Cue

Demo

One minute, and you will know whether you want it.

Features

It already knows the context.

• Two gestures, and that is the whole app

  One shortcut opens the panel beside your cursor: press Return for the main action, or a number for any of the rest. The other shortcut opens nothing at all and writes the reply straight into the field you came from. Cue never asks you to leave the conversation you are in, and it never opens a window you have to copy out of.

• A memory that crosses conversations

  Most assistants forget the moment the window closes. Cue keeps a memory of the people you talk to and the facts of your own life, and carries it into every app. Tell one person you have plans tomorrow at eight, and when another person asks about tomorrow evening, the answer already accounts for it. The memory is browsable, editable, and yours to delete a line at a time.

• Every reply shows its receipts

  Under each result Cue lists exactly what went into it: how many messages it read, whose voice it matched, which remembered facts it used, and what it knew about your schedule. An eye icon shows the raw context it pulled from the app. Nothing about the reply is invisible, so you always know why it said what it said.

• Actions you write yourself

  Rephrase, Polish, Formal, Casual, Shorten, and Translate come built in. Add your own beside them, each with its own instruction, its own model, and its own shortcut, so the one you use twenty times a day is a single keystroke away.

• You can watch it write

  Nothing spins in place while you wait. The edge of the panel lights up as the request goes out, and the reply streams in word by word, so you can start reading before it finishes and press Escape the moment you know it is wrong.

• Bring your own key

  Anthropic, OpenAI, Google Gemini, or any OpenAI-compatible endpoint, and Ollama if you would rather run a model on your own machine and stay offline. Keys are stored in the macOS Keychain, requests go straight from your Mac to the provider, and there is no server of mine in the middle.

• It reads the room

  Cue takes the conversation from the app itself, through the Accessibility API: who you are talking to, the recent messages, and your draft. No screenshots and no scraping of the screen. It recognizes WhatsApp, Messages, Mail, and Slack by name, and falls back gracefully in everything else.

• The memory stays here

  Everything Cue remembers is a file in your own Library, on your own Mac. It is never uploaded, never synced to a service of mine, and never used to train anything. Password fields are refused outright: the panel will not even open over one.

• Small and native

  A menu-bar app of a few megabytes, written for macOS with the platform's own tools. Developer ID signed and notarized, built for Apple silicon, and requires macOS 14 or later.

Screenshots

Scroll sideways to see more

Version history

• 2.8.328 September 2026

  Cue does less work in the background, so your Mac stays quieter. Proofreading in Slovak and other languages macOS cannot check no longer offers English spellings. Bug fixes. With thanks to Roman for his report.

• 2.8.226 September 2026

  Tab and the Right arrow work normally in password windows, terminals and code editors. Grey text blinks far less in Claude and other chat apps, and underlines stay in place. In Apple Mail, fixing everything no longer adds a second full stop. The fix card now says it is the mouse pointer that picks the mark. With thanks to John for his report.

• 2.8.126 September 2026

  Smarter email follow-ups: Cue's Email section lists only what someone still owes you, and a reply sent as a new email counts. Works with Proton Mail Bridge. Suggestions in Slovak and other languages macOS cannot check, with a new switch in Settings. Tab works normally on web pages, and proofreading no longer puts "?" in the middle of a sentence. Bug fixes. With thanks to John, Adrian and Roman for their reports.

• 2.8.024 September 2026

  Promises, Email and Prompts redesigned. Promises sorted by due date, with Done and Snooze. Proofreading catches "i" and "Ive" every time, and no longer adds a full stop mid-sentence. Cue now works in Thunderbird. Settings reorganised and easier to follow. Smoother windows and panel. With thanks to John for his reports.

• 2.7.923 September 2026

  Undoes what 2.7.8 broke: corrections apply again, the grey text is steady again, and Cue is back to using almost no processor. Suggestions start after your first word, and replies in chats and email use what the other person actually said. Tab takes a green correction even after a line break. Dictation lands in a second or two and never pastes twice, and Proton Mail Bridge connects. With thanks to Ulf, Paul and John for their reports.

• 2.7.822 September 2026

  Proofreading applies when you ask: the fix lands on the first hold, not the second. Steadier suggestions, so the grey text sits still while you type through it. Bug fixes.

• 2.7.720 September 2026

  Better suggestions and speed. Works in more apps and languages. Bug fixes. With thanks to Ulf, Andreas, Dirk, John, Paul and Roman for their reports.

• 2.7.518 September 2026

  Fixed Tab not working in some apps — Cue was still holding the key after you left a text field. The Cue mark now appears beside text fields in more apps. The mail list shows the real message, with no more rows of dashes, forwarded headers or signature images, and long legal footers fold away behind one line so you read the message rather than the disclaimer. Senders have their own initial and colour, Inbox and Sent show how much mail Cue holds, and search is a proper search field you can clear in one click. A message with no text of its own now says so instead of showing you the footer. Email settings tidied to match the rest of Settings. New in Settings, About: Create a report, which gathers what went wrong into one file you can email, asking first whether to include your text. Thanks to Ulf and Andreas for reporting the first two.

• 2.7.416 September 2026

  Suggestions are now about what is on your screen — the page you are working in, and the ones you were reading a moment ago — instead of only the words you have already typed. Cue no longer echoes your own draft, its own instructions, or your Settings writing-style note back at you as a suggestion. Accepting a suggestion no longer mangles the word you were part-way through, repeats the word you just typed, or splits a number. Proofreading capitalises proper names, carries a capital along with any full stop it adds, and finishes the whole fix rather than reporting that some of it went in. Arabic and other right-to-left fields no longer pick up double spaces or stray punctuation. The AI settings are now two tabs — one for your keys and downloads, one for which model does which job — with a badge on each model showing what it is best at and which suits your Mac. Five models to choose from instead of ten, and nothing you have already downloaded is removed or switched. Cue at rest went from 46% of a CPU core to under 1%.

• 2.7.314 September 2026

  Suggestions no longer flicker: the grey text stays where it is while you type instead of vanishing and coming back a moment later, and it keeps up with fast typing, appearing as you write rather than only when you pause. The continuations themselves are better too. A suggestion now follows the words you just typed instead of starting something new, and it no longer stops in the middle of a phrase. Accepting one with Tab works when you press it, including in Slack, Discord and the other apps built the same way, where the key could previously fall through into the document. A correction can no longer overwrite your sentence with the wrong text. The circle beside your text field no longer stays on screen after a check has finished. And the Cue mark beside your text field is on by default now, so the circle showing Cue reading your writing is there from the start; you can still switch it off from its own menu. With thanks to Ulf, whose detailed reports found most of what is fixed here.

• 2.7.214 September 2026

  Holding the fix key now fixes your text. It could answer with a small sound and do nothing, or say you had kept typing when you had not touched the keyboard; underneath, Cue was giving an app a tenth of a second to hand over what you had written and treating silence as proof that you had changed it, and a busy app simply needs longer. It now waits properly, asks again, and when an app really will not answer it says so instead of blaming you. That one change fixes Claude, Slack, Discord and the other apps built the same way, where holding the key failed most and suggestions got stuck mid-sentence. Suggestions also stop flickering, because grey text no longer gives up the moment a heavy app is slow to answer, and it no longer sits on top of your writing: type something different from what Cue predicted and the suggestion goes at once instead of freezing where it was. Two suggestions no longer appear on top of each other either, since a ready reply offered into an empty field was landing on the faint hint the app draws there, and now sits just below the cursor. A circle around the star shows Cue reading your text, filling as your writing is checked and disappearing when it is done. Checking itself is much faster: Cue was re-reading its own instructions before every single sentence, at more than two seconds a time, and keeps them ready now, so a page that took twenty-six seconds takes sixteen. Corrections you never asked for are gone, because Cue no longer touches file names, web links, version numbers, email addresses or user names, which it used to quietly take apart while fixing the sentence around them. Pointing at a mistake now gives you that mistake rather than the one beside it, Escape skips a single correction instead of clearing them all (press it again on the same word and Cue leaves that word alone for the rest of what you are writing), a long document is marked where you are actually looking instead of only near your cursor, and corrections come back when you return to an app rather than staying gone until the next check.

• 2.7.113 September 2026

  Suggestions that fit what is in front of you. Cue now weighs several possible continuations against the mail, page or chat on your screen before it shows one, instead of taking the first words the model produces; across twenty everyday cases it used the right detail from the screen eighteen times where it used to manage eleven, and it is no slower, the first word still arriving about a tenth of a second after you stop typing. It also speaks up more often: the checks meant to stop Cue repeating you were throwing away good suggestions too, so it sat silent while you typed, and they now catch real repetition only. The strange capitals are gone, the ones that produced “NONULL” or a name written half in capitals, which came from a missing first token when your text was read and from misread words on screen; a suggestion carrying a mangled word is now refused rather than shown. The grey text no longer flickers as you type and stays where it is when the cursor moves along the line, hammering Tab no longer gets stuck, a run ending at the end of the sentence and never leaving a stray tab behind, and a suggestion after a line break appears on the line you are on instead of the one above it. Cue can also see the other window of the app you are writing in: it used to skip every window of that app, so the message you were replying to, sitting right beside you, was invisible to it, and it now reads windows in the order they are stacked on screen and skips only the one you are typing in. Proofreading starts sooner, the first check after you move into a field taking under a second instead of nearly two, and asking Cue to check no longer answers that it finds nothing to fix merely because a suggestion was being written at that moment. You can bring your own model from Settings, where a button takes any .gguf file you already have, and two higher quality builds of the built-in model and the larger E4B model have joined the list. The Context Inspector now shows the text after your cursor, which Cue was already sending to the model but never showed you.

• 2.3.130 August 2026

  The release that goes after the two things you actually noticed. Holding to fix everything now works every time: the gesture could quietly do nothing, and it depended on what you had written rather than on anything you did, so a line ending, an invisible mark macOS puts around English words inside Arabic, or an accent typed one way instead of another was enough to stop it. Eight separate places in the app were asking whether the text was still the same and disagreeing with each other, and when they disagreed the fix was dropped in silence, after the marks had already gone from the screen. There is now one answer to that question and all eight use it, and a fix that still cannot be applied says so out loud instead of leaving you holding a key that does nothing. The same disagreement is why the underlines sometimes vanished a second after appearing. Arabic and English in one sentence is checked properly now: Cue used to hand the whole sentence over at once, the model answered by translating the Arabic into English, the answer was thrown away as a rewrite, and every real mistake in that sentence went unmarked. Each language is checked on its own, and certain Arabic corrections, hamza and the feminine ة among them, no longer depend on the model at all. Dictation keeps your words: Arabic speech was being translated into English and typed into your field, a long dictation could be cut off halfway with nothing to tell you, and a model refusal could be typed in as though you had said it. None of those can reach your text now, because if the tidy-up cannot be trusted then what you actually said goes in instead, complete. Dictation is also cleaned by default: the "uh"s, the false starts and the part you said before correcting yourself are removed before the words land, and you can turn that off in Settings under Dictation to insert exactly what was heard. Arabic now reads right to left in the writing panel, which laid it out left to right with the punctuation at the wrong end. Settings opens instantly, where opening the AI section used to stall the app for about half a second every time it redrew because it re-read a whole downloaded model file to check it. A Gmail app password is one click from the field that needs it, and the Ollama Fetch button says when nothing came back instead of looking dead.

• 2.3.026 August 2026

  Cue now has a panel for your mail. Press the mail shortcut on the message you are reading and Cue opens with the thread it belongs to, a draft written from it, and the tone you want it in: choose which of your accounts it goes out from, turn your signature on or off, and mark yourself busy so every reply says you will come back to them later, in your own words. Mail you mark sensitive is only ever shown to a model running on your own Mac. Memory keeps the facts worth keeping: the limit on what Cue remembers now drops the weakest facts rather than the oldest, questions reach facts filed under a topic instead of stopping at the untagged ones, you can make Cue forget everything it learned from one app, and deleting everything really does leave nothing behind. Replies land in the right thread, with mail matched to its conversation more accurately, threads read oldest first with your own message marked, and a mailbox that fails to sync once is no longer dropped. Your second language holds steady, with sensible defaults on first run, Persian and Spanish behaving, Arabic digits staying Arabic, and diacritics left alone unless you ask for them. Arabic is corrected as you type, hamza and taa marbuta put right in place, and Arabic typed in Latin letters comes back as Arabic. Holding to apply a fix no longer stalls partway, and grey suggestions no longer arrive shouting in capitals. And if Cue cannot read one of its own files it now sets it aside and tells you, instead of saving over it, so a single unreadable file can never become permanent loss.

• 2.2.022 August 2026

  Cue knows what day it is, and it can now hear you. Every reply carries the current date and time, message timestamps survive from WhatsApp and Messenger, and remembered facts read like a timeline, so a Tuesday message is never answered on Friday as if it had just arrived. Set your working days and hours in Settings, and replies about availability decline slots on your days off and name your actual next working day. Voice to text is built in: hold Option anywhere, speak, release, and the voice panel appears with what you said, ready to insert as spoken or rewritten through a voice action first, landing at your cursor even in web apps with an existing draft, with transcription running entirely on your Mac. Proofreading has steadier hands: fix-all no longer plants the cursor at the top of web fields, a hovered fix always answers the key, Caps Lock and the Globe key no longer swallow a press, and blank lines survive every fix-all byte for byte. Arabic is faster, and Arabic typed into an English-labelled thread is no longer rejected as the wrong language. A second language now gets its own panel: pick one in Settings and selecting text in it shows that set of actions. Also: an App Blocker section controls proofreading, suggestions and memory per app in one place, and 22 fixes from the 2.1.0 full review landed.

• 2.1.09 August 2026

  Cue can now read your work email. Sign in to your Exchange account in Settings and Cue keeps a private, read-only archive of your recent inbox and sent mail on your Mac; when you write in Apple Mail, Outlook, Thunderbird, Spark or eM Client, replies draw on the messages that belong to the thread. The archive is only ever offered to the model running on your Mac: pick a cloud model and it is refused at the source, nothing on the server is changed or deleted, and you choose how much Cue keeps, from one to six months. Memory now files what it learns under topic tags, in English or Arabic, so a conversation about food brings back the facts about food, not the facts that happen to share a word. Proofreading catches what it used to miss: a first word without its capital, names like Paris, a lone lowercase i, and a missing Arabic question mark are all flagged now, and three silent ways a check could give up and mark nothing have been closed. Spacing survives every fix everywhere: fixes are applied from the app's own copy of your text and verified letter for letter, so blank lines and paragraph breaks stay exactly as you typed them. And accepting a suggestion in the middle of a word no longer capitalises the rest of it. This release also compiles the development-only test hooks out of the shipped app entirely, so no other program on your Mac can start Cue in a mode that makes it report what it reads.

• 2.0.24 August 2026

  Proofreading, rebuilt around trust. A model answer that describes your text instead of correcting it is now refused outright, whatever its phrasing, so commentary can never be offered as a fix. Applying fixes no longer flashes the whole text in web apps: Cue replaces only the span it corrected, after proving the app really moved its selection, which also means blank lines and spacing outside the fix survive untouched. Doubled last letters are gone. Lists numbered 1-, 2- keep their numbering. A quiet pop confirms every applied fix, single or all at once. Holding to force a scan now always scans: it overrides the pause after an apply and works on short text. Rewrites keep who-did-what intact, so a question about their call can never come back as a question about yours. And actions reread the conversation just before they run, so the reply answers the newest message, not the one from when the panel opened.

• 2.0.12 August 2026

  Five fixes, all of them things you would have noticed. Taking words with Tab no longer jumps out of the field: pressing it quickly used to send the second press to the app underneath, which in Slack or a browser moves focus to the send button, so Cue now holds the key across the gap while it writes. Hold to fix everything works on the first hold; accepting a suggestion changed the text without refreshing the underlines, so the ring would fill and nothing would happen until you forced a scan and held again. Blue underlines actually arrive: the pass that suggests smoother wording was being abandoned whenever the first pass finished before you paused, which was almost always, so every sentence except the last one went red and could never go blue. Ghost text is the right size in apps that hold two text sizes in one field, such as a note with a title and a body, where Cue could draw its suggestion at the title's size over your body text and stay wrong until you clicked away and back. And memory now answers questions about specific things: ask about a particular day, in English or Arabic, and Cue finds what is on that day by date rather than by matching words, while asking directly about something it knows, a favourite food or a show, no longer causes the fact to be discarded for being too similar to the question.

• 2.0.01 August 2026

  The version Cue was being built towards. Suggestions now run on your Mac: download a model and Cue runs it inside itself, with no server and no API key, and each job picks the model that suits it, so the quick one answers while you type and the careful one writes the replies. Bring your own Anthropic, OpenAI or Gemini key instead if you prefer. New: proofreading. Stop typing and Cue checks what you wrote, sentence by sentence, marking mistakes in red and clumsy phrases in blue; point at a mark to see the whole message with every change in place, press one key to fix it, or hold that key until a ring fills to fix everything in a single edit that keeps your line breaks, bullets and spacing exactly as they were. Suggestions are now grey text drawn at your cursor in the field's own font, taken a word at a time or all at once, with every shortcut yours to rebind. Cue now works properly in Electron and browser apps, which report a caret that is not where they say it is. Downloaded models are verified against a published SHA256 before they are accepted. Quitting no longer crashes. Blocked apps are now refused by suggestions and proofreading too, not only by memory, and deleting everything Cue has learned really does delete everything. And Cue is now a product you can buy: seven days free, then a licence.

• 1.3.027 July 2026

  Cue now finishes your sentences. Pause for a moment while typing and the next few words appear where your cursor is; press Tab to take them. Click into an empty message box and a whole reply is already waiting, drafted from the conversation and everything Cue remembers about the person you are writing to. New Context Inspector shows exactly what Cue read from the screen, which memories it chose, and the precise prompt it would send. Conversations are fenced in the prompt and framed as a recording, so a message trying to give the model instructions is treated as something somebody typed rather than as a command. And Cue updates itself from now on, signed and pinned.

• 1.2.026 July 2026

  First public release. Cue lives in your menu bar and works inside the apps you already type in: WhatsApp, Messages, Mail, Slack, anywhere there is a text field. One shortcut opens a panel of actions, a second skips the panel and writes straight into the field. Underneath both is a memory that works across conversations. Every reply shows its receipts, so you can see which memories, whose voice, and what part of your schedule went into it. Bring your own key for Anthropic, OpenAI, or Google Gemini, point it at any OpenAI-compatible endpoint, or run it fully local with Ollama. Keys live in the macOS Keychain and the memory never leaves your Mac.

Common questions

• Which apps does Cue work in?

  Any Mac app with a text field, including WhatsApp, Messages, Mail and Slack. Cue lives in the menu bar and opens next to your cursor wherever you are already writing.

• Do I need my own API key?

  Only if you use a cloud model. Cue is bring-your-own-key for Anthropic, OpenAI or Google Gemini, and works with any OpenAI-compatible endpoint. You can also run it fully local through Ollama, which needs no key at all. Keys are stored in the macOS Keychain and requests go straight from your Mac to the provider.

• Can Cue work offline?

  Yes, if you point it at a local model through Ollama. Nothing then leaves your Mac at all.

• How does Cue see what I am replying to?

  It reads the conversation from the app itself through the macOS Accessibility API: who you are talking to, the recent messages and your draft. It does not take screenshots.

• Is there a free trial?

  Yes. Cue runs free for 7 days with every feature included, and asks for no account and no card.

• What does Cue need to run?

  macOS 14 or later, on an Apple silicon Mac. Cue does not run on Intel Macs. It is distributed directly, signed and notarized with a Developer ID.

Related apps

From the same studio

• Emojist

  Handles the other half of typing fast: the emoji, inserted from the keyboard without a picker.

• مِداد

  Does for Arabic what Cue does for English, entirely offline and free.

Privacy