Do I need a paid name to get the AI?

Yes. The AI features come with a paid name. A free account is a complete mailbox with webmail, IMAP and SMTP, but its mail is not read by any model and nothing from it is sent anywhere.

Does my mail leave your servers?

The analysis that happens to every message runs on our own GPUs, on an open-weights model we host. Only three things you start yourself go out: translation and reply drafts to Anthropic, the thread advisor to OpenAI.

How often can I run the advisor on a conversation?

Five runs a day per account, and 150 a day across the whole service. The other AI features are limited by a monthly message allowance instead.

What the xmail AI does with a message

Most mail does not need you. It needs a decision: this one matters, that one is a receipt, this one is someone pretending to be your accountant. With a paid name, xmail makes that decision on every message before you open it. This page describes what it does today, 21 October 2026.

The short answer

Every incoming message is read by our own model and gets a verdict: what it is, what the sender wants, how urgent it is and how risky it looks.
Three things happen only when you ask: a translation, a draft reply, and the advisor, which reads a whole conversation.
All of it comes with a paid name. A free account is a real mailbox, and no model reads its mail.
The verdict is an opinion shown to you. Nothing is deleted, filed or blocked on your behalf.

The verdict on an incoming message

A message arrives and goes through a chain of checks. The cheap ones — sender reputation, rules, a spam score, a statistical model — run on every account. On a paid name the chain continues: our own model reads the message and writes down what it is. A bill. A delivery notice. A recruiter. A contract that needs an answer. With it: who the sender appears to be, what they want, a priority, a risk score and a one-line summary you can read instead of the message.

This is not a demo path. In the thirty days to 20 October 2026 the service produced 20 425 of these analyses for 356 accounts. In the last seven days of that period, 3 309 of the 5 643 analyses went through the deep step that also extracts amounts, dates and names.

Where it runs

On our own graphics cards, on an open-weights model (Qwen 2.5) we host ourselves. The quick check sees at most the first 1 500 characters of a message, the deep step at most 6 000, and tracking links are shortened to the bare domain before the model sees them. Nothing about that step leaves our network.

The ratio is the honest measure. In the thirty days to 20 October 2026 the service made 13 627 model calls on mail: 13 564 on our own hardware, 56 to Anthropic, 7 to OpenAI. The automatic work stays in-house; only a feature you press a button for goes out.

Translating and drafting

Two buttons sit on an open message. Translatesends the subject and up to 6 000 characters of the body to Anthropic, with phone and account-number-like strings masked before sending and restored in the result. Draft a replysends the sender, the subject and up to 3 000 characters to Anthropic and gives you back a draft in your editor. You send it or you delete it; nothing goes out by itself.

Asking a question about the message you are reading is answered by our own model, on our own hardware.

The advisor on a conversation

A single message is sometimes the wrong unit. A negotiation running for three weeks, a thread where the bank details changed halfway through, a chain of replies you inherited — the advisor reads the whole conversation and gives you a risk analysis of it: what is unresolved, and what looks wrong. It sends the thread to OpenAI, up to 12 000 characters per message, the largest amount of your text xmail ever sends anywhere, and it is capped at five runs a day per account and 150 across the service.

Where the line is

The AI needs a paid name on the account — an active paid name, a captured payment, or a grandfathered account. Prices for names are on the pricing page — an ordinary [email protected] is free forever, and famous or single-word names are priced individually.

On a free account the chain stops before the model. The message is delivered, filtered by rules and reputation, and marked as not analysed. No model call is made, so nothing from a free mailbox is sent to us or to anyone else.

What it does not do

It does not move, delete, archive or block anything for you, and it does not answer mail in your name. The verdict, the risk score and the draft are shown to you and wait: an assistant you can overrule, not an autopilot we have not built.

Questions and answers

Do I need a paid name to get the AI?: Yes. The AI features come with a paid name. A free account is a complete mailbox with webmail, IMAP and SMTP, but its mail is not read by any model and nothing from it is sent anywhere.
Does my mail leave your servers?: The analysis that happens to every message runs on our own GPUs, on an open-weights model we host. Only three things you start yourself go out: translation and reply drafts to Anthropic, the thread advisor to OpenAI.
How often can I run the advisor on a conversation?: Five runs a day per account, and 150 a day across the whole service. The other AI features are limited by a monthly message allowance instead.