Ibán Ríos Español

Scanner for skills, plugins and MCP

Snitch?

Before you install a skill, a plugin, an MCP server or a repository, check whether it sends your data out, runs things behind your back or carries hidden orders for your AI agent. It is checked in your browser: nothing is uploaded.

Drop here what you want to check

A .zip, .skill, .mcpb, .tgz or a whole folder. If it lives on GitHub, its address is enough.

A folder or a single file works too. Your browser downloads it straight from GitHub; nothing reaches my site.

Or paste the text

What you check never leaves your browser, and you can verify it

  • This page is served with the header Content-Security-Policy: connect-src https://api.github.com https://raw.githubusercontent.com: the browser only lets it connect to GitHub, to download the repository you paste. Nowhere else, not even my site.
  • Check it: open the developer tools, Network tab. With a file, no request shows up; with an address, only GitHub's.
  • When you paste an address, GitHub sees your connection and which repository you ask for, as if you opened it yourself. Without an account it allows 60 lookups an hour; each check uses one.
  • No accounts, no cookies, no analytics and no AI: fixed rules that run on your computer.

What it looks for

  • Data out: capture inboxes (webhook.site, ngrok, Telegram or Discord bots…), uploads with curl, bare IP addresses, back doors, images that carry data in their address.
  • Secrets: SSH keys, cloud and GitHub credentials, browser passwords and cookies, wallets, all your environment variables, your agents' sessions.
  • Hidden code: base64 that gets run, invisible characters (and what they say), text spelled as character codes, code on a single line, compiled programs.
  • Orders for your agent: «ignore the previous instructions», «do not tell the user», tags posing as the system, comments you cannot see when the Markdown is shown.
  • What stays: scripts that run on install, hooks, permission to run any command, changes to your terminal's startup or your agent's configuration, scheduled tasks.

What it cannot see

  • It reads the text; it runs nothing. Code that downloads more code later, or heavily obfuscated code, can slip through.
  • It does not check dependencies (node_modules and the packages it installs): npm audit or OSV-Scanner do that.
  • Green means «I saw nothing odd», not «it is safe». Red means «look before you install», not «it is malware».

A tool by Ibán Ríos. Accessibility