skill-review

name	skill-review
description	Security scanner for Claude Code Skill packages. Use when the user wants to audit, review, or check the safety of a Skill before installing — e.g. "is this skill safe?", "check this skill", "scan for backdoors", or "skill-review".

skill-review

A multi-agent security scanner CLI for Claude Code Skill packages. It combines deterministic static pre-scanning with LLM-driven deep analysis to surface security risks across 7 layers before you install a Skill.

When to use

• Auditing a third-party Skill before installation
• Checking a skill directory for prompt injection, credential theft, data exfiltration, or hidden backdoors
• Evaluating supply chain risk of a Skill's npm/PyPI dependencies
• CI/CD integration to block high-risk Skills automatically

How it works

The scanner runs in two phases:

1. Pre-scan (deterministic, no LLM) — walks all files and flags: symlinks, suspicious filenames (Unicode confusables, shell metacharacters), large files, binary executables, invisible characters, ANSI escape sequences, JS obfuscation patterns, and hardcoded URLs.

2. LLM Analysis — an Explore Agent reads each file and performs 7-layer analysis:

   • Layer 1: Prompt Injection (direct injection, jailbreak, remote prompt loading)
   • Layer 2: Malicious Behavior (credential theft, data exfiltration, sandbox escape)
   • Layer 3: Dynamic Code Loading (remote execution via fetch+eval, curl|sh, etc.)
   • Layer 4: Obfuscation & Binary (obfuscated scripts, compiled binaries)
   • Layer 5: Dependencies & Supply Chain (npm/PyPI/CLI tool inventory, typosquat detection)
   • Layer 6: System Modification (global installs, profile changes, cron jobs)
   • Layer 7: Code Quality (hardcoded secrets, insecure configs, vulnerable code patterns)

   An optional Deep Analysis Agent then verifies URLs, checks dependency metadata on registries, and inspects binaries.

3. Deterministic Scoring — each finding is scored based on its type and severity. The overall risk level (safe/low/medium/high/critical) and recommendation (install/caution/do_not_install) are computed deterministically, not by the LLM.

Installation

cd <skill-review-dir>
npm install

Configuration

Create .env and fill in your LLM provider details:

Variable	Description	Default
OPENAI_API_BASE	LLM API base URL (OpenAI-compatible)	required
OPENAI_API_KEY	API key	required
OPENAI_API_MODEL	Model name	gpt-4o
NPM_REGISTRY_URL	npm registry for dependency checks	https://registry.npmjs.org
PYPI_INDEX_URL	PyPI index for dependency checks	https://pypi.org

Alternatively, pass a JSON config file via --config.

Usage

# Standard scan (pre-scan + LLM explore)
node index.mjs <skill-dir>

# Pre-scan only (no LLM, fast)
node index.mjs --pre <skill-dir>

# Deep analysis (pre-scan + explore + deep verification of URLs/deps/binaries)
node index.mjs --deep <skill-dir>

# JSON output, save to file
node index.mjs --json -o report.json <skill-dir>

# Chinese language report
node index.mjs --lang zh <skill-dir>

# Verbose logs to stderr + log file
node index.mjs -v --log scan.log <skill-dir>

Options

Option	Description
<skill-dir>	Path to the skill directory to scan (required, positional)
--config <file>	Path to JSON config file
--pre	Run pre-scan only (no LLM calls)
--deep	Enable deep analysis phase
--lang <lang>	Report language (default: English)
--json	Output raw JSON instead of text report
-o, --output <file>	Save report to file (default: stdout)
--log <file>	Save detailed logs to file
-v, --verbose	Stream detailed logs to stderr
-h, --help	Show help

Output

The text report shows each layer with a risk score (0-10), star rating, and up to 5 findings per layer. The JSON output contains the full structured result with all findings, layer scores, overall risk, and recommendation.

Risk levels: safe (0) / low (1-3) / medium (4-6) / high (7-8) / critical (9-10)

Recommendations: install (safe/low) / caution (medium) / do_not_install (high/critical)