skilly 0.0.28

Manage agent skills.

skilly

Manage Agent Skills from the command line or Python.

skilly creates specification-compliant skills, installs skills from GitHub or Python dependencies, and keeps managed skills up to date.

Installation

Run with uvx:

uvx skilly --help

Run with npx:

npx @xelandernt/skilly --help

Install the native CLI with Homebrew from this repository's tap:

brew tap xelandernt/skilly https://github.com/xelandernt/skilly
brew install xelandernt/skilly/skilly

Homebrew installs the CLI only. Install the Python package when you need the import surface as well:

pip install skilly

The npm package ships the same native Rust CLI behind the executable name skilly. The current npm support matrix is macOS arm64/x64, Linux x64 (glibc), and Windows x64.

Quick Start

Scan the current Python project for skills shipped by its dependencies:

uvx skilly scan

Download a skill from GitHub:

uvx skilly download https://github.com/example/project/tree/main/skills/code-review

Inspect installed skills:

uvx skilly list

CLI Commands

Command	Purpose
uvx skilly scan	Find skills provided by the project's Python dependencies.
uvx skilly download <github-url>	Install one or more skills from GitHub.
uvx skilly list	Browse, update, or remove installed skills.
uvx skilly update	Preview available updates; add --yes to apply all updates.
uvx skilly remove <name>	Remove an installed skill by directory name.
uvx skilly skillsmp search <query>	Search SkillsMP and install a selected result.
uvx skilly create	Create a valid skill through a terminal wizard or explicit options.

Use uvx skilly <command> --help for all options.

Create Skills

Interactive terminals open a full-screen editor for the required and optional Agent Skills fields. Ctrl+S creates the skill, Ctrl+X cancels, and Enter in the multi-line editors inserts real line breaks.

For scripts and automation, provide the required name and description:

uvx skilly create deployment-checks \
  --description "Validate deployment readiness before a production release." \
  --instructions "# Instructions

Run the deployment checklist and report blockers." \
  --metadata owner=platform \
  --with-scripts \
  --yes

Existing skills are rejected by default. Use --overwrite to replace one atomically. See uvx skilly create --help for the complete contract.

Install Dependency Skills

uvx skilly scan reads pyproject.toml and the project's .venv, then offers skills shipped by direct, development, and optional dependencies:

uvx skilly scan

Filter named groups and extras when needed:

uvx skilly scan --group dev --exclude-extra docs

Install GitHub Skills

uvx skilly download https://github.com/example/project

When a repository contains multiple skills, select one or install all:

uvx skilly download https://github.com/example/project --skill-name code-review
uvx skilly download https://github.com/example/project --all

Destinations

Management commands accept the same destination options:

Option	Destination
--local	Project-local skills directory.
--global	User-global skills directory.
--claude	Claude skills directory.
--codex	Codex skills directory.
--copilot	GitHub Copilot skills directory.
--directory <path>	Explicit directory; overrides all other destination options.

Without destination options, skilly uses .agents/skills.

Set SKILLY_DIRECTORY to change that default without passing --directory every time:

export SKILLY_DIRECTORY="$HOME/.config/skilly/skills"
skilly list

--directory, --local, --global, --claude, --codex, and --copilot override SKILLY_DIRECTORY when provided explicitly.

uvx skilly download https://github.com/example/project --global --codex
uvx skilly list --local --claude

Destination Mapping

skilly resolves destination paths before use. Relative paths become absolute from the current working directory, and --directory "~/.copilot" expands ~ to your home directory.

Flags	Resolved destination
none	SKILLY_DIRECTORY when set, otherwise .agents/skills
--local	.agents/skills
--global	~/.agents/skills
--claude	.claude/skills
--claude --global	~/.claude/skills
--codex	.codex/skills
--codex --global	~/.codex/skills
--copilot	.github/skills
--copilot --global	~/.copilot/skills
--directory <path>	Exactly that directory, after ~ expansion and absolute-path resolution.

--directory overrides --local, --global, --claude, --codex, and --copilot.

GitHub Authentication

Authenticated requests have higher GitHub API rate limits. Use the first available token:

export SKILLY_GITHUB_TOKEN=ghp_your_token
# or GITHUB_TOKEN / GH_TOKEN

GitHub-fetching commands also accept --github-token.

Python API

SkillRepository is the stateful management interface. Bind a destination, project settings, or custom filesystem once and reuse it:

from pathlib import Path

from skilly import ProjectSettings, Skill, SkillRepository

repository = SkillRepository(
    directory=Path(".agents/skills"),
    project=ProjectSettings(
        dependency_groups=("dev",),
        optional_dependencies=("docs",),
    ),
)

created = repository.install(
    Skill(
        name="code-review",
        description="Review code for correctness and maintainability.",
        content="# Instructions\n\nReview the proposed change.",
    )
)

for match in repository.scan_project():
    print(match.available.name, match.status)

Use focused discovery functions when no repository state is needed:

from skilly import discover_installed_skills, discover_venv_skills

installed = discover_installed_skills()
dependency_skills = discover_venv_skills()

The SkillsMP client returns typed results directly. Async methods expose the same result types without blocking the event loop:

from skilly.skillsmp import ClientSettings, SkillsMp, SkillsMpSearchQuery

client = SkillsMp(settings=ClientSettings(base_url="https://skillsmp.com/api/v1"))
result = client.search(SkillsMpSearchQuery(text="python", limit=5))
print(result.data.skills[0].github_url)

Development

Install development dependencies and the editable extension:

just install

Run the required quality gates:

just lint
just test
just typecheck

The TypeScript launcher package also has dedicated helpers:

just ts::build
just ts::test
just ts::typecheck
just ts::smoke
just ts::publish-dry-run

License

MIT