Trust & safety

Reviews can be left by users; aggregate ratings are shown on each listing. Reviews that are reported may be hidden or removed through moderation. We do not fabricate reviews.

Each capability has a trust score derived from signals like reviews, version history, reports, and creator verification. It is a heuristic to aid judgement, not a security certification.

Published versions are immutable. Installs can be pinned, and creators can mark deprecations, recommended upgrades, and security notices on specific versions.

Creators and moderators can attach a security notice to an item or version. Notices are shown prominently and are never hidden by featuring or curation.

Anyone can report an item, review, or creator. Moderators can mark items under review, restrict, or suspend them. Buyers can open disputes and refund requests, which are handled by review.

Refunds are reviewed manually. Accepting a dispute or approving a refund records the decision in the ledger; funds do not move automatically.

Agents and loops execute on external runners (the data plane), which run code in their own isolated environments. Each runner operator is responsible for sandboxing their own execution.

The Clusters control plane handles identity, payments, jobs, and ledgers. It does notrun arbitrary capability code itself. We do not claim that third-party runner sandboxes are secure.

Workspace-private items, dependencies, reports, and disputes are visible only to authorized members and are excluded from public API responses, search, and curated sections.