MailUGet early access

Now in private beta

Your inbox,
already handled.

You do not have 400 emails. You have four decisions, buried in 400 emails. MailU does the triage, writes the replies, and — when you trust it to — sends them.

Private beta. Auto-send stays off until you turn it on. See how it decides

app.mailu.ai/today
This week's mail400 messages
  • ApproveBudget request for the team offsiteFinance
  • ReplyCan we move the launch to next week?Product team
  • DeclineThursday 3pm clashes with your reviewCalendar
  • ChaseSigned contract, no reply since MondayNew client

396 filed without you · 4 waiting on you

Works with the mailbox and the model you already have

The problem

Every mail client sorts by when it arrived.

None of them sort by whether it is your problem. So the work was never the reading — it is the triage before the reading, done by hand, every morning, forever.

The drafting engine

It writes the reply

Six stages, one of which costs a frontier model. Newsletters are identified from raw headers before any model runs — that is most of a real inbox, handled for nothing.

  • Triage — what each email wants, how urgent it is and who it involves, worked out on your fast model

  • Research → Draft → Evaluate — a real agent graph with a bounded loop-back; only the text a human reads runs on the frontier model

  • Three alternatives — one draft makes your job editing. Three make it choosing

  • Learns from the diff — the untouched AI text is kept beside your edit — that gap is the strongest signal in the system

Autonomy

You keep the switch

Auto-send is off until you turn it on. On, it still has to clear your threshold, survive the intent blocks and outlast an undo window you control.

  • Confidence threshold — default 0.85 — below it, a human decides

  • Hard-blocked intents — legal, HR and credentials can never be auto-sent, checked at drafting and again at dispatch

  • Burst guard — one person, one email at a time — the rest defer, never drop

  • One master switch — stops drafting, sending and chasing instantly, including drafts already queued

Chief of Staff

Chat that acts

One surface that answers and acts. Ask it anything, tell it to do anything, speak to it and hear the answer — audio is never written to disk, database, or a log line.

  • Read runs silently — search mail, read a message, check availability

  • Mutate offers an undo — archive, snooze, save a view, draft a reply

  • External needs your tap — sending and booking are never run by the loop

  • Tiers read from code — a prompt-injected tier still hits the confirmation gate

How it works

How a message becomes a reply.

Six stages. Only one of them spends a frontier model, because only one of them produces something a person reads.

  1. 01

    Blackhole

    Newsletters and bulk mail are identified from raw headers before anything else runs. Free, instant, and it is most of a real inbox.

    No model
  2. 02

    Triage

    What the email wants, how urgent it is, who and what it mentions, and any money involved. If the model is unsure, the message is left for you rather than guessed at.

    Fast model
  3. 03

    Attachments

    PDF, CSV, JSON and text parsed into the message context and searchable immediately. Sending that text to a model is a separate switch, off by default.

    No model
  4. 04

    Threading

    Every message joins its conversation, and related threads merge into one topic. Long threads keep a rolling summary, so a model only ever reads that and the latest few messages.

    No model
  5. 05

    Research → Draft → Evaluate

    A real agent graph with a bounded loop-back. Research and Evaluate run on the fast model; Draft runs on the frontier one, because it is the only output a human reads. Evaluate scores confidence and lists what it could not verify.

    Fast, then frontier
  6. 06

    Dispatch

    It clears your confidence threshold and your autonomy settings, or it waits for you. Hard-blocked intents are re-checked here, not just when the draft was written.

    No model

27s

From an email arriving to MailU having read it

0.85

Confidence a draft must reach before it may send itself

180s

To undo anything MailU sent on its own

3

Drafts to choose between, not one to rewrite

Security and control

Built for the inbox you can’t afford to get wrong.

An assistant with access to your mail has to earn it. These hold on every account, from the first day.

Nothing sends without your say

Auto-send is off until you switch it on. Legal, HR and password mail can never send itself, however confident the draft.

Your AI key, your bill

Bring Anthropic, OpenAI or Google. See what every step costs, and nothing is billed per message.

Encrypted where it rests

Mailbox passwords, sign-in tokens and API keys are encrypted at rest. Your browser never holds a readable session token.

A record of everything

Every action MailU takes on its own is written to an audit log you can read, and what it learned can be erased in one step.

Questions

The things worth asking first.

MailU is in private beta. Leave your email on this page and you’re on the early-access list; we bring people in in small groups so every mailbox gets proper attention, and we’ll write to you when your spot opens.

Private beta

Four decisions, not four hundred emails.

MailU is opening to early adopters in small groups, so every mailbox it handles is handled well. Leave your email and we’ll let you know the moment your spot opens.

No spam — one email now, one when you’re in. Already have access? Sign in