The Markdown-first AI meeting notetaker

Yes. Your recordings live in your own vault folder, transcription and speaker detection run on your computer, and your audio never leaves it unless you choose to contribute a recording. With local models (the default), your transcripts, summaries and notes never leave it either. We have no cloud servers to store your meetings, no account, and no access to your conversations.

Three things can send meeting text elsewhere, and each one is your choice: a cloud model you pick for summaries, polish or the assistant; a connection such as Linear or Notion, when one of its tools is used; and an AI app such as Claude Desktop that you connect, which sends what it reads to its own AI service. Separately, you can send a meeting to the developer yourself: with a bug report, or by contributing a recording to help improve transcription. All of these are spelled out in the network disclosure.

No, and Samurai Scribe makes no such claim. There is no official HIPAA certification for software: compliance is about how a practice handles patient information, and an app is one part of that. Samurai Scribe does not sign Business Associate Agreements, because the app is built so that your patients' data never reaches its developer. What can be set out is what the software does, so you and whoever handles compliance for you can decide.

What the app does

• Your audio is never sent anywhere unless you choose to contribute a recording to help improve the app, which asks you to confirm that everyone in it agreed. With local models (the default), nothing from your meetings is sent either: there is no account and no sync, and nothing reaches the developer unless you send it.
• From version 0.63.0, the app's logs leave out meeting titles, people's names, email addresses, file names and what was said, so a bug report's logs carry none of them, and logs from earlier versions are deleted when it first starts. A meeting's text reaches the developer only if you pick that meeting under Include a meeting in a bug report.
• Recordings, transcripts and notes are ordinary files in your vault folder. The app does not encrypt them itself and has no passcode of its own, so they are protected only by your computer's login and disk encryption.
• There is no audit log of who opened, exported or deleted a meeting.
• Deleting a meeting moves its folder to your Trash or Recycle Bin, where it stays until you empty it, and removes what the app kept of it elsewhere: the text sent to the AI model, its cached audio and waveform, and any link import of it. Two things can stay behind: any assistant chat that used it, until you delete that chat, and what the pictures in it were read as, kept on your computer for up to 30 days.
• Dictated text is kept out of your clipboard history. The app keeps its own history of what you dictated, on your computer: 30 days on a new install, and everything on an install from before 0.63.0, until you change Keep dictation history in Settings → Dictation.

If you work with patient information

• Turn on FileVault (Mac) or BitLocker (Windows).
• Keep your vault out of iCloud Drive, OneDrive, Dropbox and any other synced folder, unless your agreement with that service covers it. The default vault is in Documents, which many computers sync, so check yours.
• Keep summaries, polish and the assistant on a local model. A cloud model sends meeting text to its provider; choose one only if you have a BAA with that provider. An Ollama on another computer counts as a cloud model, and the app marks it as one.
• The same goes for connections such as Notion or Linear, and for AI apps such as Claude Desktop that you let read your meetings.
• When you report a bug, leave Include a meeting at None, and keep patient details out of what you write.

Every connection the app can make is listed in the network disclosure, so this can be checked rather than taken on trust.

Yes. Recording, transcription, speaker detection, summaries, dictation and the assistant all run on your own computer, on its CPU or GPU, and keep working with Wi-Fi off once your models are downloaded. Transcription and speaker detection always stay on your machine. Summaries, polish and the assistant use the built-in local model by default; you can choose a cloud model for any of them instead, and the first time you do, the app tells you what will be sent and to whom.

Left alone, the app makes only a handful of network requests: a once-a-day update check (carrying anonymous usage data, which you can switch off), downloading an update if you choose to install one, downloading models, and — only if you buy Pro — checking that your licence is still valid. Connect a calendar and it will sync that too, read-only, to your machine. Import a video from a link and it fetches that video, plus — the first time, after you agree — the open-source downloader it uses, which then updates at most once a day and only when you import. Every one of them is listed, with what it carries, in our network disclosure.

Instead of inviting external bots to join your virtual meetings, Samurai Scribe records your system audio and microphone feeds directly at the OS layer. This allows you to capture Zoom, Google Meet, Teams, or Slack calls privately without displaying a bot in the call. When you first launch the app, macOS or Windows will ask you to grant permission to record audio. Windows needs nothing beyond that; macOS asks for a little more.

What it asks for today. To hear the other people on a call, the app needs to capture the audio your computer is playing. On macOS that currently means granting Screen Recording, which is broader than what we actually use — Samurai Scribe only ever reads the audio stream, and never captures your screen, your camera, or any video. macOS has since added a narrower "System Audio Recording Only" permission, and we are moving to it so the app stops asking for more than it needs. On Windows, capturing desktop audio needs no special permission at all.

What that permission can and can't lead to. Permission to hear your audio is only a risk if there is somewhere for that audio to go. There isn't. Your audio is never transmitted anywhere unless you choose to contribute a recording, and with local models (the default) nothing transcribed or summarized from it is either — there is no account, no sync, and no cloud processing of ours to send it to. Every network request the app is capable of making is written down in our network disclosure, and it's meant to be verified rather than believed: point Little Snitch or Wireshark at the app, record a full meeting, and you should see exactly that list and nothing else.

On third parties. The macOS build is signed with our Developer ID and notarized by Apple, so the copy you download is the one we built, and macOS refuses to run it if anything has altered it. The Windows installer, and the programs it installs, are signed through Microsoft’s Artifact Signing service under an identity Microsoft verified, so Windows can confirm the copy you run is the one we built, and a modified file no longer shows a valid signature. The heavy lifting is done by well-known open-source components — whisper.cpp for transcription, ONNX Runtime for speaker detection, ffmpeg for audio — and these are compiled into the app rather than fetched at runtime, so nothing can be swapped underneath you after you install. The one exception is the optional video downloader (yt-dlp, with Deno), which the app fetches from its official releases only when you install it or import a link, checks (yt-dlp against its maintainers’ signed checksums, Deno against a hash built into the app), and runs as a separate program.

Local by default, and there is nothing else to install. Samurai Scribe ships with its own AI engine (llama.cpp) and downloads a capable default model — Qwen 3.5 (4B), which writes clear summaries in over 200 languages — the first time you need it. Summaries, dictation polish and the assistant all use it, on your machine.

Prefer your own setup? Point any of them at Ollama or LM Studio, or at a hosted model (Anthropic, OpenAI, Gemini, OpenRouter or a custom address), all from one model manager. For the assistant, a larger model than the built-in one does better on long, multi-step requests.

Ask it about your meetings and notes, in a panel beside any meeting (⌘J on Mac, Ctrl+J on Windows) or full width in the Projects view. Answers cite the exact moment they draw on — click a citation to open the meeting at that moment, ready to play. It can suggest notes and edits, which wait for you to keep or discard in Review mode (the default) and can be undone later. It drafts emails for you to open in your mail app or Gmail; it never sends anything itself. The assistant is free.

Claude Code, Claude Desktop, Codex and Cursor can read your meetings too, over MCP, and that's free as well. Settings → AI Apps gives you the setup for each: a one-line command for Claude Code and Codex, or a block to paste into Claude Desktop's or Cursor's config file. Samurai Scribe needs to be open while they use it.

Folder rules decide what each app can see and change, and by default the changes they suggest wait for your review in Samurai Scribe. Connecting one sends nothing from Samurai Scribe itself, but the app you connect sends what it reads to its own AI service, under its own terms.

OpenAI Whisper natively supports 99 languages, including English, Spanish, French, German, Japanese, Mandarin Chinese, Italian, Portuguese, Dutch, Korean, Arabic, Hindi, Polish, and Swedish.

The app includes automatic spoken language detection and formats your local AI summaries in the same language as your conversation. NVIDIA Parakeet, free, recognizes 25 European languages, English included, automatically.

Translation comes in two forms. The assistant can translate a finished transcript into any language you name: open the transcript's ⋮ menu, choose Translate transcript… and type a language, such as German or Japanese. Every line keeps its time and speaker. Like any change the assistant suggests, it waits for you to keep or discard it in Review mode (the default), and a kept one can be undone later. It's free, and it works on any transcript, Parakeet's included. It runs on the assistant's model, the built-in local one unless you chose another, so a larger model does better on long meetings. Anything it can't translate is left as it was, and it tells you.

The other is an optional Translate to English toggle, which turns speech into English as it's transcribed (on any Whisper model, free ones included; Parakeet does not translate).

Yes. Drag a recording onto a meeting, or pick it from your files: MP3, M4A, WAV, FLAC, OGG, AAC, Opus and other audio, or the audio track of an MP4, MOV, MKV or WebM video. It gets the same speaker-labeled transcript and summary as a live recording, all on your computer, so a voice memo or an interview from your phone never leaves it. Importing a whole folder or many files at once, each as its own meeting, is part of Pro.

For video online, paste a link. The built-in downloader supports 1,000+ sites, including YouTube, TikTok, Instagram, Vimeo and podcast feeds: it downloads the media directly onto your computer, transcribes it locally, and saves clean Markdown into your vault. Nothing is uploaded, and the app sets up the downloader for you.

Not any more. Since version 0.65, new installs start on the free version, which has no time limit and stays free. To see what a Pro transcription model does on your own audio first, download it in Settings → AI Models → Transcription and try it in Compare models, side by side with the model you use. A trial started before 0.65 finishes its ten days; when it ends, the app swaps any Pro model for a free one (Parakeet Ultra for Parakeet v3) and tells you.

Pro is a one-time payment (see pricing), with a 7-day money-back guarantee. You can activate it on up to 2 machines simultaneously. If you upgrade your computer or need to switch machines, you can deactivate the license from the app settings on your old machine to free up a slot for the new one.

No. Meetily and a few others are, and if that is your deciding factor you should use one of them. What we offer instead is verifiability: the network disclosure, and output you can read without us — plain Markdown files in a folder you choose.

Skills are short Markdown files in the open Agent Skills format, and you can share your own in the public skills repository on GitHub (MIT-licensed).

A moderate to modern computer is recommended for the fastest, hardware-accelerated transcription results (such as Apple Silicon Macs or Windows PCs with compatible GPUs). However, transcription is fully supported on almost any computer using CPU fallback.

macOS

• Apple Silicon (M1, M2, M3, M4, etc.) with native Metal GPU acceleration. An Intel build is not available yet.
• macOS 13 (Ventura) or higher — required for system-audio capture. Microphone recording and file uploads also work on macOS 12.
• 8 GB RAM minimum.

Windows

• Windows 10 or 11 (64-bit editions).
• 8 GB RAM minimum (16 GB recommended).
• Compatible GPU (NVIDIA, AMD, or Intel) with Vulkan support is highly recommended for hardware-accelerated inference. (Most modern NVIDIA, AMD, and Intel graphics drivers include Vulkan by default — no separate CUDA or SDK installation required.)
• The installer adds Microsoft's WebView2 runtime if it's missing; the notes editor needs it.

Linux

• Not available yet. If you'd use it, add a 👍 to the Linux issue on GitHub — that count decides whether it gets built.

Storage

• Approximately 1.5 GB of free storage for the app and a speech-to-text model.
• About 3 GB more for the built-in AI model that writes summaries and runs the assistant, downloaded the first time you need it. Skip it if you use Ollama, LM Studio or a hosted model instead.
• Models are downloaded and managed directly within the app. Choose which models to keep locally and monitor disk usage from the built-in model manager.