SkillSafe CLI

The open-source Python CLI for SkillSafe — the secured skill registry for AI coding tools. Scan, save, share, install, and verify AI skills with security scanning and dual-side cryptographic verification.

Works with Claude Code, Cursor, Windsurf, Codex, Gemini CLI, OpenCode, Cline, Roo, Goose, Copilot, Kiro, Trae, AMP, Aider, and more.

Quick Start

No dependencies — uses only Python stdlib. Requires Python 3.8+.

pip3 install skillsafe

# Install a public skill (no account required)
skillsafe install @anthropics/pdf

# Sign in to save and share your own skills
skillsafe auth
skillsafe scan ./my-skill
skillsafe save ./my-skill --version 1.0.0
skillsafe share @myname/my-skill --version 1.0.0

Install

pip (recommended)

pip3 install skillsafe

AI coding tool

Tell your AI coding tool:

Install skillsafe from https://skillsafe.ai/skill.md

Commands

Authentication

Command	Description
auth	Sign in via browser. Saves API key to ~/.skillsafe/config.json
whoami	Show current auth status, namespace, and masked API key

Scanning & Security

Command	Description
scan <path>	Run a 12-pass security scan (AST analysis, secrets, injection, structural mimicry, composite patterns, surplus functionality)
scan <path> -o report.json	Save scan report as JSON
bom <path>	Generate a Bill of Materials — inventory of files, APIs, capabilities
bom <path> -o bom.json	Save BOM as JSON

Publishing

Command	Description
init [path]	Initialize a skillsafe.yaml manifest in a skill directory
lint [path]	Validate a skillsafe.yaml manifest
save <path> --version <ver>	Save a skill privately to the registry
share @ns/name --version <ver>	Create a share link (requires email verification + scan report)
share @ns/name --version <ver> --public	Share with public visibility (discoverable via search)
share @ns/name --version <ver> --expires 7d	Share with expiration (1d, 7d, 30d, never)
yank @ns/name --version <ver>	Yank a version — blocks future downloads

Installing

Command	Description
install @ns/name	Install to .agents/skills/ and auto-symlink to detected agents
install @ns/name --version <ver>	Install a specific version
install @ns/name --tool claude	Install directly into .claude/skills/
install @ns/name --tool cursor --location global	Install to a tool's global skills directory
install @ns/name --skills-dir ./custom	Install to a custom directory
install @ns/name --no-symlink	Install to .agents/skills/ without creating symlinks
install <share-link>	Install via a shr_ share link or URL

No account required for public skills. Authenticated installs enable dual-side verification.

Discovery

Command	Description
search <query>	Search public skills
search <query> --sort trending	Sort by popular, recent, verified, trending, hot
search --category "code review"	Filter by category
info @ns/name	Show skill details, versions, and verification status
list	List all locally installed skills across all agent directories

Updating

Command	Description
update @ns/name	Upgrade a specific installed skill to the latest version
update --all	Upgrade all installed skills
update --all --tool claude	Upgrade all skills for a specific tool
update --all --dry-run	Preview upgrades without applying

Demos

Command	Description
demo <json> @ns/name --version <ver> --title "My demo"	Upload a demo recording

Evals & Benchmarks

Command	Description
eval @ns/name --version <ver> --eval-json results.json	Upload eval results
eval @ns/name --version <ver> --pass-rate 95 --test-cases 20	Upload eval metrics directly
benchmark @ns/name --version <ver> --runs 10 --avg-time 2.5	Upload benchmark results

Vault (Backup & Restore)

Command	Description
backup <path>	Back up a skill directory to the encrypted vault
restore @ns/name	Restore a skill from the vault
restore @ns/name --tool claude --location global	Restore to a specific tool directory

Import & Claim

Command	Description
import <url>	Import a skill from a GitHub or ClawHub URL
claim github.com/owner/repo	Claim a GitHub repo as your skill on SkillSafe

Security Model

SkillSafe uses dual-side verification:

1. Publisher scans before sharing
2. Consumer re-scans after download
3. Server compares both reports

Tree hashes (SHA-256 of archive content) detect tampering. Verdicts: verified, divergent, critical.

The scanner runs 12 analysis passes including AST parsing, regex pattern matching, credential detection, prompt injection analysis, and composite behavioral patterns.

File Structure

pyproject.toml             # Package metadata
src/skillsafe/             # Python package
    __init__.py            # Public API: Scanner, SkillSafeClient
    __main__.py            # python -m skillsafe support
    _cli.py                # CLI implementation (stdlib only)
SKILL.md                   # Skill definition (source of truth)
LICENSE                    # MIT

License

MIT