Find Skills with OpenAgentSkill

Turn the user's task into an evidence-backed shortlist. Do not recommend a skill only because its repository is popular.

Workflow

1. Resolve the task with the public endpoint:

   GET https://www.openagentskill.com/api/agent/resolve?task=<encoded-task>&agent=<agent-name>

2. Read selected, recommendation_lanes, policy_decision, and at least three entries from alternatives when they exist.

3. Before installing, open the selected Skill's urls.audit and inspect its repository source, license, permissions, setup requirements, and freshness.

4. Use install_plan.command only when the policy allows it. If human review is required, present the risks and ask for approval. Never install a blocked Skill.

5. Run one narrow verification task in a sandbox or low-risk workspace.

6. Report the result to the returned feedback endpoint with its unique event_id. Report success only when installation and the verification task both succeeded.

CLI

The official release can perform the same workflow:

npx --yes https://github.com/Leon-Drq/openagentskill/releases/download/cli-v0.3.0/openagentskill-0.3.0.tgz find "<task>"

Use resolve for the complete decision object and add <slug> --dry-run before a reviewed install. Anonymous telemetry can be disabled with --no-telemetry or OPENAGENTSKILL_TELEMETRY=0.

Decision rules

• Prefer task fit, safety, verified outcomes, and maintenance over raw stars.
• Treat repository stars as a project-level popularity signal, not proof that every nested Skill has equivalent adoption.
• If no candidate fits, say so and use the agent's native tools instead of forcing an unrelated recommendation.