Hold fn. Speak.
Read it before
it types.
Fast enough for a Slack reply, careful enough for the email you can’t get wrong. Your voice never leaves this Mac, and nothing gets typed until you’ve read it.
Dictation is free forever.
Works offline too.
Live transcription and Tidy cleanup, always free. AI cleanups are optional — 100 free, then bring your own key, or switch them off.
macOS 26 · Apple silicon
Audio never leaves this Mac
Two ways
Some things you send.
Some things you read first.
Both start the same way — hold fn and talk. What changes is how you let go, and whether anything reaches the app before you’ve read it.
For the reply you’d have typed in five seconds
Priya Raman2:34 PM
Numbers for the review are in the doc now.
Dan Whitfield2:37 PM
Nice. I still owe you the churn breakdown — tomorrow morning.
Arun2:41 PM
Can we move the review to Thursday?
Hold fn. Say it. Let go — it’s in Slack before your hand is back on the keyboard.
For the draft you want to think out loud and then structure it
Q3 planning
We shipped the onboarding rewrite last week. Open question is what goes out next, and in what order.
Hold fn, tap space, then let go and keep talking. It waits in the panel until you’ve read it and pressed Return.
Four versions
Press them. This is what they actually do.
Every version is cached after the first run, so flipping back and forth is free — that’s the point of having buttons at all. Try them: these behave the way the real ones do.
Original always stays. Tidy, Polish and Structure each have a switch in Settings, and Tidy needs Apple Intelligence turned on.
That’s the whole gesture. Want it on your Mac?
Works where you work
If you can type there,
you can talk there.
Speaklively’s panel opens without stealing focus, so the app you were in stays the app you’re in. When you press Return the text arrives at the caret — no click back, no window to dismiss, nothing to put away.
~/speaklively % swift build
[1/2] Compiling SpeaklivelyCore
[2/2] Compiling Speaklively
Build complete! (12.34s)
~/speaklively %
Text goes in the same way ⌘V does, so there’s nothing to set up per app. Password fields are the one exception — it copies instead, and tells you.
What stays and what goes
Your voice never leaves this Mac.
Your words leave only when you press.
Private enough for the document you can’t afford to leak, and it still works on a train with no signal. Hearing you and tidying you up both run on Apple’s on-device models, so there’s nothing to send in the first place.
Your voice
Never written to disk — not a temp file, not a cache.
The transcription
Apple’s on-device models, running here.
Tidy’s cleanup
Also here, using the model built into this Mac.
Your last 100 dictations
Plain text, on this disk. Switch history off and nothing is written at all.
All of it works with the Wi-Fi off.
The text of one dictation
Only the transcript you pressed Polish or Structure for. Nothing else, and never the recording.
It travels through a relay I run, which counts what the call cost so a beta token can have a ceiling, and keeps nothing else — no transcript, no prompt, no reply. Paste your own Anthropic key in Settings and it skips the relay entirely.
One exception, and it ships off. Vocabulary suggestions are the only feature that sends more than the single transcript you pressed for — they send a slice of your history, so Claude can learn the words you keep having to correct. The switch is in Settings, off, and you can leave it there.
Once a day, it asks whether there’s a newer version. That’s the one thing that goes out without you pressing anything. It carries which build you’re running, a random number made when you installed — so I can tell forty people running it from one person opening it forty times — and how many words you’ve dictated since it last asked. A count, never the words themselves: the number 412, and nothing about what any of them were. It’s the only way to know whether anyone finds this useful, because dictating with Polish off never touches a server at all. Nothing about your Mac, and nothing you dictated. The switch is in Settings under Privacy, and turning it off stops all of it — including the app saying which install it is when it checks for updates. Updates themselves carry on; they are a separate switch now.
No account. No upload. Want a build?
Before you ask for one
Ask the awkward ones first.
What it costs once the free dictations run out, which Macs it won’t run on, what leaves your Mac and exactly when. These are the answers I’d give you in a reply, and the ones that aren’t settled yet say they aren’t.
What happens after I put my details in?
The download starts straight away. The AI features are switched on from inside the app, in Settings: press “Set up access”, type your address, and a six-digit code arrives by email — you never handle a key at all, and your first 100 dictations are free. Your name and address are kept on their own, with no IP address or browsing history attached, and are never sent to website analytics. I use them once, to ask how you’re getting on. That question is the whole point of a private beta.
What will it cost me later?
The parts that run on your Mac — dictation, live transcription, Tidy, your history — are free forever. Not free for now: free, permanently. They cost me nothing each time you use them, so there is nothing there I could honestly charge for, and a price that appears later on something that was always costless is exactly the move I would be annoyed by. The AI cleanups are the one part that does cost me money, every single press. Those are free for your first 100 dictations. After that it is $39 a year — though checkout is not open yet, so for now you can point them at your own Anthropic key so nothing of yours comes near me, or switch them off. Everything in the first list carries on exactly as it did.
Which Macs won’t run it?
Anything before macOS 26, and every Intel Mac. Tidy needs Apple Intelligence switched on as well, which some otherwise-eligible Macs have turned off. That’s a narrower requirement than anything else in this category, and it’s also the reason it can transcribe live without the network.
Which languages does it handle?
macOS 26 ships two on-device transcribers and Speaklively uses whichever one covers your language — 54 between them, Hindi included. The Whisper-based tools do 100+, so this is genuinely fewer. If your language isn’t in Apple’s set, this is the wrong app for you today.
What exactly leaves my Mac, and when?
Your voice never does. Transcription and Tidy both run on-device. When you press Polish or Structure, that one transcript goes out — through a relay I run, which counts what the call cost and keeps no transcript, prompt or reply — and on to Claude. Nothing goes anywhere until you press for it.
Can I use my own Anthropic key?
Yes, in Settings. Your key takes precedence over the one Speaklively hands you whenever both are present, which means no cap, no metering, and nothing of yours passing through my relay at all.
What if I don’t want it using Claude at all?
Turn Polish and Structure off in Settings and it never reaches for the network. Dictation and Tidy carry on working with the Wi-Fi off, and Original — what you actually said — is kept whatever else is switched on.
Can I get a dismissed dictation back?
Your latest 100 are kept on this Mac as plain text, dismissed ones included. Turn history off in Settings and nothing is written to disk at all. The audio never is, either way.
WHY SPEAKLIVELY
Everyone else types it and hopes. This one shows you first.
Fast enough to think out loud, careful enough for the thing you’d hate to send twice. It runs on your Mac — so it works offline, and your voice never leaves the machine.
Read it before it types · Nothing leaves this Mac · Works offline
Dictation free forever — macOS 26 on Apple silicon.