ClawHub

ClawHub is the launch set's marketplace-shaped review. Marketplaces are attractive because they turn agent capability into a catalog. They are dangerous for the same reason: every easy install path becomes a supply-chain boundary. An agent skill is not a harmless theme pack. It can shape what an agent reads, calls, installs, or sends. That means the marketplace's trust model is not a secondary feature. It is the product.

The review score is therefore split. Capability and Momentum get credit for the ecosystem role. A healthy catalog can make useful agent behavior easier to discover, and a marketplace can create incentives for specialist skills that one core project would never build. Setup & DX also benefits when installation paths are clean. But Safety & Control and Reliability are deliberately conservative because ClawBlog's own coverage of skill-marketplace attacks makes provenance the central issue, not a footnote. A catalog without verification is not a moat. It is an attack surface with search.

The draft should be read as a security-first marketplace review. The questions before publication are concrete: are packages signed or otherwise provenance-checked? Are maintainers verified? Is there automated scanning? Is there a human review queue for sensitive capabilities? How fast can malicious or abandoned skills be removed? Can operators inspect permissions before installation? Can an enterprise block risky sources? The score should move only when the receipts answer those questions, not when the catalog count looks impressive.

This draft should be edited against the latest ClawHub security posture before publication. If scanning, signing, review queues, and removal workflows have improved, the Safety & Control score can move. If they have not, the verdict should get sharper. ClawLab should test a benign skill, a stale skill, and a deliberately suspicious install path in a controlled environment. The ideal ClawHub review is not anti-marketplace. It is pro-marketplace with the correct condition attached: distribution earns trust only when provenance, permissions, and removal workflows are strong enough for agents that act in the world.