Review an AI Skill before you install or publish it

An AI Skill can look like a harmless instruction file while reaching into credentials, environment variables, local files, external APIs, update scripts, or copyrighted material.

AI Skill Safety Review performs a static, evidence-based review of a Skill, Agent, plugin, repository, ZIP, or pasted instruction package. It inventories the artifact, checks provenance and licensing, scans for secrets and personal data, maps permissions, detects prompt injection and update risks, and turns the findings into a decision you can act on.

What you receive

• PASS, REVIEW, BLOCKED, or UNKNOWN verdict
• Highest-priority risks with file or section evidence
• Network, filesystem, shell, login, environment, and external-action permission map
• Provenance, version, checksum, license, and redistribution assessment
• Masked secret and personal-data findings
• A safe next step and explicit review limitations

Best for

• Creators installing Skills from GitHub or a marketplace
• Teams reviewing internal Agent packages before sharing them
• Publishers removing secrets and private data before release
• Buyers who want a concise decision with plain-language evidence

How to use

Install the downloaded Skill in a compatible Agent Client. Upload or point the Agent Client to the package you want to review. You can also paste instructions or provide an accessible official repository or release URL. Add the intended action, such as install, adapt, or publish, when relevant.

Privacy and data flow

The Skill makes no network calls and sends no reviewed material to the Publisher. Your own Agent Client and selected LLM provider process the content under their terms. Do not intentionally submit live API keys, private keys, customer records, regulated data, or other secrets. If a sensitive value is found, normal output masks it and recommends revocation or replacement.

Safety boundary

This is a defensive static-review tool. It does not execute, install, exploit, bypass, or modify the reviewed package. It cannot prove that software is free of security, privacy, licensing, or operational risk. Independent review remains necessary.

License

You may inspect and modify the downloaded files for personal or authorized business use. Redistribution, resale, republication, sublicensing, AI model training, and unlawful use are prohibited. Full terms are included in the package. You remain bound by Capafy's Terms of Service and checkout terms.

Functional test

Example input: Review a package that stores an auth token, calls one API domain, can run pip, and can self-update files inside its own directory. The official download provides no source repository or published checksum.

Expected result: REVIEW or BLOCKED with explicit findings for token storage, network access, subprocess use, self-update scope, missing provenance evidence, and mitigations. The package must not be executed.