The conversations that matter happen around a table. Taking notes pulls you out of them, and meeting bots join calls, not rooms. Oats sits on the table with one microphone. While you talk, the screen shows the shape of the conversation and nothing that competes with the people in front of you.
What it does
Watch the conversation take shape
Press ⇧⌘O and talk. One line draws the conversation live: a notch where the topic changes, a mark where someone asks. Its height is how loud the room is. The sidebar steps aside while you record, and under the clock two quiet counters keep the words and the questions so far. Speech is transcribed on your Mac with whisper.cpp, on the GPU, and M marks a moment.
Topics name themselves
Each topic gets its name from what was said. When the room comes back to one, the line goes back too.
The conversation flow under the line keeps one lane per topic, in the order raised, with the minutes spent on each. Graph shows the same topics by when each was raised and how unsettled it still is. Oats finds the topics on your Mac and names each with a phrase said in it, and the writing model names it more exactly as it grows. Under the lanes are topics that came up with these in earlier conversations and have not come up yet.
Every question, caught as it's asked
Each question shows up the moment it's asked, and changes when the room answers it, or doesn't. Ask again before anyone leaves.
On the line, a question is a mark. With the transcript shown, it is also a card that says how it ended, in place, while people are still talking. A question asked twice is kept as a signal, not dropped as a repeat.
| Card | When |
|---|
| Answered | Someone gave an answer. |
| Unknown | The room said it did not know. Oats opens a web search for the question. |
| Unanswered | The room moved on. |
| Open | Nobody has replied yet. |
Unsure answers, checked on the spot
When an answer sounds unsure, Oats looks it up in your notes and past conversations while people are still talking.
It is off by default: turn on the context check in Settings and choose a folder of notes. Under the question, Oats says That's right, Wait, that's wrong, or Found it when the room did not know, with the answer and the file it came from. When the folder and your past conversations have no answer, Oats can also ask Claude Code, if you turn that on too: it reads your GitHub, Linear, Notion and Slack connections, read only, and the question and the reply go to Anthropic, and the search text to those services.
Written up when you stop
Title, summary, talk time for each person, and the transcript with every voice told apart.
The summary starts by itself when you stop, with the threads the conversation followed, the decisions and the action items, next to how each question ended. It is written by a model on your Mac, which Oats downloads by itself, or with your own API key. The transcript, the voices, the questions and the shape of the conversation need no writing model.
Find anything anyone said
Search every transcript at once, even for part of a word. Chinese and Japanese too.
Search reads every transcript on your Mac, including the oldest. It finds a word inside a longer one, so "rice" finds "price", and it works in Chinese and Japanese, which are written without spaces. Each search says how many matches it found and how long it took, and each match says when in the conversation it was said. Press / to search from anywhere in the app.
See what keeps coming back
Conversations on the same topic link up, so you see it come back across weeks.
The board lays out every conversation as a graph and as a timeline, so a topic that keeps coming back shows as the arcs between the days it came up. Each conversation is sorted into Work, Personal or a group you make by the model that writes your summaries, from its title, its topics and the start of its summary, never from the transcript.
Privacy you can check
On this Mac, audio and transcripts never leave it: transcription runs on whisper.cpp with Metal, and summaries on a local model. There is no Oats account and no Oats server. With your own API key, transcripts go to your provider for summaries, and with an OpenAI key the audio goes to OpenAI for transcription.
One thing is on by default, and you can turn it off in Settings: when a question meets a clear "I don't know", Oats opens a web search for the question text in your browser.
The sidebar counts every network request since launch. Calls to the speech, writing and search servers that run on your Mac are not counted, because they never leave it. The network allowlist lists every destination.
Small guarantees
- Nothing is lost. The transcript is saved as each sentence is finished, so a crash or a closed lid costs at most the sentence in progress. ⌘Q finishes and saves first.
- A dead microphone is reported while it still matters, not at the end.
- A conversation that captured nothing is not saved as a blank note.
- Stopping and starting again within half an hour continues the same conversation, so a coffee break does not split one meeting into two notes.
- Every key is named where you use it: on the button, in its tooltip, or in its menu.
Keyboard
| Key | Action |
|---|
| ⇧⌘O | Start or stop a conversation, from any app |
| M | Mark this moment |
| ⌘\ | Show or hide the sidebar while recording |
| / | Search conversations |
| ⌘F | Find in this conversation |
| Esc | Back |
| ⌘, | Settings |
The record shortcut can be changed in Settings.
Install
- Download Oats for Apple Silicon or Intel. It needs macOS 12 Monterey or later. The speech models ship inside, so it transcribes offline from the first launch: on Apple Silicon, Whisper
small.en for English and multilingual base for every other language; on Intel, base.en and base.
- Open the DMG and drag Oats to Applications.
- Open Oats. This version is not notarized by Apple yet, so macOS says it cannot verify the app. Click Done.
- Open System Settings, then Privacy & Security. Scroll down and click Open Anyway next to "Oats was blocked", and confirm with your password or Touch ID. macOS asks only once.
The SHA-256 checksums are in SHA256SUMS.txt. On an empty Home, open the example conversation, a nine-person launch review, to explore everything before you record.
Tell people before you record. Some places, for example California, need everyone's consent.
Build from source
You need Node.js 24 and the Xcode command line tools.
git clone https://github.com/alexgaoth/oats.git
cd oats
npm ci
npm run setup:check
npm run dev
Developing Oats has the full setup, the Linux notes and the test suites. The public page lives in site/.
How it works
Oats is an Electron app. The interface is React. The main process runs the local speech server (whisper.cpp), an optional local language model (llama.cpp) for summaries, local embeddings for related-conversation search, and SQLite with full-text search for the transcripts. The topic tracker and the question detector are plain code over the transcript, so they cost nothing to run for an hour and never need a network. Architecture maps the modules, DESIGN.md is the design reference, and the product description explains each behaviour in depth.
Principles
- Reliable like a pencil. It works with no network, no model and no account, it never loses a conversation, and it needs nothing explained.
- Alive, not busy. Nothing moves at rest. Motion answers what you did, and the recording view stays calm while people talk.
- Honest networking. What leaves the machine is listed, counted and switchable.
Acknowledgements
Oats grew out of OpenWhispr, an MIT-licensed desktop dictation app, and UPSTREAM.md records the provenance. Transcription uses whisper.cpp, local models run on llama.cpp, the interface is set in Inter, and the wordmark in EB Garamond. Topic names prefer distinctive words by their rank in English, from the Google Books Ngram corpus as counted by google-books-ngram-frequency (CC BY 3.0). The speed check that picks the speech model for a Mac reads one utterance from LibriSpeech (Panayotov et al., CC BY 4.0): test-clean 7127-75946-0024, read by Bill Kneeland for LibriVox.
Parakeet Ultra, a model you can choose on Apple Silicon, is moondream/parakeet-ultra (CC-BY-4.0), a post-training of NVIDIA's Parakeet TDT 0.6B v3 (CC-BY-4.0), in the Core ML build of FluidInference. It runs on FluidAudio (Apache-2.0), behind Silero VAD (MIT).
License
MIT
