Skip to main content
Someone sent you a link and you’re not sure about it. Your agent scans it across 70+ security engines, reads the actual page content, captures a screenshot so you can see it without clicking, then Claude gives a plain-English verdict. ~$0.08, self-contained — just the URL.
This is a durable gateway demo — it inspects the live URL right now across real security engines. A sandboxed client can’t fake a 70-engine scan or a real screenshot.

The prompt


What runs

  1. POST /virustotal/v1/scan — scan the URL across 70+ antivirus engines (~$0.02)
  2. POST /jina/v1/read — read the destination page as clean markdown (~$0.02)
  3. POST /screenshot/v1/capture — capture the page so you can see it safely (~$0.02)
  4. POST /anthropic/v1/messages — Claude gives the verdict + recommendation (~$0.02)

Run it

Claude Desktop (MCP)

Paste the prompt with any URL. The agent runs the scan, read, and screenshot in parallel before the verdict.

SDK


Expected output


Extend it

  • Add IPinfo (/ipinfo/v1/lookup) on the resolved host IP for geo + hosting provider
  • Scan attached files too — VirusTotal also takes a file hash
  • Batch a whole inbox of links in parallel; flag only the ones that come back caution/avoid
  • Pipe flagged links into Resend (/resend/v1/emails) for an automated security digest