Skip to content

Explanation: Core Concepts

Agent Skill

An Agent Skill is a packaged unit of capability for an AI Agent. It is defined by a SKILL.md and may contain supporting files (scripts, prompts, schemas).

skr treats skills like container images. They are: - Versioned: Using tags (e.g., v1.0.0, latest). - Portable: Can be pushed to standard OCI registries. - Layered: Deduplicated storage of content.

Skill Sources & Storage

skr considers Git the standard, common methodology for installing Agent Skills. You just point skr install at a public repository, and it handles the rest.

Behind the scenes, skr utilizes the Open Container Initiative (OCI) specification for storage and local caching. When a skill is cloned from Git, skr automatically unpacks it locally into standard OCI formats: - Manifest: Describes the skill content (config + layers). - Config: Metadata about the skill (creation time, author, etc.). - Layers: The actual file content (tar.gz).

This underlying OCI architecture allows for seamless scaling: While Git is the common path, you can use remote OCI registries (like GitHub Packages ghcr.io, Docker Hub, or Harbor) as a "value add." Authors can publish pre-built artifacts to registries to accelerate downloads and freeze distributions for deployment environments, bypassing the need for git cloning entirely.

The Global vs. Local Scope

  • System Store: A global cache of all downloaded/built artifacts on your machine.
  • Project Scope: When you run skr install, skills are "installed" into your project (referenced in .skr.yaml and synced to .agent/skills).
  • Lockfiles: The .skr.lock file stores precise digests corresponding to the tags in .skr.yaml, guaranteeing reproducible skill resolution on subsequent skr sync operations across team environments.