MailVerdict

Is this email phishing? Get a verdict, not a vibe.

Forward an email. Get a calibrated phishing verdict in under a second.

Start 7-day free trial See how it works

How it works

Step 1

Forward the email

Send the suspicious message to your MailVerdict inbox, call the REST endpoint from Power Automate, or ask your Copilot Studio agent. Raw .eml is best: headers give us SPF, DKIM and DMARC.

Step 2

Jev decides

Deterministic header and URL checks run first. Then Jev, a purpose-built judge model by TypeSafe, returns a calibrated probability plus attack type: credential harvest, payment fraud, malware, executive impersonation.

Step 3

You get the verdict

PHISHING, SUSPICIOUS or LEGITIMATE, the probability behind it, ranked indicators, and a plain-language explanation for the grey zone. Structured JSON for automation, a card for humans.

What a verdict looks like

PHISHING credential_harvest
93% calibrated probability · 640 ms
  • Link text says paypal.com but points to paypa1-secure.com
  • Reply-To domain differs from From domain
  • SPF failed for the sending server
  • Extreme urgency: "account suspended within 24 hours"
The message impersonates a payment provider and pushes you to sign in on a look-alike domain that the sender does not control. Do not click, reply or open attachments.
curl -X POST https://your-deploy.vercel.app/v1/analyze \
  -H "x-api-key: $MAILVERDICT_API_KEY" \
  -H "content-type: application/json" \
  -d '{"subject":"Account suspended",
       "from":"PayPal <alerts@paypa1-secure.com>",
       "body":"Sign in within 24h: https://paypa1-secure.com/login"}'

{ "label": "PHISHING",
  "probability": 0.93,
  "attackType": "credential_harvest",
  "indicators": [ ... ],
  "explanation": "...",
  "latencyMs": 640 }

Same result over MCP at /mcp for Copilot Studio: tools analyze_email, explain_verdict, get_policy.

Why calibrated beats an LLM guess

A probability you can set policy on

When Jev says 93%, roughly 93 of 100 such emails are phishing. That lets you pick thresholds (auto-quarantine above 0.9, ask a human between 0.1 and 0.9) and know the trade-off you are making.

Consistent, not chatty

General chat models drift with prompt wording and are confidently wrong on edge cases. A dedicated judge answers the same fixed questions every time. The LLM only writes the explanation, and only when the verdict needs one.

Fast and cheap by design

One judge call per email, sub-second latency, no giant prompts. Deterministic checks (auth results, link mismatch, dangerous attachments) surface as indicators alongside the model's view.

Pricing

$2.99
per month, per mailbox
  • Unlimited forwards (fair use)
  • Copilot Studio + Power Automate ready
  • Verdict + probability + indicators
  • 7-day free trial
Start free trial

Prefer to run it yourself? Self-host free (MIT) on GitHub.