A free meeting copilot in a 0.8 MB download.

Live transcripts, answers mid-call and notes afterwards, on the Apple Intelligence already in your Mac.

Price$19.99/moFreeStealth modeup to $149.99/moFreeDownloadgigabytes, or their cloud0.8 MB

curl -sL https://github.com/JiyauddinSaiyad/sotto/releases/latest/download/Sotto.zip -o /tmp/Sotto.zip && ditto -xk /tmp/Sotto.zip /Applications && open /Applications/Sotto.app

Paste in Terminal, no security warning · Download · GitHub · Apple silicon, macOS 26+

Help during the call

A small glass panel that floats over every app, hears both sides, and answers when you ask.

00:23Listening

Listens to both sides. Your mic is “You”, the call’s audio is “Them”, transcribed live on your Mac with nothing sent anywhere.

Assist · heard the last 4 minutes

Answers in a second or two. Press ⌘↵ for Assist, tap “What should I say?”, or ask anything about the conversation.

Why we built it

Your Mac already has a speech recognizer and a ~3B-parameter language model. Apple put them there. Almost nobody uses them.

Meanwhile, AI meeting copilots charge $19.99 a month, upsell “stealth” at up to $149.99, and stream your calls to their servers to run models you’re renting. That felt backwards. So we asked one question: how far does the AI that’s already on your Mac go?

That’s why it’s free. The AI is already built into your Mac. We pay for no servers, so you pay us nothing.

Further than we expected. 2.06% word error rate. First word of an answer in 0.4 s. Notes in about five seconds. All of it in a 0.8 MB app, because the heavy parts were never ours to ship.

It isn’t a frontier model, and we won’t pretend it is. It’s great for summaries, follow-ups and “what was that number again?”, not for deep reasoning. Everyone else on the call shows up as “Them”, because Apple’s speech APIs don’t separate speakers yet. Names get misheard. Only English has been tested. We’d rather tell you now than have you find out mid-call.

Windows: coming soon Copilot+ PCs ship their own on-device model (Phi Silica), so the same idea works there. Tell us you want it →

Benchmarks

Why does a 0.8 MB app work? Because the models already shipped with macOS. Here’s how they run on the cheapest Mac that has them.

M4 MacBook Air16 GB · fanlessmacOS 27.0.1medians of 5 runsreproduce: bench/ →

Language model

Apple Foundation Models, on-device ~3B · tokens counted with Apple’s own tokenizer

42tok/s sustained decode
1,127tok/s prefill
0.4s to first token
WorkloadInOutTTFTPrefillDecode
Short answer11330.42 s–55 tok/s
Long answer174340.37 s–42 tok/s
Recap, 1K-token transcript8181360.79 s1,032 tok/s35 tok/s
Recap, 3K-token transcript2,5781332.29 s1,127 tok/s26 tok/s

Context window 8,192 tokens. First request after launch: 3.75 s cold start. On a fanless Air a hot machine runs up to ~40% slower; we saw it.

Speech to text

Apple SpeechTranscriber · LibriSpeech test-clean, the set every ASR paper reports

2.06%word error rate
49×faster than real time
0MB of models shipped
Word error rate, LibriSpeech test-clean · lower is better
Sotto0 MB2.06%
Whisper Small*466 MB3.74%
Whisper Base*142 MB5.42%
Whisper Tiny*75 MB7.88%

Ours: all 2,620 utterances, 52,576 words, 5.4 hours of audio in 6.6 minutes. *Whisper rows are a published third-party run on an M2 Pro, not ours; it measured Apple’s recognizer at 2.12%. Whisper Large-v3 is reported around 2–2.7%, so roughly a tie, but it’s a ~3 GB model.

0.8 MBdownload
1.6 MBinstalled
~85 MBRAM mid-meeting
0bytes of audio leave your Mac
0dependencies

The only one that needs nothing

Open-source meeting tools make you download a model or paste an API key. Paid ones make you pay monthly. Sotto does neither.

Open-source apps

Open-source appMac downloadWhat you set up before it worksWorks out of the box
Sotto0.8 MBNothing. It uses the AI built into macOS.Yes
Meetily45 MBA Whisper or Parakeet model, plus Ollama or an API key for summariesNo
Anarlog201 MBA local model server, your own API key, or their hosted planNo
cheating-daddy92 MBA Gemini API keyNo
Pluely13 MBYour API key, an AI CLI, or a paid planNo
free-cluelyBuild from sourceA Gemini API key or OllamaNo
OpenCluelySetup scriptA Gemini API key, plus a local Whisper installNo

The six most-starred open-source alternatives on GitHub, checked on Oct 26, 2026 against each project’s README and latest release.

Paid apps

Prices from each company’s own pricing page on Oct 26, 2026, monthly billing; “–” means not offered on that page. †Parakeet AI prices by region; this is what its page showed us. Spotted something out of date? Tell us.

Notes, before you’ve closed the call

Stop the meeting and Sotto writes a summary, action items and decisions, saved as Markdown in ~/Documents/Sotto.

2026-10-26 09.41.md

Action items

    2026-10-26 09.41.md

    Weekly launch sync

    2026-10-26 09.41.md

    Decisions

      Quiet by design

      sotto voce: in a quiet voice. An assistant that whispers to you, not the room.

      Meeting participants (4) No bots in the call
      MKMaya K.maya@example.comHost
      DODev O.dev@example.comSpeaker
      SLSam L.sam@example.comSpeaker
      YouYouyou@example.comSpeaker

      Nothing joins your call. Sotto hears your Mac’s audio, so no note-taker bot shows up in the participant list.

      What they see when you share

      Legal signs off Thursday, so the earliest launch is Tuesday.

      What you see

      Hide from screen share. One click on and a dotted border shows it’s on. Meeting apps don’t all honour it the same way, so check your share preview first.

      Hide 04:12

      Drag the panel

      +hides it

      Out of your way. Floats over every app and full-screen Space, never steals focus. ⌘\ hides it; drag it anywhere.

      0bytes

      sent to a server. There isn’t one.

      Private by default. Audio, transcripts and notes stay on your Mac. No account, no cloud, no API key.

      Zoom Google Meet Microsoft Teams Slack Webex

      Works with anything that makes sound. No integrations, no plugins, no calendar access. If it plays through your Mac, Sotto can hear it.

      0.8MB

      The whole download

      Local Whisper models alone weigh 75 MB to 3 GB. Sotto ships none; it uses the speech recognizer and language model macOS already has.

      $0

      Forever

      No subscription, no usage limits, MIT licensed. Your Mac already paid for the AI.

      0.4s

      To the first word of an answer

      Measured on an M4 MacBook Air. A full recap of the call lands in about five seconds.

      Real-time transcript04:12

      Questions

      Is it really free?

      Yes. Sotto is open source under the MIT license. There’s no account, no subscription and no server, because the AI is the one Apple already put on your Mac.

      Which Macs does it run on?

      Any Mac with Apple silicon (M1 or later) on macOS 26 or later, with Apple Intelligence turned on in System Settings → Apple Intelligence & Siri. Sotto checks at launch and tells you plainly if something’s missing.

      What permissions does it need?

      Two: the microphone, to transcribe you, and System Audio Recording Only, to transcribe everyone else. It hears audio only and never sees your screen. If the other side isn’t showing up, turn Sotto on in System Settings → Privacy & Security → Screen & System Audio Recording.

      Can people on the call tell?

      No bot joins, so nothing appears in the participant list. Hiding the panel from screen share is opt-in, and meeting apps don’t all honour it the same way, so check your share preview before relying on it. And be a good human: in many places you need consent to record or transcribe a call.

      How good is it, honestly?

      It runs on Apple’s ~3B-parameter on-device model, which is what makes it free and private, and also sets the ceiling. It’s genuinely useful for summaries, follow-ups and “what was that number again?”, not for deep reasoning. Everyone except you is “Them”, because Apple’s speech APIs don’t separate speakers. Names get misheard; add them under Settings → Names & terms and the notes spell them right.

      Which languages?

      Your Mac’s language, if Apple’s speech recognizer supports it. Only English has been tested so far, so reports from other languages are very welcome.

      Your Mac already has the AI. Put it to work.

      Free, 0.8 MB, and nothing else to download. Ever.

      Install in one line. Paste this in Terminal. Files fetched with curl aren’t quarantined, so macOS opens Sotto straight away.

      curl -sL https://github.com/JiyauddinSaiyad/sotto/releases/latest/download/Sotto.zip -o /tmp/Sotto.zip && ditto -xk /tmp/Sotto.zip /Applications && open /Applications/Sotto.app

      Or download it the regular way

      1. Download Sotto.zip and open it.
      2. Drag Sotto into Applications.
      3. Open it. macOS will say it can’t verify the developer, because Sotto isn’t notarized yet. Go to System Settings → Privacy & Security and click Open Anyway. Once.