Clean it now.
Catch what lands.

Every day comes back
as one page.

Each message SiftMail sorts is given a verdict. Most are ordinary and need nothing from you. What is left is a short numbered list — sender, subject, the reason it is on the list, and what it wants from you. The same list is the daily digest: one computation, one set of numbers, no second opinion.

Sample data · The daily digest on Pro, once you switch it on · 14:00 UTC or your own hour

You see the before and
the after in one sitting.

Connect, and SiftMail reads and scores the 500 most recent messages in your inbox. Until you finish or dismiss the free deep clean, opening the dashboard prepares it without waiting to be asked: it counts the promotions, the bulk mail and the no-reply mail older than ninety days in your inbox, reading up to 2,000 messages in each.

The sweep runs as a preview. It shows you the count and a sample of the messages behind it, and it changes nothing until you apply. Applying archives: the mail leaves the inbox and stays in All Mail. The one way SiftMail sends mail to Trash is the Heavy mail screen, on messages you pick and confirm, at most 200 at a time, and Gmail keeps them in Trash for 30 days.

Sample preview · Not your mail

The deep clean is free, and you can run it again whenever the inbox fills back up. You see a preview first; nothing moves until you confirm. Confirming archives the matching mail — it leaves the inbox, it is not deleted, and Undo puts it back. Shadow mode stays on until you change it, so new mail is scored and flagged and nothing else. On Pro you can switch Shadow mode off and auto-quarantine on, and high-confidence threats get the Sift/Quarantine label on every sync. A label is all it adds: the message stays where it landed. The daily digest is Pro too, once you switch it on.

Then it keeps watch.

Every sync, SiftMail scores what just landed with the same engine (up to 100 messages a day on Free, no daily cap on Pro). Threats are flagged, bulk is told apart from the mail that needs you, and a message that wants a reply stays in front of you for up to seven days, whether or not you were at your desk the day it landed. Scoring never moves a message out of your inbox. When it fills back up, run the deep clean again; it is free every time. On Pro, switch on the daily digest and one email a day says what happened.

If SiftMail acted, it says acted. If it only flagged, it says flagged. It never claims the other one, because a clean only counts the messages Gmail confirmed it moved. And every clean records what it changed, so Undo puts that run back.

Auto-quarantine, the daily digest and 3 mailboxes are Pro · $15/mo · Gmail

Every flag shows its work.

A risk score on its own is a number you are asked to trust. SiftMail prints the signals instead, in the same words the product uses everywhere else. Each one is checkable in the message headers, by you, which is the entire point of printing them.

The gate runs the other way too. A message only reaches the digest's most trusted slot — the service alert — on a curated infrastructure sender, an alert-shaped subject, and a DMARC pass. And a DMARC-aligned sender you have real history with does not get called a threat on a hunch: the guard stands the flag down, and the digest stands it down with it. One inbox, one story.

Promotion to a critical alert requires a DMARC pass. Never without.

We mark it.
We don't take it.

In Gmail, quarantine adds one label — Sift/Quarantine — and stops. The message stays in your inbox, in its thread, in the place you last saw it. Release takes the label back off. Which means the worst thing a false positive can do to you is disagree with you in writing.

Archiving is one of two actions that take mail out of the inbox. A cleanup you confirm archives promotions, bulk mail and no-reply mail older than ninety days. SiftMail also has a recurring maintenance setting that would archive recent inbox mail the classifier rates safe to archive, but never starred or important mail, mail that needs a reply, or mail a person wrote. It is off by default, and the app has no switch for it yet, so Pro does not include it today. The other action is Trash, on messages you pick and confirm on the Heavy mail screen. Most tools in this category ask you to trust that they took the right thing. A quarantine takes nothing, so there is nothing here to trust.

SiftMail for Gmail — in-inbox risk badges · Built at 3.4.0, not yet listed on the Chrome Web Store

Half museum.
Half liability.

The same reader, pointed backwards. Your mailbox comes back as strata — what year, what is in it, and what a stranger holding your password would find first. A passport scan from 2016. A tax form. A temporary password somebody typed into an email and never thought about again. Sitting in the same result set: the oldest photograph anyone ever sent you.

Red is reserved for identity-level exposure; if everything were critical, nothing would be. Old phishing that already failed is trash, not a siren — it is muted and grouped for review, never presented as an emergency a decade late. No model runs on this path at all, and the build fails if anything on it so much as imports one.

No model runs on this path · Findings stored per-user, visible only to you · Deletion happens in Gmail, by you

Here is exactly what
leaves your mailbox.

Headers and a snippet go to Anthropic when our own scoring cannot settle a message. Up to 4,000 characters of the body go on ingest — one pass that sorts it, spots sensitive content, and decides whether it wants a reply. A message can be sent again, for example after a failed attempt or when a later sync picks it up. If you connect Slack, SiftMail posts an alert to the webhook you supply when a message the SiftMail for Gmail extension checks gets one of the labels you chose. The alert carries the sender, subject, risk score, label and the signals behind the score, which can include the AI's short assessment.

We are not going to tell you this runs on your device, because it does not. We are not going to tell you we never read your mail, because reading it is the job. It is read by automated systems rather than by staff — with the narrow exceptions our privacy policy sets out, which is also the only complete list. Tokens are encrypted at rest with AES-256-GCM.

Four scopes. Nothing else is requested.