Everything that makes you good at your work gets said
out loud and then disappears: decisions, nuance, promises, the small details
that matter about people. Notes catch a fraction of it. Listen turns the
conversations you choose to keep into a memory of people, projects and
decisions, where every answer links back to what was actually said.
No account · No meeting bot · Your audio stays on your devices
Calls on the Mac, rooms and memos on iPhone
Every answer opens the moment it came from
Works with the AI tools you already pay for
Free to record, for ever
This looks like an Intel Mac, and Listen will not run on it.
Listen transcribes on your own machine, and that needs the Neural
Engine Apple introduced with the M1 in late 2020. There is no Intel
version, and there is no version that would work.
Open the Apple menu and choose About This Mac. If it says Chip
rather than Processor, this guess is wrong and you can
download
Listen as normal.
Two minutes, sound on. A demo library: the people, the companies and the voices are invented.
The problem
The best context never makes it into your notes.
The promise somebody made at the end of a call. Why the
team rejected the first plan. The detail a client mentioned three months
ago. The exact words behind a decision everybody now disagrees about.
These are the things that make your next move better, and they are the
first things ordinary notes lose.
You rebuild old decisions from fragments. The
reason was said out loud once and written down nowhere.
You walk into recurring conversations cold. Ten
minutes before the call you cannot remember what was left
unresolved.
You teach your AI the same context every time. The
first prompt is always the project, the people and the history,
typed again.
How it works
Listen. Remember. Use.
Three ordinary things, and the middle one is the part
nothing else does.
ListenA call on your Mac, with nothing joining it, or a conversation in
a room on your iPhone. It asks the moment it starts, and
“No” throws the audio away.
RememberIt works out who spoke, and builds a memory of the people,
projects and decisions you choose to keep. Each detail carries the
sentence it came from.
UseAsk it, here, and get an answer with the receipts attached. See
the whole library at once. Open the week and find out what actually
changed.
A page per person
Every detail says where it came from.
A person's page is not a list of their recordings. It is
what Listen has heard about them, and each line carries the recording, the
speaker, when it was said and when it became true. Press one and you are
looking at the sentence itself.
People
What they are responsible for, what they prefer,
who they work with, and what changed since last time.
Projects
Decisions, the reasons behind them, what is
still open, and when each of those moved.
Promises
What you said you would do, and what somebody
said they would do for you.
Your thinking
The notes you wrote yourself, which are
the one thing no transcript can infer.
The source
Every detail keeps the sentence it came
from, who said it and when, so you can open it and check.
Off until you ask, per person. Nothing is generated
in the background for anybody you have not switched on, and adding a
person never calls a model. There is a daily limit and a global pause.
Corrections survive a rebuild, and so does anything you pin, hide or
exclude.
A detail can be out of date without being wrong, and the page
says which. What somebody said themselves is marked differently from what
was said about them.
The galaxy
Your whole library, in one picture.
Four shells around this Mac: your recordings, your notes,
the people in them and the conversations you have had with Listen about
any of it. A line is drawn wherever the library already records a
relationship between two things.
Search in the sidebar and the picture narrows to what matches
and what that connects to. Click a star to see what it is joined to;
double-click to go there.
Nothing here is inferred. A note that names a meeting
is a line. Somebody speaking in one is a line. A question you asked about
one is a line. There is no similarity edge and no "these happened the same
week" edge, so every line is a row you can go and read.
Distance from the centre is the kind of a thing and nothing else, not how
important it is and not how recent, because there is no measure of
importance here that is not a guess, and a guess drawn as a distance reads
as a fact.
Drawn from what is already on the disk and reaches
no network. It stops on its own when the window is hidden, under Reduce
Motion and in Low Power Mode, and Settings turns it off entirely.
This week
What changed, and what to do about it.
A deck of cards where the sidebar sits, with the galaxy
beside it: the week's new conversations, people and notes arriving into
the library you already had. Every card carries something to do. Name the
speakers nobody labelled. Accept or correct a detail. Open somebody you
have not spoken to in months.
Two of the cards look forward rather than back: what was
promised, in the words somebody actually used, and what you already know
about whoever you are seeing next.
The window is a named period rather than a rolling seven days, because
"last 7 days" ending on a Tuesday afternoon is nobody's week and two
readings an hour apart were two different reviews. This week, last week,
this month, this year, Monday to Sunday. A finished period ends at its own
last second, so it can be compared with the one before it.
It is derived, not stored. Nothing is generated on a
schedule and nothing advances itself. A person's page has the same deck
about them over the last ninety days.
What you do with it
Three questions you cannot answer today.
Ten minutes before the call
“Catch me up before I speak to Maria again.”
What you last agreed, what she is waiting on, and the thing
she mentioned in passing that you meant to follow up. Each line opens
the conversation it came from.
When an old decision comes back
“Why did we decide against annual pricing?”
The reasons, in the words they were argued in, with the
date and the meeting. Not a summary somebody wrote afterwards.
When your AI needs to know something
“Draft the brief using what customers actually
said.”
Your agent reads the relevant conversations over MCP and
writes from them, instead of asking you to paste a folder of transcripts
into the prompt.
Ask
Answers with the receipts attached.
Ask one conversation or all of them, in the app, out loud
if you would rather. Every claim comes back with a numbered
reference, and pressing it opens the recording at the moment it
came from. If a model answers without reading anything, Listen says so.
Retrieval is local: keywords and Apple's on-device semantic search, and it
can tell you what it knew at a date in the past.
You choose the model: your own assistant subscription, your own API
key, or a model running on your own Mac, in which case the question never
leaves the machine.
And it reads from the tools you already use
The same retrieval is behind a command line and an MCP server, so
Claude Code, Codex or anything else that speaks MCP can read your library
directly rather than a folder of transcripts you paste into a prompt.
claude mcp add listen -- /usr/local/bin/listen mcp
An agent can write notes and tags. That is all. It
cannot rename a speaker, correct a transcript, retitle a recording or
delete one, and where it disagrees with what Listen believes about
somebody it proposes a correction rather than applying one. A wrong note
sits harmlessly beside the recording that disproves it. A wrong transcript
edit is a fact that is simply gone, so those edits stay with you.
Underneath
The record it is all built on.
A memory is only worth as much as what it was built from.
Recording, transcription, the voices and the search over them are free,
for ever, with no meeting count and no trial clock.
Record
Both sides of the call, and no bot in it.
Your microphone on one track, everything your Mac plays on the
other. Nothing joins the call, so it works on
WhatsApp and FaceTime as well as it does on Zoom, Meet and Teams.
You do not have to remember to start it: Listen spots the call,
begins recording, and only then asks whether you want it.
It is already recording when it asks, so the answer costs you nothing
Say never once and it stops asking about that app
Asks to record audio, not your screen
Transcribe
Written up before you've made coffee.
Transcribing happens on your Mac's own chip, about 240 times
faster than real time on an M4 Max. Play a meeting back and the
transcript follows along, sentence by sentence, so you can find the
bit you half remember without listening to the rest.
Fix one sentence without touching the paragraph around it
Reads 25 languages, and works out which one it's hearing
Teach it the names and jargon it keeps getting wrong
Speakers
Name someone once. Listen knows them next time.
Listen works out who spoke and puts names on the turns. Tell it
who someone is once, and it recognises them in the next meeting they
turn up in. On a call it never has to guess at your own voice.
Works for a room too, with the laptop on the table
Fix a whole speaker, one paragraph, or one sentence
Voiceprints are biometric data, and yours never leave your Mac
Notes
Your notes stay yours.
Write your own notes while the meeting is still running, which is
when they're worth the most: “we should upsell them” is the kind of
thing no transcript will ever hold. Ask an agent for a summary and it
writes a separate one. It can read yours. It cannot change
them.
No New Note button. Open a recording, click Notes, start typing
One note can cover four meetings, and outlive any of them
Tag a meeting in your own words, and ask for that tag later
Dictate
Talk anywhere on the Mac, and it types.
Press fn + left shift wherever you're typing,
say what you want written, and press it again. The words land where
the cursor is, in an email, in Slack, in a search bar, and on the
clipboard as well.
The same model, and the same vocabulary, your meetings use
On macOS 26, an optional pass that drops the ums
Needs Accessibility. Recording meetings never does, so it never asks
Your devices
Start on one Mac, pick up on the other.
Turn on sync and your transcripts, notes, people and tags follow
you between Macs, through your own private iCloud
rather than anybody's server. Listen seals all of it on the device
first, with a key Apple never holds. There's nothing to pair and no
code to scan.
The audio travels too. An hour of meeting is about 61 MB
No device deletes its copy until another one has it
Anything a sync removes is kept for fourteen days
Works with
Listen records the audio your Mac is already playing, so it needs
nothing from the app on the other end. That includes the ones no notetaker
bot can reach.
The iPhone app is in TestFlight. It records the
conversations that happen in a room rather than on a call, transcribes
and separates the voices on the phone itself, and writes into the same
memory: a name your Mac already knows arrives with the turns. The Mac
keeps the heavier transcription model, the calendar and the MCP server.
Join
the beta →
Trust
A memory this personal should be yours.
There is no Listen server, so there is nothing to breach
and nothing to hand over. Your meetings are saved on your own computer, and
the recording, the transcript and the names on it are all made there. No
account, and none of it ever reports back: the only statistics are
anonymous counts, turned off in Settings whenever you like, and
public
in their entirety.
Nothing is kept quietly. Listen asks on screen the
moment it starts, and “No” deletes the audio. Say never once
and it stops asking about that app at all. Between calls it is watching
for a call to start, not listening to the room.
The memory is yours to correct. Every detail it
learns carries the sentence it came from, so a wrong one is visible
rather than buried, and it can be edited, retired or deleted. Building
it is off until you turn it on.
It works with the Wi-Fi off. None of it waits on
a connection, so losing the network does not cost you the meeting.
Your recordings are ordinary files. A folder you
own, that you can move, back up or delete. There is nothing to export,
because nothing was ever taken.
Keep everything, for as long as you like. Your
library is a folder on a disk you already own, so no history expires and
nothing is held back to be unlocked later.
Three things can send anything out, and each is off until you turn it
on: keeping your library in step across your own devices, putting a
question to a hosted model, and building the memory, which reads passages
to whichever model you point it at. Listen seals anything that travels
before it leaves the device, with a key we never hold, so there is nothing
readable at the far end. Point all three at a model running on your own
Mac and none of it leaves the machine at all.
None of which has to be taken on trust. Listen ships
a machine-readable list of every connection it is allowed to make, inside
the app, which firewall tools like Little Snitch read straight out of it.
Point one at Listen and watch. What runs
where sets out each connection and how to check it yourself, and there
are answers ready for a privacy or
HIPAA review.
Somewhere with rules about this? Listen can be
locked down centrally through the same device profiles an IT team
already uses: sync off, questions answered only by a model on the
machine, no dictation history, backups pointed wherever policy says.
Deploying under HIPAA or GDPR has the detail.
Where it sits
What kind of memory do you want?
Categories rather than named products, because naming one
dates this table the week they ship something. Check any row against the
tool you actually use.
A meeting summariser
A notes app
Memory inside one AI
Listen
Starts from real conversations
Yes
Only what you type
Only what you tell it
Yes
Carries people and projects forward
Rarely
By hand
Inside that one vendor
Yes
Links a claim to what was said
Sometimes
Sometimes
Rarely
Yes, every one
You can correct what it learned
Limited
Yes
Limited
Yes
Readable by the AI you choose
Limited
Varies
Usually not
In the app, and over MCP
Runs without a vendor server
No
Varies
No
Yes
Every row is a claim about a category, not about a
named product, and is worth re-checking against whatever you are comparing
it with.
Who it is for
Conversations that cannot go to a server.
Some work is sold by the hour of attention, and some rules
out a meeting bot before anybody has read the pricing. Either way this was
built for you rather than adapted to you.
Download it, record one conversation, name the people in
it, then ask what was said. The memory grows from there. Recording is free
for ever and the source is public; what it
costs has the rest. Apple silicon, macOS 14 or later, and the speech
model is about 2.5 GB the first time you run it.
