ClawBlog

Concept

AgentSkill

A packaged capability — and a packaged trust decision.

A packaged unit of capability an agent loads to gain a new action — OpenClaw’s building block, distributed via ClawHub. Convenient and composable, but each skill runs with the agent’s access, which makes provenance a security question.

Review

Not yet scored

No ClawScore review has been published yet.

Coverage

0 stories

No linked coverage yet.

Sources

1 link

Official and reference links for deeper verification.

Discovery

Noindex

Readable here, but withheld from sitemap until coverage deepens.

An AgentSkill is a packaged capability an agent installs to gain a new action: send a message, query an API, run a workflow. It is the building block of the OpenClaw ecosystem (100+ skills, with more than 13,700 listed on ClawHub) and the reason agents can be extended without rewriting the framework.

The model’s strength and weakness are the same fact: a skill runs with the agent’s connected access. That makes a skill a packaged trust decision, not just a packaged capability. ClawHavoc — the typosquatting supply-chain attack — was the demonstration: installing a skill by name match is installing a contractor with your keys. Skill curation (pinning, reviewing, scoping) is the operator discipline the model demands.