Last updated 18 July 2026 · This page states the whole policy in plain
English — there is no longer, scarier version behind it.
The core promise
Fabled Flow processes your voice entirely on your own computer. Your audio is never
uploaded, never stored on our servers, and never used to train anything. Dictation works
with your network switched off.
The iron rules
Three promises, engineered in — not policies, but rules the code itself cannot break,
each held in place by an automated test suite that every release must pass:
1
No audio stored
Recordings are transcribed and released. There is no folder of your voice
anywhere. The short local troubleshooting window described below is the one
disclosed exception — and it never leaves your machine.
2
No word lost
Formatting may only rearrange — never delete, replace, or invent words. AI
formatting output is verified word-by-word against what you said, and rejected if
even one word went missing. Every refinement started life as a real dictation that
went wrong on a real day, kept verbatim as a test the product must pass forever.
3
Common words are untouchable
Your dictionary can teach new words, but ordinary, correctly spelled English can
never be silently rewritten into something else — that boundary is enforced
structurally on every code path.
What stays on your device
Audio: recordings are transcribed on-device and released. A small rolling window
of your most recent recordings (up to thirty) is kept locally for troubleshooting only,
overwritten as you dictate. We never receive these and have no way to access them —
nothing is ever transmitted automatically. The only audio that can ever leave your Mac is a
recording you make inside the Feedback box and choose to send with a report.
Transcripts: your dictation History is a local file on your device. You can delete
entries, or everything, at any time.
Dictionary and snippets: local files on your device.
Correction notices (optional, on by default): right after Fabled Flow pastes
your text, it re-reads that one text box once — about twelve seconds later, and never
again — to notice whether you fixed a word by hand. The comparison is
dictation-by-dictation: only the words you just dictated are compared against your
fix of them. Whatever else is in the box is used for one thing only — finding where
your dictation landed — and is then immediately forgotten: never stored, never
analyzed, never part of any note. Only the text box Fabled Flow itself pasted into is ever
read, never your screen, and starting a new dictation cancels the check entirely. Every
step happens on your Mac.
What gets noted, and where it lives: the before-and-after of your fix (say,
"are" → "our") is saved in a small file on your Mac, with a count of how often the
same fix repeats. Your Mac also sorts each note into one of three kinds: a new word worth
learning (a name, a brand), a sound-alike mix-up ("are"/"our"), or an ordinary word swap.
Separately, your personal Dictionary entries keep two counters — how often each entry
genuinely fixed something, and how often you had to fix something an entry should have
caught. Every bit of this analysis happens on your Mac. Nothing is ever added to your
Dictionary automatically. These notes exist so future accuracy improvements are
designed from real evidence — and so features like personal sound-alike fixes can one
day be offered to you, from your own patterns, rather than imposed on everyone.
What leaves your Mac: numbers only — how many notes exist and of which kind,
plus the two counters, riding the same anonymous usage statistics described below. The
words themselves, your text, and your audio never leave your device. Turn all of it off
any time in Settings → Dictation ("Notice my corrections").
This is not screen recording: Fabled Flow cannot see your screen — no
pixels, no image of anything, no view of other apps or other windows. It takes no
screenshots and never asks for the macOS Screen Recording permission. It reads the text of
one box, through the same accessibility channel it uses to paste your words there and
verify they landed: the one box we pasted into, once, twelve seconds after our own paste,
then immediately discarded. For comparison: during our July 2026 head-to-head benchmark
testing, we found Wispr Flow's app storing a screenshot of the screen with every
dictation in its local database on our test Mac. That is screen recording, per
dictation, silently. We read one box's text once and forget everything but your own
fix.
The app's network traffic — the complete list
One-time model downloads — the speech models fetch once from public model hosts.
Update check — once a day the app asks for the latest version number. Nothing about
you or your machine is sent.
Anonymous device counters — once a day the app reports: a random device identifier,
device model, app version, macOS version, interface language, and words-dictated counts.
No names, no emails (unless you sign in — below), no audio, no text content, no location.
Error reports you choose to send — the in-app Feedback box sends only what it shows
you: your description, recordings you make inside the feedback box while recreating a
problem, and a technical log. Nothing is attached silently.
If you sign in (optional)
Signing in links your devices to your email address for device management and future
sync. The email address is the only personal data stored, and sign-in is never required
for dictation. Sign out removes the link. Authentication and the device registry run on
Supabase; login codes are delivered by email.
What we will never do
Upload, store, or listen to your audio.
Send your dictated text anywhere.
Sell or share any data with third parties.
Add tracking, ads, or analytics SDKs to the app.
This website
This site is a static page. It sets no tracking cookies and runs no analytics. Download
links are served by GitHub, and fonts by Google Fonts — both see standard web-request
metadata (your IP address) as any web host does.
Fabled Flow