Changelog

Return was answering new questions with an old answer

Type a reply cue, press Return, read the result, then change your mind and type a better cue. Pressing Return again re-inserted the first reply instead of writing a new one, and nothing on screen said so. Found by Mike Beanland, who tried three times, wrote down every step with screenshots, and sent the lot in.

• Return runs the cue you can see. With a finished reply on screen, Return used to mean insert it into Mail, whatever you had typed since. So editing your keywords and pressing Return quietly handed you the same reply again. Now an edited cue counts as a new request and Return writes the new reply. Emptying the box still means insert, and Shift+Return is still a new line.
• Updates can look after themselves. Settings, Updates now shows that MailTwin checks about once a day, and lets you switch that off. New: it can also download and install a new version for you, applied the next time you quit. It stays off until you turn it on. Two people spent days this week on bugs that were already fixed in a build they did not know existed.

Current provider defaults. Official catalogs checked 25 July 2026: Claude Opus 4.8 ($5/$25) and OpenAI GPT-5.4 ($2.50/$15) per million input/output tokens. Only an exact deprecated gpt-4o selection migrates; unrelated custom models stay untouched.

Two small things you told us about, fixed the same day

Build 92 shipped launch at login but only told existing users where the switch was; the founder's own first launch showed that a card describing a toggle is not an ask. And Alf of Faq-Mac sent a screenshot of our menu bar icon disappearing on a teal menu bar. Both are put right here.

• Open MailTwin at login — your call, right here. The switch below is the real thing, not a pointer to Settings. Flip it and MailTwin registers with the system's own Login Items, visible in System Settings like any other app; leave it off and nothing changes. It stays available in Settings → App either way.
• A menu bar icon you can see on any menu bar. The status icon is now a template image, so macOS adapts it to your menu bar — light, dark, or any wallpaper tint — instead of our fixed brand colours washing out on the wrong background. Spotted by Alf of Faq-Mac on a teal desktop, his fourth find in a week. The colourful sparkle still lives in the Dock and everywhere identity belongs.

Current provider defaults. Official catalogs checked 25 July 2026: Claude Opus 4.8 ($5/$25) and OpenAI GPT-5.4 ($2.50/$15) per million input/output tokens. Only an exact deprecated gpt-4o selection migrates; unrelated custom models stay untouched.

A slow mailbox no longer costs you the whole index

Voice learning used to give each batch of sent mail fifteen seconds and threw everything away when one ran over — which on a big IMAP mailbox could mean everything, every time. This is the third of three indexing bugs found by Alf of Faq-Mac, whose two screenshots did what our own tests had not.

• Indexing survives slow batches and keeps what it has. Each batch of sent messages now gets forty-five seconds instead of fifteen, and a batch that still runs over keeps every message already read instead of discarding the run. Re-indexing also finds the real Sent folder more reliably and reads newest-first, so the most recent — most representative — mail is learned first. If indexing ever failed for you, try it again in this build.
• Open MailTwin at login, if you want it. A new option in Settings → App (and at the end of onboarding) starts the assistant automatically after a restart, so it is already in the menu bar next to Mail. In MailTwin fashion it is off until you switch it on, applied through the system's own Login Items, and visible in System Settings like every other login item.

Current provider defaults. Official catalogs checked 25 July 2026: Claude Opus 4.8 ($5/$25) and OpenAI GPT-5.4 ($2.50/$15) per million input/output tokens. Only an exact deprecated gpt-4o selection migrates; unrelated custom models stay untouched.

The button that finds your Sent mail was lying

Diagnose sent mailboxes reported “(none found)” for every account, whatever the truth — on the one screen you open when indexing finds nothing. Found by re-reading a reviewer's screenshot.

• Diagnose sent mailboxes tells the truth again. Since build 88 this diagnostic called an internal routine that was never included in its own script. The resulting error was swallowed silently, so it answered “(none found)” for every account no matter what Mail actually held — sending people to chase a problem they did not have. It now reports real mailbox names and message counts, finds the folders Exchange hides from its own mailbox list, and answers in under two seconds.

Current provider defaults. Official catalogs checked 25 July 2026: Claude Opus 4.8 ($5/$25) and OpenAI GPT-5.4 ($2.50/$15) per million input/output tokens. Only an exact deprecated gpt-4o selection migrates; unrelated custom models stay untouched.

Voice learning finds your real sent mail

On many accounts — Exchange especially — the folder Mail lists is not where your sent mail actually lives. MailTwin could time out, index a years-old archive, or learn nothing. Found on a real 125-mailbox account and fixed end to end.

• The right Sent folder, read the right way. Discovery now asks Mail for the true Sent folder directly — including the hidden ones Exchange never lists — binds it so it cannot silently switch mid-read, and reads newest messages first, so the voice it learns is who you are now, not who you were years ago. Finding your folders went from timing out to a fifth of a second on a real four-account setup.
• Onboarding now teaches the voice. A new setup step explains that every account learns its own separate voice and offers to start the one-time, on-Mac reading — only when you press the button, with live progress and a skip. Voice Studio remains its permanent home.

Current provider defaults. Official catalogs checked 25 July 2026: Claude Opus 4.8 ($5/$25) and OpenAI GPT-5.4 ($2.50/$15) per million input/output tokens. Only an exact deprecated gpt-4o selection migrates; unrelated custom models stay untouched.

The panel stopped saying things that were not true

Tone and Length looked like labels with no settings. My Style kept offering to learn a voice it had already learned. The footer reported a fault when you had simply not picked an email yet.

• Tone and Length are controls again. They rendered as "Tone:" and "Length:" with nothing after them — no current setting, no arrow, nothing to suggest you could click them. The menu was clipping its own label, so the value and the chevron were cut off while the words survived. Both now show what they are set to, and open when you click.
• My Style stops asking for something it already has. It could not tell "this account has no voice yet" apart from "no email is selected", so with nothing selected it offered to learn your voice forever, however many times you had taught it.

Current provider defaults. Official catalogs checked 25 July 2026: Claude Opus 4.8 ($5/$25) and OpenAI GPT-5.4 ($2.50/$15) per million input/output tokens. Only an exact deprecated gpt-4o selection migrates; unrelated custom models stay untouched.

Voice learning now works in every language

If you run Apple Mail in anything other than English, MailTwin could not learn your writing voice at all. A reviewer in Spain found it. This build fixes it.

• MailTwin can find your Sent mailbox again. Voice learning reads your sent mail once to learn how you write. It looked for a mailbox called Sent — but Mail translates that name, so on a Spanish Mac it is Enviados, on a German one Gesendet, in Japanese 送信済み. Anyone running Mail in one of the other fourteen languages MailTwin ships in was quietly unable to learn a voice. All fifteen are now recognised, along with the folder names mail servers use.

Current provider defaults. Official catalogs checked 25 July 2026: Claude Opus 4.8 ($5/$25) and OpenAI GPT-5.4 ($2.50/$15) per million input/output tokens. Only an exact deprecated gpt-4o selection migrates; unrelated custom models stay untouched.

Errors that say what to do next

A developer testing MailTwin sent annotated screenshots of everything that confused him. He was right about all of it, so this build is his.

• No more “AppleScript error”. When something went wrong in Mail, MailTwin showed you the name of the technology it failed in, followed by an untranslated system message. It now tells you what happened, that nothing was changed, and what to try — and a test refuses to build if any error message names an implementation detail again.

Current provider defaults. Official catalogs checked 25 July 2026: Claude Opus 4.8 ($5/$25) and OpenAI GPT-5.4 ($2.50/$15) per million input/output tokens. Only an exact deprecated gpt-4o selection migrates; unrelated custom models stay untouched.

Every claim checked against the code

We audited every feature claim MailTwin makes — in the app, on the website, and in its privacy policy — against what the code actually does, and corrected everything that had drifted.

• The shortcut list now matches reality. Removing shortcuts in an earlier build left the cheat sheet, Help and onboarding advertising keys that did nothing. They now list only shortcuts that exist, and a test fails the build if that ever drifts again.

Current provider defaults. Official catalogs checked 25 July 2026: Claude Opus 4.8 ($5/$25) and OpenAI GPT-5.4 ($2.50/$15) per million input/output tokens. Only an exact deprecated gpt-4o selection migrates; unrelated custom models stay untouched.

An interrupted cleanup can no longer go unnoticed

If Inbox Cleanup is interrupted partway, the work waiting for your review used to announce itself in a menu entry that no longer exists. It now marks the Inbox Cleanup tab directly.

• The Inbox Cleanup tab tells you when something is waiting. A cleanup interrupted partway leaves work that needs your review, or that MailTwin must verify against Mail before trusting. That state used to be visible in the menubar menu, which no longer lists Inbox Cleanup. The tab now carries an amber dot for review and a red one for verification.

Current provider defaults. Official catalogs checked 25 July 2026: Claude Opus 4.8 ($5/$25) and OpenAI GPT-5.4 ($2.50/$15) per million input/output tokens. Only an exact deprecated gpt-4o selection migrates; unrelated custom models stay untouched.

Help opens, and there is only one menu

Choosing Help used to move focus and show nothing. The window was there all along — it had collapsed to no size. That is fixed, and the two menus MailTwin shows are now built from one list, so they cannot list different things again.

• Help actually opens. Choosing Help activated MailTwin and then presented nothing, because the Help window had collapsed to no usable size and was ordered in front at zero height. It now keeps its size, with a floor so it cannot collapse again.
• One menu, shown in two places. The menubar icon menu and the assistant panel’s gear menu were written separately and had drifted apart, listing different commands in different orders. Both are now built from a single list.

Current provider defaults. Official catalogs checked 25 July 2026: Claude Opus 4.8 ($5/$25) and OpenAI GPT-5.4 ($2.50/$15) per million input/output tokens. Only an exact deprecated gpt-4o selection migrates; unrelated custom models stay untouched.

Fewer things, and the ones that matter work

Learning your voice was quietly broken: pressing Re-index failed for everyone, in every pane. That is fixed. Alongside it, MailTwin now asks you to remember one keyboard shortcut instead of a dozen, and stops listing the same thing in two different menus.

• Learning your voice works again. Pressing Re-index used to fail with an AppleScript error, so MailTwin could never read your sent mail and never learned your voice. The script it used could not compile at all. It is fixed, and MailTwin now compiles its own scripts in testing so this class of failure cannot ship again.
• One shortcut to remember, not a dozen. The only shortcut you need is ⌘⇧M, because it is the one that works from inside Mail where your hands already are. Return still sends your keywords and inserts the finished reply. The rest are plain menu items now, and ⌘? is no longer claimed by two menus at once.
• Menus that agree with each other. Inbox Cleanup and Scratchpad are tabs, so they no longer appear as menu items too. The menubar menu and the panel menu now list the same things, and the + and − buttons in My Actions say what they do.

Current provider defaults. Official catalogs checked 25 July 2026: Claude Opus 4.8 ($5/$25) and OpenAI GPT-5.4 ($2.50/$15) per million input/output tokens. Only an exact deprecated gpt-4o selection migrates; unrelated custom models stay untouched.

One window for everything that is not writing

MailTwin does one thing: it writes your replies, in your voice, per mailbox. Everything that supports that now lives together in a single MailTwin Tools window, and everything that got in the way of it has been fixed.

• Inbox Cleanup and Scratchpad moved in. Inbox Cleanup and Scratchpad are tabs in the MailTwin Tools window rather than windows of their own. At the top level they suggested MailTwin was an inbox manager; it is a reply writer. Their keyboard shortcuts still work and now open the right tab.
• Voice Studio can finally start learning. Voice Studio now has the button that teaches it: start indexing your sent mail, watch the live progress and the sentence being learned, and stop whenever you like. Until now the controls lived in Settings, so the tab about your voice could not begin one.
• The menus work. MailTwin now installs a real menu bar, so Help opens Help instead of doing nothing, and Cut, Copy, Paste, Select All and Undo work in every MailTwin window.

Current provider defaults. Official catalogs checked 25 July 2026: Claude Opus 4.8 ($5/$25) and OpenAI GPT-5.4 ($2.50/$15) per million input/output tokens. Only an exact deprecated gpt-4o selection migrates; unrelated custom models stay untouched.

Now you can see what it saved you

MailTwin has always saved you time. From this build it tells you how much: a running total in Workspace, and a line in the trial banner so the price has something concrete to sit against. The estimate is deliberately cautious, and the whole calculation is written on screen so you can check it rather than take our word for it.

• A running total of the time you got back. Workspace now shows how much writing time MailTwin has saved you, and across how many replies. It stays hidden until there is a meaningful number to show, so it never greets you with a boastful zero.
• An estimate built to under-claim. A reply is measured against two minutes of writing it by hand, then the time you actually spent in MailTwin is subtracted. Drafts you threw away count for nothing, and regenerating shrinks the figure instead of doubling it, because one reply is one reply however many attempts it took.
• The number never leaves your Mac. The counter is stored on your Mac as a plain total, with no subject, sender, account, or message identifier beside it. Nothing about it is sent anywhere, including to us.

Current provider defaults. Official catalogs checked 25 July 2026: Claude Opus 4.8 ($5/$25) and OpenAI GPT-5.4 ($2.50/$15) per million input/output tokens. Only an exact deprecated gpt-4o selection migrates; unrelated custom models stay untouched.

Nobody should need a manual

Setting up MailTwin now walks you through every step: it shows you where to click, catches your free Groq code from the clipboard, and moves on by itself the moment you flip a switch. Reply and My Style explain themselves, and errors say what to do next.

• Getting your free code, guided. Getting your free Groq access code is now a guided walkthrough — MailTwin even catches the code from your clipboard the moment you copy it, so there is no field to hunt for.
• Screens that move on by themselves. Permission screens show you exactly which switch to flip and move on by themselves the moment you flip it — no “click Continue when you’re done”.
• Reply and My Style say what they do. Reply and My Style now say what they do: My Style writes in your voice, and tells you when it still needs to read your sent mail once.
• Every error ends with the next step. Every error message now ends with what to do next, in all fifteen languages.

Current provider defaults. Official catalogs checked 25 July 2026: Claude Opus 4.8 ($5/$25) and OpenAI GPT-5.4 ($2.50/$15) per million input/output tokens. Only an exact deprecated gpt-4o selection migrates; unrelated custom models stay untouched.

Expand keywords, working the way it should

Expanding keywords in Mail’s reply window now works: MailTwin selects your keywords the way Mail itself does, still proving the selection before writing. When something does go wrong, you finally see the real reason instead of a generic message.

• Expand keywords completes. Expand keywords now completes in real Mail reply windows — the quoted message is still never rewritten, so its indentation and formatting stay exactly as Mail made them.
• Proof before a single character. MailTwin still proves it has selected exactly your keywords before a single character is written; if it can’t, it writes nothing and leaves your keywords untouched.
• The real reason, every time. Failure messages now carry the real reason and a diagnostic code, including when Mail is still running an older MailTwin extension.

Current provider defaults. Official catalogs checked 25 July 2026: Claude Opus 4.8 ($5/$25) and OpenAI GPT-5.4 ($2.50/$15) per million input/output tokens. Only an exact deprecated gpt-4o selection migrates; unrelated custom models stay untouched.

Your quoted message, exactly as you left it

Expand keywords now replaces only the keywords you typed. The quoted message below them is never rewritten, so Mail’s own indentation and formatting survive untouched.

• The quote is never rewritten. Expanding keywords no longer touches the quoted message — its indent bar, links, and formatting are preserved because those bytes are never written.
• Proof before a single character. MailTwin proves it can select exactly the keywords you typed before writing anything. If it can’t, it writes nothing and tells you, and your keywords stay exactly as you typed them.
• Your signature, untouched. Your signature is preserved the same way — by never being rewritten.

Current provider defaults. Official catalogs checked 25 July 2026: Claude Opus 4.8 ($5/$25) and OpenAI GPT-5.4 ($2.50/$15) per million input/output tokens. Only an exact deprecated gpt-4o selection migrates; unrelated custom models stay untouched.

Type it rough. Send it right.

MailTwin is radically simpler: a first-run panel that is just keywords and Return, an onboarding that shows the magic before it asks for anything, and keywords you can now type straight into Mail’s own reply window.

• Keywords, straight into Mail. Type a few keywords in Mail’s own reply window, press the MailTwin button and choose Expand keywords. They become the finished reply in place — above the quote, with your signature untouched.
• A first panel with one job. New Macs start in a simple panel: type what you want to say, press Return, press Return again. More… reveals every other action, and Settings → App → Behavior carries a Simple panel switch. A panel you already use is unchanged.
• Setup shows you the result first. The new welcome runs three sloppy keyword examples and the replies they turn into, clearly labelled Preview — example output. It makes no AI call and touches no network.
• A free Groq key in about three minutes. Setup opens the Groq page, you paste the key back, and MailTwin checks that it answers before moving on — or start instantly on-device with Apple Intelligence. No provider is ever activated silently.
• Fail-closed, exactly as before. If the AI invents a number, a date, or a sign-off, nothing is written and your keywords stay exactly as you typed them. Every expansion is receipted locally in Activity.

Current provider defaults. Official catalogs checked 25 July 2026: Claude Opus 4.8 ($5/$25) and OpenAI GPT-5.4 ($2.50/$15) per million input/output tokens. Only an exact deprecated gpt-4o selection migrates; unrelated custom models stay untouched.

Your replies, at the speed of Return

Build 75 makes the reply loop dramatically faster: Return generates and Return inserts, Mail prepares your reply window while the AI writes, and your indexed voice becomes the default — every action still explicit, receipted, and undoable.

• The whole reply on the keyboard. Press ⌘⇧M, type a few keywords, press Return — and Return again to insert. The hotkey lands with the keyboard already in the composer; Shift-Return still makes a new line.
• Insert is instant. Mail prepares your reply window while the AI writes, so Insert has nothing left to wait for. Prepared windows are receipted in Activity, tidied away if you change your mind, and the behavior has a Settings toggle.
• Your voice is the default. Once an account’s voice is indexed, Reply (my style) becomes that account’s default reply — and MailTwin keeps honoring a manual override you set.
• Reply all, exactly when needed. A Reply-all toggle appears only when the original message has more than one recipient. Mail opens its own reply-all window and the AI addresses the group.
• Insert & Archive closes the loop. One click inserts the reply and files the original — only after the insert actually succeeded. Receipted in Activity and undoable straight from the toast.

Current provider defaults. Official catalogs checked 23 July 2026: Claude Opus 4.8 ($5/$25) and OpenAI GPT-5.4 ($2.50/$15) per million input/output tokens. Only an exact deprecated gpt-4o selection migrates; unrelated custom models stay untouched.

A calmer place to run your mail day

The fast floating assistant now has a local workspace beside it: a place to see what needs attention, review exactly what MailTwin did, and choose the context it may use.

• Today + Activity. See due and waiting items, replies confirmed by your last Refresh, quick actions, recent work, and local receipts labelled as success, failure, suggestion or learned choice.
• Reply-aware Waiting On. Press Refresh in Workspace to run one bounded Sent/incoming-header metadata check. Replied requires exact RFC ancestry. Failure changes no status and triggers backoff; nothing runs automatically, in the background, or at launch.
• Three Inbox Cleanup views. Clean Now stages a plan; Keep Clean lets you inspect or forget learned sender choices; Cleanup History shows receipts and exceptions. Nothing moves in the background.
• Multilingual Voice Studio. Tune greeting, brevity, directness, warmth, never-use phrases and chosen examples separately for each account, with controls in all 15 languages.
• Editable, reviewable compose. Edit generated text before inserting it. Improve Draft can show a red/green comparison, and rewrite validators still reject invented subjects, sign-offs and facts.
• Structured action review. Edit title, owner, due date and Reminders/Calendar destination, deselect anything, then export the reviewed set once; duplicate lines are removed.
• Cited Relationship Memory. Search the local Sent index, inspect source/date/recipient citations, open the exact message when its locator is available, and explicitly add at most three excerpts.
• My Actions. Save reusable instructions with notes and optionally assign one unique ⌘⌥1–9 shortcut. The menubar or shortcut only opens the instruction visibly in the assistant; it never runs AI or sends mail automatically.

Sent-mail indexing is always your action. New installs default to 200 messages per account (up to 2,000). Index or Re-index runs locally only when you start it, in five-message Mail batches. Mail may be briefly busy; Stop prevents later batches but cannot interrupt the Apple Event already running. Nothing indexes automatically — not at launch, not on a timer, not in the background; the onboarding invitation and Voice Studio start it only when you press the button. After updating, re-index once only if older Relationship Memory results need exact source details or View in Mail.

Current provider defaults. Official catalogs checked 16 July 2026: Claude Opus 4.8 ($5/$25) and OpenAI GPT-5.4 ($2.50/$15) per million input/output tokens. Only an exact deprecated gpt-4o selection migrates; unrelated custom models stay untouched.

Meet your calmer inbox

Your choice of AI and your separate work and personal voices are still the heart of MailTwin. This release also helps you clear the pile before you write — with you firmly in charge.

• New: Inbox Cleanup. Review up to 150 messages from one account Inbox using metadata only. Keep, mark read, archive, junk, or trash one message or a whole sender group; bodies and attachments stay closed.
• No mystery automation. Past choices may create suggestions, but they are visibly highlighted, easy to clear, and nothing moves until you review and apply the plan.
• MailTwin where you write. The real MailKit compose extension now puts Improve and My Style inside Mail’s compose window, with metadata-only context until you explicitly act.
• Improve and My Style that can’t go rogue. In the compose window, Improve makes small, careful edits and My Style reshapes your wording into your own voice — both keep your signature, links, numbers, and dates intact. If the AI invents a subject or a sign-off, or changes a link, a number or a date, MailTwin rejects the rewrite and your draft stays untouched.
• Trust polish everywhere. Exact-message “View in Mail,” screen-aware windows, visible staged actions, careful rechecks, receipts, recovery, verification, and Undo.
• A fresh 30-day trial for early users. If you installed before this release and never bought a license, build 73 grants one new 30-day full trial — enough time to try Inbox Cleanup and the compose extension properly.