysis, with a separate step for judging intent.

• skillhub

  Open-source Python CLI that lints, searches and installs skill packs from a curated registry.

• skrun

  Open-source Node CLI that imports and syncs skills across Claude Code, Cursor and OpenCode.

Checked 2026-10-10