Skillager

Skillager is a local CLI for discovering, reviewing, searching, and exposing agent skills without loading every skill into every chat.

discover -> review -> search metadata -> expose only what the task needs

Skills can come from project folders, Python environments, npm packages, Cargo crates, native agent skill directories, or shared collections. Skillager keeps them searchable after review, then writes native, stub, or router skills for the agent when you choose to expose them.

Quickstart

uv tool install skillager
# or: pipx install skillager

# Optional: register a personal/team collection, such as Superpowers.
skillager collection add ~/skills/workflows --name workflows

# Run per project.
skillager setup --agent codex

Then restart your agent in the project and have it run:

skillager working --agent codex --json

setup discovers local/project skills, package-provided skills, collections, and native agent skills. It scans them and only makes content available after your review. Skillager is installed once as a user tool; it does not need to live inside every project environment.

Core Model

Skillager keeps these choices separate:

Choice	Meaning
Approval	You reviewed this skill at its current content hash.
Curation	A project groups approved skills into tags like gis, workflows, or release.
Exposure	Skillager writes native, stub, or router skills for an agent and project.

Tip

Approval is not exposure. Approved skills are searchable; expose only what a project or task needs.

Exposure Modes

Mode	Best For	What The Agent Sees
native	Normal agent skills, as if Skillager were not involved	Full reviewed skill body
stub	Named skills you want available without loading the body	Tiny activation handle
router	Larger tags or one-off skill sets	One compact multi-skill router with supporting metadata

Routers do not load full skill bodies. They list reviewed members and activate one on demand:

skillager expose --tag workflows --mode router --agent codex --scope project
skillager activate workflows/pr-review --from-router workflows

Metadata commands stay metadata-only: working, list, search, show without --content, tag show, tag list, doctor, and summary JSON do not print full skill bodies.

Common Commands

Task	Command
Review or refresh a project	skillager setup --agent codex
Diagnose state	skillager doctor --agent codex
Repair Skillager working artifacts	skillager doctor --agent codex --fix
Approve a skill	skillager review approve workflows/pr-review
Expose a tag as a router	skillager expose --tag workflows --mode router --agent codex --scope project
Expose explicit skills as one router	skillager expose workflows/pr-review workflows/release-check --mode router --agent codex --scope project
Expose one skill as a stub	skillager expose workflows/pr-review --mode stub --agent codex --scope project

Read-only allowlist examples for agent permission prompts: codex, claude. Keep mutating commands user-run unless you intentionally delegate them.

Collections

Collections are user-global skill sources. A collection can be a personal repo, a company-maintained repo, or a public skill repo like Superpowers. Tags are project-local curation, usually maintained by the agent after setup.

skillager collection add ~/skills/workflows --name workflows
skillager setup --collection workflows --agent codex
skillager tag add workflows --from-collection workflows --sync
skillager expose --tag workflows --mode router --agent codex --scope project

For fully trusted personal or company collections:

skillager setup --collection workflows --bulk-approve --agent codex
# same path, more fun:
skillager setup --collection workflows --yolo --agent codex

After review, collection skills are searchable from any project on your machine.

Package Authors

Python libraries, npm packages, and Cargo crates can ship skills in .agents/skills/:

your-package/
  .agents/skills/
    fastapi-usage/
      SKILL.md
      skillager.yaml
      references/
      scripts/

Skillager discovers installed package skills from project Python environments, node_modules, and Cargo.lock-selected crates without importing packages, running package scripts, or invoking Cargo. Users still review skills before activation or exposure.

skillager.yaml is optional and structured-only. Put searchable prose in SKILL.md; use the manifest for audience, activation, compatibility, and package-target metadata. For CI:

uvx --from skillager-linter skillager-lint .

See the package author guide for details.

Safety

Skillager's scanner is deterministic, local, and imperfect. It flags common agent-risk patterns such as instruction overrides, hidden prompt requests, credential paths, download-and-execute flows, secret exfiltration language, encoded blobs, and oversized content.

Human review decides availability. Signatures are provenance evidence, not safety signals: a verified signed skill still needs normal review before activation or exposure.

Docs

• User guide
• Agent CLI guide
• Skill repositories
• Package author guide
• Safety model
• Release runbook
• Security policy

External contributions are not being accepted yet while the early API and workflow settle.

Development

uv run python -m unittest discover -s tests
uv run python -m unittest discover -s packages/skillager-linter/tests
uv run --python 3.13 python scripts/check.py
uv build packages/skillager-linter
uv build

Skillager is released under the MIT License.