security check
Nobody reads what their agent installs. We do.
Your agent installs AI skills, MCP servers and npm packages on a single chat message. Paste what you're about to install and see exactly what it does before it ever runs on your machine.
Paste the install command you were about to run. We read the published files as text — we never run them.
try:
Automated risk assessment. Not a guarantee of safety.
What we look at
One careless install can hand a stranger your SSH keys, cloud credentials and every file on disk — or plant hidden instructions your agent will follow without ever showing you. A check takes seconds.
npm package security check
- Install scripts that fetch or run code on your machine
- Code that touches SSH keys, cloud credentials, tokens or wallets
- Where it sends data, and whether those places are known
- Release age, provenance, maintainer changes and known advisories
MCP security scanner
- Every tool it offers, and what each one claims to do
- Hidden instructions and invisible characters in tool descriptions
- Tools that steer your agent towards other tools or secrets
- A fingerprint of the tool list, so you are told when it changes
Recent checks, with real findings
Every report below came from a live package or server — read the findings before you install.
loading…
Common questions
- Does the check run the package?
- No. Every check is static: we read the registry metadata, the published tarball as text, and — for MCP — the tool definitions the server advertises. Nothing is installed and nothing is executed, so checking a malicious package cannot harm you or us.
- What does each verdict mean?
- On the list means nothing notable turned up. Worth a look means there are findings a human should read before installing — a lifecycle script, an unfamiliar outbound domain, a known advisory. Turned away means something high or critical was found. A verdict is an informed automated opinion, not a certification.
- Can I check a private MCP server?
- Yes. Switch to MCP mode and paste a one-time token. It is used for that single request, never written to our database, and the resulting report stays private to you instead of appearing in the public list of recent checks.
- Why was a minified bundle not flagged as obfuscated?
- Because ordinary minification is not an attack. We separate real obfuscation — mangled hex identifiers, self-decoding strings, packers — from normal build output, so honest packages are not punished for shipping a bundle.
- What happens after I approve something?
- The exact version is pinned and you get a ready-to-paste config for Claude Code, Claude Desktop, Cursor, VS Code, Windsurf or Zed. If you leave watching on, we re-check it and alert you when a new version, a changed tool definition or a fresh advisory appears.
- Is there an API?
- Yes — a free HTTP API at
/api/public/v1/check, capped at 10 checks per hour, plus in-page WebMCP tools so an agent can check a package before it installs it.
Checking is the first gate. Once something is running, the MCP gateway governs what it can actually call, and the API catalog lists the servers published behind it.