Agents and skills
The console includes separate registries for Agent SDK and Skills SDK packages. Both are beta package-management surfaces. A package begins disabled and untrusted.
Agent SDK packages
Section titled “Agent SDK packages”The Agent SDK page validates and records declarative Agent SDK v0.3 manifests. Lifecycle controls can enable, disable, or quarantine a registered package.
This surface is a registry and inspector—not a general Agent SDK execution engine. Package metadata does not bypass:
- extension execution controls;
- sandbox selection;
- capability policy;
- network restrictions;
- secret boundaries;
- tool approvals.
Keep package instructions, integrity metadata, and provenance reviewable before enabling it.
Skills SDK packages
Section titled “Skills SDK packages”The Skills SDK page validates a Skill SDK v0.1 manifest and its rendered SKILL.md, stores immutable managed package versions, verifies persisted integrity, and controls lifecycle state.
An integrity failure blocks enablement.
Discovered skills
Section titled “Discovered skills”Ódinn can discover existing workspace and imported SKILL.md files as unmanaged packages. Discovery does not silently:
- install them into managed storage;
- trust them;
- enable them;
- inject them into a model;
- execute their instructions.
Review unmanaged content as untrusted input.
Practical review checklist
Section titled “Practical review checklist”- Is the source known and expected?
- Does the manifest match the package contents?
- Is integrity clean?
- Are requested capabilities necessary?
- Does the package instruct the model to expose secrets or bypass policy?
- Is there a clear disable and rollback path?
For executable tools and MCP servers, use Extensions and MCP. For imports, see OpenClaw and Hermes imports.