Talking to Moe

Four ways in, and they all end in the same place: Moe listening, the notch lit green.

• Hold right ⌥ and talk. Let go when you are done. This one always works and never mishears its name, because it does not listen for one. It needs Accessibility; until that is on, Moe says so wherever you look, on the island when you hover, at the top of the open panel and above the chat box, each with a Fix button, and the warning goes the moment it is on. Settings › Voice › The key you hold to talk changes the key to right ⌘, right ⌃ or fn / 🌐, and the new key works straight away. Right ⌃ is only on some keyboards, and fn works only if macOS is not already using it: set System Settings › Keyboard › Press 🌐 key to Do Nothing, or macOS will act on it too. Everywhere Moe names the key, it names the one you picked. A key held as part of a shortcut doesn't start Moe: hold fn and press Delete, or right ⌘ and Tab, and the press is dropped: nothing is recorded, no picture is taken, and a reply it paused carries on.
• Double-tap the talk key and talk with nothing held, for when your hands are busy or you have a lot to say. The notch breathes and says Listening — tap to send. Pause as long as you like to think: Moe never decides you have finished. Tap the key once and what you said goes to Moe, exactly as if you had held the key and let go; if you said nothing, nothing is sent. Esc drops it without sending, and so does locking the screen or the Mac going to sleep. This is only when Moe is idle: a double-tap while Moe is talking or working stops everything.
• ⌃⌥Space to start and stop, hands free. It needs no permission at all. The row for it in Settings › Voice, Press ⌃⌥Space and ask, offers ⌥Space, ⌘⌥Space and ⌃Space instead, for when another app already owns it. A new shortcut takes effect when Moe restarts.
• Say "Hey Moe": Settings › Voice › Say "Hey Moe", off by default. The menu-bar icon has the same switch, Always Listening ("Hey Moe").

Two switches under While you talk go with the key. Moe sees my screen when I ask is on: a picture of the screen under your cursor, taken as you press the key, goes with your words, so "what do I do here" just works. It is never saved. Move my mouse to what Moe points at is off: Moe only draws a ring around the thing it means, and it never clicks.

Moe starts reading your turn back the moment you stop talking, not the moment it decides you have finished, so when you stay quiet the answer starts sooner. Say anything more and the whole turn is read again; nothing is ever taken from a half-heard sentence.

Staying in the conversation#

Keep the conversation going, in Settings › Voice, is off to start with: Moe answers one question and stops listening. Switch it on and, when Moe finishes answering, it keeps listening for a few seconds so a follow-up needs no key and no name; you just keep talking. How long it keeps listening under it chooses 8, 12, 16, 20 or 30 seconds; it starts at 8. Longer is friendlier in a quiet room and worse in a busy one, because everything said in that window is something Moe has to judge.

Also after I hold the key decides whether the window opens after a hold-to-talk turn too, or only after the name and the shortcut. Holding the key is meant as a one-off question, so this is off unless you want otherwise. Both rows, and When it goes quiet below, are greyed out until the conversation switch is on.

Not everything you say is for Moe#

The hard part of an always-listening assistant is not hearing you; it is working out which of the things it hears were meant for it. In the follow-up window Moe weighs several signals rather than matching a phrase: how long ago it spoke, whether it had just asked you a question, how loud you are compared with the room, whether the voice is the one that started the conversation, and what the words themselves suggest. In English it reads the words closely; in other languages it goes by the name, the timing, the loudness and the voice. If the answer is still unclear it asks the Mac's own on-device judgement (on macOS 26 with Apple Intelligence), and without that it takes a turn that looks like a request and leaves one that does not.

It also knows its own voice. If the microphone picks up Moe's own speech, that is echo and not a new turn; it never answers itself.

Whoever opens a conversation owns it. The moment you say "Hey Moe" or hold the key, Moe takes a reading of that voice and uses it for the rest of the conversation, so a second person talking beside you is not answered, and a guest who opens their own conversation is. Someone else can still join by name: saying "Moe" gets through regardless. This needs no setup, because the voice it compares against is the one that just spoke.

Teach Moe your voice, under Your voice in Settings › Voice, learns what you sound like from a few sentences read aloud. It is optional; with it, "is this the person who started this conversation" becomes a measurement instead of a guess. Press Read one, read the sentence on screen, and after three the row reads Moe knows your voice. Moe keeps learning from the turns you open by name. Nothing is recorded; what is kept is a description of the voice, on this Mac. Forget my voice appears once you have read one: Moe stops using how a voice sounds as a signal, and nothing else changes.

Interrupting#

Brushing the hold key does not cost you the answer. The sound stops the moment the key goes down, which is what you want when you have something to say, and the reply is kept, not thrown away, until what you do next says whether you meant to replace it.

Talk over Moe and it stops. What you say next changes what you asked for rather than replacing it: ask for five things, correct one of them halfway through, and Moe keeps the other four. Withdraw it outright, "never mind", "forget it", "scratch that", and it drops the lot. Amending is the default because it is the one you can undo.

Interrupt Moe by talking, off to start with, lets you talk over it without a key. Start speaking while it is answering and it goes quiet, then stops, if it was you. A voice across the room, or the television, and it simply comes back up a moment later, having interrupted nothing. Each reply can be paused this way once; a false alarm does not get a second go until the next turn.

The same is true while Moe is still thinking: pressing the shortcut again mid-thought puts you back in control. Saying "Hey Moe" or pressing the shortcut never waits; those stop it immediately.

Pausing what Moe is saying#

Tap the hold key while Moe is talking and it stops there. Tap it again and it carries on from the same word. Nothing is re-said and nothing is skipped. The island says paused — tap again to carry on so you can tell a pause from a problem.

That is a tap, a quick press. Holding the key is still "let me say something": the sound stops the same way, and then your words either replace the answer or, if you say nothing, let it carry on. Double-tapping while Moe is talking or working stops everything outright; double-tapping while it is idle is how you talk hands-free.

Pausing mid-sentence#

A breath is not the end of your turn. When you stop after something that sounds complete, Moe starts working on it straight away but keeps listening until the moment you would hear its first word. Go on talking inside that window and nothing is lost: the half-answer is thrown away unheard, what you said before the pause and after it are joined into one request, and that is what Moe answers. Stay quiet and the answer arrives exactly when it would have anyway. This applies to turns you start with the name or the shortcut; a held key ends the turn when you let go.

Ending a conversation#

Say you are done in whatever words you would use, "that's all", "thanks", "I'm good", and Moe closes the conversation. This is Moe's own judgement of what you meant, not a phrase it matches.

With the conversation switch on, if it has been quiet for a few seconds and Moe is not waiting on you, it offers to step out rather than sitting there lit. When it goes quiet chooses how: Ask on the island (a tick and a cross by the notch, the default), Ask out loud, Both, or Just stop listening.

The name#

"Hey Moe" runs entirely on your Mac, on a small model that listens only for those two syllables; the room is not transcribed. It is off by default, and there are honest limits worth knowing:

• It sometimes misses the name, and it sometimes fires on a phrase that sounds like it ("hey Joe", "a moment"). Its training was synthetic speech, and no recording of your voice or your room went into it.
• When it fires on something that was not the name, the checks above usually stop anything from happening, but the light comes on for a moment.
• Hold-to-talk and the shortcut are unaffected by any of this.

Teach Moe its name. Under the Say "Hey Moe" switch in Settings › Voice, Teach records you saying "Hey Moe" twelve times and eight phrases that sound like it, each prompt shown on screen, about two minutes in all. Moe then measures how well it hears you on this Mac and picks the setting that catches your voice without answering the near misses, and tells you what it found in words: how many of your "Hey Moe"s it now catches and how many of the near misses it lets through, against how it was before. The row then reads Moe knows how you say its name, and Teach again repeats it, in a different room, say. The clips stay in Moe's folder on this Mac; nothing is sent anywhere. Two more buttons appear once you have recorded: Send these to the Moe team opens a new message to hello@moebot.app in your mail app with your clips zipped and attached, so the next version of the wake word learns from real voices in real rooms, and nothing goes until you press Send; Forget my recordings removes the clips from this Mac and puts the wake word back to the setting it shipped with.

What Moe says while it works#

During a long job Moe says what it is doing, "Going through your mail", "Searching your files", and repeats the current one every so often so you know it is still there. Anything it does through a connected service says which service, whether the service is on this Mac or reached through Composio. The switch for this is Tell me what it's doing while it works in Settings › Replies.

Work silently. When a long job is more talk than you want, hover the notch while Moe works and press Work silently. For the rest of that job Moe stops saying what it is doing, and when the job is done it says only the answer. The step it is on stays on the island as text, so you can still see it is working. Two things are never silenced: a message waiting on the confirmation sheet for you to press Send, and anything that went wrong. Press it again, it now reads Talk me through it, to hear the updates for the rest of the job. It lasts one job, and the button only appears while a spoken question is being worked on; a typed one is already quiet.

Language#

Moe starts in your Mac's own language, the first language in System Settings › Language & Region that Moe can hear, and in English when none of them is one it knows. Settings › Voice › Language changes it. Talking to Moe is tuned for English: in another language Moe waits for a moment of silence to know you have finished, and "Hey Moe" is an English phrase, so the shortcut is the reliable way in; the banner at the top of the tab says so. The fast engine hears English and about two dozen European languages; other languages use a universal engine that downloads once, about 600 MB.

Moe answers in the language you set it to, and stays there. A stray foreign phrase in a turn, which speech recognition sometimes produces, is treated as the mishearing it usually is rather than a reason to switch.

Addresses and numbers#

Numbers arrive written: say "twenty five thousand four hundred dollars" and you get "$25,400". Email and web addresses are written too. "Email izzy three at hqpulse dot ai" becomes izzy3@hqpulse.ai; "example dot co dot uk" becomes example.co.uk; letters said one at a time run together. It only happens where the shape is unmistakable, "dot" followed by a real domain ending, so "a dot matrix printer" stays an ordinary sentence.

Moe's own voice#

Moe speaks with Kokoro, which ships inside the app: twenty-eight voices, male and female, British and American, all running on your Mac with nothing sent anywhere. Settings › Voice › How Moe sounds picks the set of voices and the voice, and Use and preview plays a sample.

Two more sets are offered there. PocketTTS has twenty-six voices and is the only local one that starts speaking before it has finished thinking. It is not in the app: the Download row says what it costs, 939 MB, once, kept on this Mac, and its button fetches it with a bar, a percentage and a Cancel. Until it is here Moe keeps speaking with Kokoro. Stop a download and what has arrived stays; the next Download carries on from there. Premium is the best voice Moe has, the most natural and the quickest to answer, rendered in the cloud and streamed back as it speaks. It is part of the Away plans; picking it on the free plan opens Settings › Upgrade.

Each set starts on a voice chosen by ear, and you can change it whenever you like: Kokoro begins on Heart, PocketTTS on Peter, and Premium on Clive, one of three Premium voices.

If Moe cannot use the voice you picked, it tells you. When Premium is out of reach, no network, a plan that does not include it, this month's allowance spent, Moe speaks with the voice on your Mac instead and a line under the button says so: which set you chose, which voice you are actually hearing, and why. Your choice is kept, so Moe goes back to it by itself the next time it can.

Premium voice and hearing, on paid plans#

On the Away and Away+ plans Moe can hear and speak through the cloud: sharper ears and a more natural voice than the ones that run on your Mac. Each plan includes a monthly allowance, counted for the whole account rather than per Mac and reset on the 1st:

Moe never holds the keys to those services: it asks for a single-use token each time it opens a connection. When the allowance is used up, or if anything goes wrong on the way, no network, a slow answer, a service down, Moe carries on with the voice and ears on your Mac without asking, and nothing you said is lost. Premium hearing covers English and about twenty other languages; in any other language the Mac's own hearing is used.

Use the sharper voice and hearing that come with my plan, under Settings › Voice, is on by default and does nothing on the free plan: what decides is the plan, not the switch. Turn it off and, paid or not, nothing leaves your Mac. The line under it says which is true right now: Cloud voice: on · 41% of this month used, Cloud voice: not part of your plan, Cloud voice: off, this month's allowance is used up · back on a date, or that Moe's cloud could not be reached just now.

Checking the microphone#

Check this microphone and these speakers, under Can Moe hear you while it talks?, says a sentence out loud, listens to how much of it comes back, and tells you whether your speakers are leaking into your microphone. Stay quiet for the eight seconds it takes. If they are leaking, Moe hears itself, and interrupting by voice gets harder; headphones, or the built-in microphone and speakers together, fix it. It also says when a call app has taken echo cancellation away; that is expected, and it comes back when the call ends.

Giving Moe a photo or a file#

Press + beside the message box in Chat and choose where it comes from:

• Photos… — from your library, through Apple's own picker, so Moe sees only the photos you choose and needs no access to the rest.
• Files… — anything on your Mac, folders included.
• Take a Screenshot — drag across the part of the screen you mean, the same as ⇧⌘4. Escape cancels.
• Paste from Clipboard — whatever you last copied, if it is a file or an image.
• Recent on this Mac — the three newest things in Downloads and on your Desktop, one click each.

Or drag files anywhere onto the conversation. What you have added waits above the box, with an × to take any of it back, until you send, with words or on its own.

The panel under the notch has the same +, without the recent files, and you can drag a file onto the notch itself: it opens as the file arrives and stays open until you let go. Once something is waiting in its box it stays open, so you can say what the file is for.

Moe opens what you send where it is on your Mac, so a large file costs nothing to send. Up to four photos per message are shown to Moe as pictures. Photos, pastes and screenshots that did not already have a file of their own are kept in an attachments folder inside Moe's own folder.

Links to what Moe read#

Moe is spoken, so it never reads a web address out loud, but the links are there to click. Under an answer that involved the web, Sources lists the pages Moe actually opened, taken from the steps it ran rather than from what it says it did. When Moe cites something in the answer itself it writes a real link, live in Chat and in the panel under the notch; the spoken version says the label and skips the address. Ask for "the links" and you get exactly these.

Reading the chat#

In Chat, what you said is on the right, in a tinted bubble, and Moe's answers are on the left under its name. Notes that belong to neither of you, like Stopped., sit in the middle. Under an answer, a small chip says what Moe did to get there, 3 steps, and opens into the list; code Moe writes gets its own scrolling box. While a spoken conversation is open, a bar above the box says whose turn it is, Your turn — just talk, no wake word needed, Thinking…, Speaking…, with End conversation beside it.