Source: https://fnflow.app/private-dictation-mac-work

# Private dictation on a Mac at work: what actually leaves your machine
Before anyone dictates into a work Slack or a coding agent, someone on the security side asks the obvious question: where does the audio go? For cloud dictation the answer is a provider's server. For FnFlow it's nowhere, but the app does use the network for a few other things. Here's the complete list, how it compares with cloud tools in general, and a checklist your team can verify for itself. It describes what the app does and isn't legal or compliance advice.

## Audio and text are separate questions

Cloud dictation sends your voice to a server. Wispr Flow's privacy page says its transcription always happens in the cloud and the audio is deleted once it's transcribed, unless you turn on cloud storage. Claude Code's /voice streams recorded audio to Anthropic's servers. Apple's dictation runs on the device for many languages, and Keyboard settings tell you whether yours does. When it doesn't, Apple says what you dictate is sent to its servers.

Keeping audio local doesn't settle the question by itself. VS Code's desktop dictation recognizes speech on the device, yet by default it sends the transcript text to a Copilot language model for cleanup. So ask two things of any tool: where the audio is recognized, and where the text goes afterwards.

## What FnFlow keeps on the Mac

Everything sits in ~/Library/Application Support/FnFlow:

- Audio of each dictation in Recordings/*.caf, readable only by your user.
- Transcripts in Recordings/*.json, with the text, length, engine and the name of the app the text went into.
- Speech models in Models/, plus your terms in vocabulary.json.

The Keep recordings setting deletes old dictations after 7, 30 or 90 days, or never. Deleted recordings go to the Trash. Removing the folder removes everything. Debug logging in Console.app has timings and errors and never what you said.

## When FnFlow uses the network

- The larger model. After install, FnFlow downloads large-v3-turbo (874 MB) in the background, only on Wi-Fi or Ethernet, never in Low Data Mode or over a phone hotspot. The request goes to fnflow.app, which redirects to huggingface.co and counts that a download started, with the date, app version and country. The file is checked against a SHA-256 checksum before use.
- Anonymous usage counts. Short one-word events go to fnflow.app, such as "installed", "set up", milestones like the 1st or 50th dictation, setup steps and a once-a-day "dictated today". Each carries only the app version. There's no ID, no text and no settings, and the site keeps no IP address. Turn it off in Settings → General.
- Update checks. About once a day the app fetches appcast.xml from fnflow.app. Updates are signed and verified before installing, and automatic checks can be switched off in the same place.
- On macOS 26, if you click Prepare Language for Apple's speech engine, macOS downloads its own language model.

Whisper itself runs as a small local server bound to 127.0.0.1, so no other machine can reach it. Report a Problem only opens an email draft in your mail app, and you decide whether to send it.

## What's easy to overlook

The paste goes through the clipboard. FnFlow puts the text there briefly, sends ⌘V and restores what you had before. If your Mac runs a clipboard history tool, it may record dictated text the way it records any copy.

The destination still matters. Text pasted into Slack, ChatGPT or a coding agent goes wherever that app sends it, and local dictation can't change that.

Permissions are worth reviewing too. FnFlow asks for the Microphone and Accessibility, and on some Macs Input Monitoring to notice the dictation key.

## A checklist for security-minded teams

- Read the [privacy page](privacy) for the version you're installing. It states which version it applies to.
- Decide on the outbound calls. Usage counts and automatic update checks can each be turned off in Settings → General.
- Allow or block the model download on purpose. It needs fnflow.app and huggingface.co. Without it, FnFlow keeps working on the smaller built-in model.
- Watch the traffic yourself with your firewall or proxy while someone dictates. Recognition shouldn't produce any.
- Check the listener with lsof -nP -iTCP -sTCP:LISTEN during a dictation. Whisper's server should show up on 127.0.0.1 only.
- Set Keep recordings to match your retention policy, and check whether your backups include the user's Library folder.

For the technical background on running Whisper on the Mac, see [how to run Whisper locally](run-whisper-locally-mac). For a direct comparison with a cloud tool, see [FnFlow vs Wispr Flow](wispr-flow-alternative).

## Questions

**Does FnFlow send my audio or text anywhere?**
No. Recording, recognition, pasting, history and playback make no network requests. The text goes only where you paste it.

**Can FnFlow work with no internet at all?**
Yes. A Whisper model ships inside the installer, so dictation works offline from the first launch. Without a connection you won't get the larger model or updates.

**Is there an account or a server-side history?**
No. There's no sign-in, and every recording and transcript stays in a folder on your Mac.

**Does FnFlow meet our compliance requirements?**
We can't answer that for you and don't make compliance claims. This page and the privacy page list exactly what the app stores and sends, so your security team can check that against your own rules.
