Agents shouldn't be monoliths.

What we believe

• Skills should be composable. A skill declares its inputs and outputs in a machine-readable manifest, so agents can chain them without human glue code.
• Permissions should be explicit. Every skill on this marketplace declares exactly what it can touch. Anything not declared is denied at runtime — and every secret:read is logged.
• Versions should be pinned. Installs resolve to exact versions. Authors can't silently change what your agent runs; breaking changes ship as new major versions.
• The spec should be open. The agentskills.io format is public and vendor-neutral. Skills built for one agent host work on any host that implements the spec.

Trust & safety

Every listing passes through human review before it goes live. Reviewers validate the manifest with askill validate, diff declared permissions against actual behavior, run the skill's own test suite, and check for exfiltration, prompt-injection payloads, and dependency confusion. Skills that touch secrets or the network get a second look.

Ratings and install counts are computed from verified installs only — authors can't inflate their own numbers. If a skill misbehaves after publication, report it and we'll pull the listing while we investigate; version pins mean your agents keep running the last good copy either way.

Where we're headed

Today: a curated catalog, one-command installs, and a spec anyone can implement. Next: dependency resolution between skills, signed manifests, and a community review track so trusted authors can publish faster. If you're building agent infrastructure, we'd love to hear what the spec is missing.

Contact

Questions about a listing, the spec, or a review decision? Write to hello@anthropicskills.com. For security reports, use security@anthropicskills.com — we aim to acknowledge within 24 hours.