Skip to content

Skill Resolvers

resolve_skill(uri) fetches a skill from a URI and returns a handle with the extracted local path. The AgentSkillStepConfig.skill field uses the same resolution.

from mmar_carl import resolve_skill

skill = resolve_skill("github://anthropics/skills/skills/pdf@main")
print(skill.local_root)   # ~/.cache/mmar_carl/skills/github/<key>/skills/pdf

Supported schemes

Scheme	Resolves from
`github://owner/repo/path@ref`	A GitHub tarball (no `git` needed).
`local://` / plain path	A local directory.
`https://`	A remote archive.
`module://pkg.path`	A Python package.

SkillResolverRegistry dispatches by scheme; register custom resolvers to extend it. GithubResolver downloads tarballs from codeload.github.com and caches them under ~/.cache/mmar_carl/skills/github/<key>/.

Integrity (SHA pinning)

Pin a skill to a known SKILL.md digest so a compromised upstream is caught:

skill = resolve_skill(
    "github://anthropics/skills/skills/pdf@main",
    sha256="<hex-of-SKILL.md>",
    trust_policy="sha_pinned",
)

With trust_policy="sha_pinned", CARL verifies the SHA256 of the local SKILL.md after extraction and raises SkillIntegrityError on a mismatch. Force a fresh download with GithubResolver().resolve(uri, force_refresh=True).

Discovering installed skills

from mmar_carl import SkillLoader

skills = SkillLoader().catalog_all()   # SKILL.md dirs + the agent-skills library

SkillManifest parses the SKILL.md frontmatter, including get_allowed_tools() / get_allowed_tool_names() for the allowed-tools list.

See also

AgentSkills overview · Sandboxing