Features
Everything Flit does.
You hold a key and talk. Finished text lands where your cursor already is, about 0.7 seconds later, without a single byte leaving the machine. Everything below is a variation on that one sentence.
Mail 9:41 Can we meet at 10:30 AM on Friday to go through the Q4 numbers?
0.7s after keyup
so can we meet at 1030 am on on friday to go through the q4 numbers
0.7s
from keyup to formatted text
0.15s where formatting is off
99
languages, every one local
25 of them stream live
0
bytes leave your Mac
no account, no telemetry, no server
How it writes
Finished text, not a transcript.
Punctuation, capitals and paragraph breaks are applied before the words reach your cursor, and the filler is dropped. The pass runs on your Mac: Apple Intelligence on macOS 26, or a local model through MLX on everything older.
- Tone follows the app, or you pin it.
- Work vocabulary written as work writes it: B2B, API, Q1, 24/7, LinkedIn.
- Every word you said survives: output that loses one is discarded.
- Terminals and editors are never reworded.
- If the model fails, the plain transcript lands anyway.
One sentence, two apps
Slack
shipping this today, will ping you when it’s live 🎉
Mail
This is shipping today. I will let you know as soon as it is live.
You said the same thing both times. The style rule for each app did the rest.
You say
You get
“open paren”
(
“dollar sign”
$
“new paragraph”
¶
“all caps on”
ABC
“dash dash force”
--force
“scratch that”
⌫
Punctuation and commands
60 commands, and not one of them is a guess.
Fixed rules, not a model. The same phrase produces the same character every time, which is the only way this is usable in a terminal. Paired symbols close themselves, and “literally comma” types the word instead.
There is a searchable reference inside the app, because 60 is more than anyone memorises on the first day.
Developer mode
It reads your project, not your code.
Dictating into an editor or terminal, Flit indexes the open workspace and repairs spoken identifiers against what is really there. Casing comes from your own repository, so camelCase stays camelCase.
- File and directory names only. Contents are never read.
- The index lives in memory and never touches disk.
- Dictating into Claude or ChatGPT shapes it as a prompt.
Cursor You said
open voice snicket store dot swift
Flit wrote
open VoiceSnippetStore.swift
The engine misheard it. The workspace index knew that file existed, so the repair is deterministic rather than a guess.
Your words
It learns your vocabulary, and asks first.
Fix a word by hand and Flit notices about 6 seconds later, then offers to remember it. Nothing enters your dictionary without you saying yes. Terms carry usage counts, so the ones you actually say are the ones that win.
Snippets work the other way round: say a short trigger, get the whole block. Matching is fuzzy on the raw transcript, so it fires even when the trigger itself is misheard.
Voice Snippet
VoiceSnippetStore
Snippet trigger
“book a call”
Expands to
Happy to find a time. Here is my calendar:
cal.com/your-link
Analytics
Quiet numbers, beautifully kept.
The History tab keeps a running count of the words you have dictated and the writing time you got back, computed on your Mac and shown in 3 calm tiles. It informs and never nags: no streaks, no notifications, no guilt mechanics.
The counters survive history pruning, so clearing transcripts never resets what you have saved.
History · your numbers
128k
words dictated
17.2h
of typing saved
0
streaks, nags or reminders
Computed locally from your own history. Leaves the machine never.
Privacy
There is no server to trust.
Flit makes no network request to dictate. Not for the audio, not for the text, not for a license check, not for analytics. There is no account because there is nothing to hold one against, and screen context is held in memory for exactly one dictation.
Check it yourself
Turn off the wifi and dictate. It works exactly the same, because there was never a request to make. Point a network monitor at it and watch nothing happen. That takes a minute, and it is worth more than any certification we could buy.
Read the privacy pageThe full list
And roughly 35 smaller things.
None of these sells the app on its own. Together they are the difference between something you try and something you keep.
Activation
- Hold to talk, or tap once to lock it on
- Double-tap for hands-free, or arm voice activation
- Any key with a modifier, Fn by default
- Escape cancels anything, at any point
- Silence ends a hands-free utterance after a pause you set
- Review before insert: Return sends, Escape discards
The HUD
- A floating panel, a glass pill in 6 corners, or the notch
- Live transcript while you speak, roughly every 320ms
- Dark glass, or follow the system appearance
- Sound cues on start and stop, if you want them
- Test a phrase from Settings without leaving it
Text handling
- Tone: auto, casual, neutral or professional
- Paired symbols close themselves
- Spoken emoji, in the apps where you meant it
- “Literally comma” types the word, not the symbol
- Times and dates resolved by fixed rule
- Filler words dropped without touching your phrasing
- Edit a selection by voice instead of appending
Your data
- Transcript history is capped, prunable, and yours to disable
- Audio deleted right after transcription unless you keep it
- Copy-only mode needs no Accessibility permission
- Your clipboard is restored after Flit pastes
- Dictionary and snippets are plain files you can copy between Macs
- Words dictated and time saved, computed locally, no streaks
Models
- Parakeet on the Neural Engine as the default
- Apple’s own speech recognition, nothing to download
- Whisper in several sizes, for accuracy over speed
- A 1 GB formatting model via MLX for Macs without Apple Intelligence
- An optional larger formatting model for Macs with 32 GB of memory or more
- Every model shows its size, and can be deleted and refetched
- Engines switch automatically when one cannot handle a language
Fit and finish
- Launch at login
- Pick your input device
- Onboarding that explains each permission before macOS asks
- If formatting fails, the plain transcript is inserted anyway
What it does not do
Where the focus ends.
Flit is built to do one thing completely rather than four things partly. Here is the edge of that decision, stated up front.
Transcribing files
Flit is live dictation: you hold a key and speak. It does not take audio files, recordings or meetings, and that focus is why it is fast.
Windows, iPhone or iPad
macOS only, on Apple silicon. Every decision in the app is a macOS decision, which is why it fits this well and why it will not follow you.
Syncing between Macs
Your dictionary and snippets are plain files you can copy across. There is no sync, because sync needs an account and a server.
Team licensing
Personal is one person on up to 2 Macs. Teams are licensed per Mac, from 5 Macs, with volume pricing on the pricing page and one key for everyone. More than 50 Macs? Write and we will talk.
Your voice was always faster.
Not out yet. $29 USD once when it is, with no account, no subscription and nothing uploaded.
No spam, no sharing. One email when Flit ships, and you can leave in one click.
macOS 14 or later · Apple silicon