Oats 0.2.0

Oats hears the whole room now, names what the conversation was about, and links the conversations that come back to the same subject. Setup sets up the models for you, and summaries start by themselves.

Download

• Apple Silicon (M1 and later): Oats-arm64.dmg, about 910 MB.
• Intel: Oats-x64.dmg, about 610 MB. It is built for Intel, but nobody has tested it on an Intel Mac yet.

Oats needs macOS 12 Monterey or later.

The first time you open it

This version is not notarized by Apple yet, so macOS stops it once.

1. Open the DMG and drag Oats to Applications.
2. Open Oats. macOS says that it cannot verify the app. Click Done.
3. Open System Settings, then Privacy & Security. Scroll down and click Open Anyway next to "Oats was blocked".
4. Confirm with your password or Touch ID. macOS asks only once.

What is new

Setup asks where Oats thinks. A new step offers two choices:

• On this Mac: private and free. When you continue, Oats downloads the models that suit your Mac: Whisper Turbo (574 MB) for speech on a Mac that is fast enough, and a writing model for titles and summaries. The writing model is Gemma 4 E4B (5.4 GB) on a Mac with 16 GB of memory or more, and Qwen3.5 4B (2.7 GB) on a Mac with less. Gemma 4 E4B made up the fewest facts in our summary tests. Each download shows its progress, with Cancel.
• With an API key: paste an OpenAI or Anthropic key. Oats asks the provider whether the key works before it saves it.

A short tour. After setup, five slides show what Oats does with a conversation. Each slide is a part of the app itself, shown with the example conversation. Skip it, or go through it with Next or the arrow keys.

One way to choose models. Settings › AI models has one choice: On this Mac, or With an API key. Under it, one row for speech and one for writing say which model Oats uses, why, and whether it is ready. To choose models yourself, open Advanced: speech and writing use the same list there.

Upgrading. On this Mac with no writing model yet, the first launch of 0.2.0 starts its download, with a notice and Cancel. A model you chose yourself is kept.

Summaries start by themselves.

• After a download: a conversation that finished while the writing model was still downloading is summarized when the model is ready.
• After a key: with an API key chosen but none saved yet, conversations wait, and they are summarized when you add one.
• When you open one: a conversation that has no summary starts one when you open it. There is no Summarize button.
• After a failure: a summary that fails quickly is tried once more. If it fails again, the conversation says why and offers Summarize again. A failed summary is never saved: before, the transcript could be saved in its place.

The field, redrawn. The field backdrop is drawn for each screen now. The sky, a low sun (a moon in dark mode) and the wheat stand around the content, never behind the words. The white disc behind the content and the sideways scrollbar are gone. The switch is in Settings › General › Advanced.

Every voice in the room. We replayed public meetings recorded by one microphone on the table through the app:

• Far voices. Oats now hears a voice across the table. In a room with steady background sound, it sends 85% of the speech to transcription, against 23% before.
• Quiet microphones. A quiet microphone gets its segments raised before transcription.
• Loud rooms. In a room that stays loud, the steady sound is taken out first.
• Sound labels. Lines that are only a sound label, such as "[inaudible]", are left out of the transcript.

A better speech model for your Mac.

• Apple Silicon: Oats uses Whisper small.en for English.
• Fast Macs: Oats times the GPU once. On a Mac that is fast enough, it downloads Whisper Turbo (574 MB) and switches to it when the download is ready. The download never starts while you record.
• Your choice wins: a model you chose yourself is never changed.

English by default. The setup screen has a Language row, and Settings changes it at any time.

Topics that say what they are about.

• On your Mac: a topic is named with a phrase someone said, not with two frequent words.
• With a writing model: the writing model gives each topic a short name.
• Live with OpenAI or Claude: if your writing model is OpenAI's or Anthropic's, a fast model checks each change of subject while you talk. A new topic then shows after about a minute, not four.

Summaries you can trust.

• Defaults: with an OpenAI key the writing model is gpt-6.1-sol, and with only an Anthropic key it is Claude Opus 5.5. These were the most accurate in our tests.
• Claude: Claude models work again as the writing model.
• No made-up attributions: the summary no longer guesses who said what when the transcript has no speaker labels.

Conversations that link by meaning. The board links conversations by what their topics mean, not by shared words, so unrelated conversations no longer link on words like "meeting".

Fixed. GPT-OSS 20B downloads again from Settings.

Known limits

• Long conversations on this Mac. The writing model on this Mac summarizes a conversation of up to about an hour. A longer conversation says that it is too long for the model on this Mac. With an API key, it is summarized.

Privacy

There is no Oats account and no Oats server.

On this Mac, audio and transcripts stay on your Mac. The models download from Hugging Face.

With an API key, these requests go to your provider, with your key:

• Audio, for transcription, if you have an OpenAI key.
• Transcripts, for titles and summaries. Opening a conversation that has no summary also sends its transcript.
• The live topic check: at most 3,000 characters of the conversation around each change of subject. It runs only with OpenAI or Anthropic.
• The link check: topic names and short gists, never transcripts. It sends only conversations whose summary that same provider already wrote.
• The key check: when you save a key, Oats asks the provider whether the key works.

The sidebar counts every network request since launch, and the model downloads are counted too. You can turn the web search for an unanswered question off in Settings.

Tell people before you record. Some places, for example California, need everyone's consent.

The SHA-256 checksums are in SHA256SUMS.txt.

Oats 0.1.0

The first public version of Oats: a high-signal tool for high-signal conversations.

Download

• Apple Silicon (M1 and later): Oats-arm64.dmg
• Intel: Oats-x64.dmg. It is built for Intel, but nobody has tested it on an Intel Mac yet.

Oats needs macOS 12 Monterey or later.

The first time you open it

This version is not notarized by Apple yet, so macOS stops it once.

1. Open the DMG and drag Oats to Applications.
2. Open Oats. macOS says that it cannot verify the app. Click Done.
3. Open System Settings, then Privacy & Security. Scroll down and click Open Anyway next to "Oats was blocked".
4. Confirm with your password or Touch ID. macOS asks only once.

What it does

• Press ⇧⌘O in any app to start or stop. Oats records the conversation in the room with one microphone and transcribes it on your Mac.
• Each question gets a verdict in place: Answered, Unknown, Unanswered or Open.
• New: the context check. Turn it on in Settings and choose a folder of your team's notes. When an answer in the room is unsure, Oats checks it against that folder and your past conversations, and says "That's right" or "Wait, that's wrong", with the passage it used. It uses the writing model you chose.
• Optional, off by default: when the folder has no answer, Oats can ask Claude Code to look in GitHub, Linear and Notion, read-only.
• When you stop: a title, a summary, the threads, talk time per person, and how each question ended.
• Search every transcript, and see all your conversations on one board.

Titles, summaries and the context check need a writing model: a local one that you download in Settings, or your own API key.

Privacy

Audio and transcripts stay on your Mac. There is no Oats account and no Oats server. Summaries run on a local model, or on your own API key if you add one. When a question meets a clear "I don't know", Oats opens a web search for the question text in your browser. You can turn that off in Settings. The sidebar counts every network request since launch.

Tell people before you record. Some places, for example California, need everyone's consent.

The SHA-256 checksums are in SHA256SUMS.txt.