๐Ÿ‡ช๐Ÿ‡บ EU-resident ยท GDPR-aligned

What InFlow sees, what it stores, what leaves your Mac.

InFlow is built around a single rule: nothing leaves your Mac unless you asked it to translate that exact piece of text. No background scraping, no telemetry pings, no analytics SDK, no account, no cloud sync.

Last updated 2026-06-02 ยท matches v0.1.0
At a glance

Six things to know, before the long read.

Dictation: 100% on-device

Voice input runs through Parakeet, fully on-device on the Apple Neural Engine. Your microphone audio is never uploaded, never written to disk after the recording ends, never used to train any model.

Translation: EU-resident LLM

The selected text or transcript is sent over HTTPS to your LLM provider. The default is Mistral โ€” a French company, EU-hosted, GDPR-friendly. Prefer another provider? Bring your own key for OpenAI, Anthropic, Groq, or your own self-hosted endpoint.

No telemetry, no analytics

InFlow does not ship any analytics SDK, crash reporter, or usage tracker. The only network calls are: to your LLM provider for translation, one-time Parakeet model download on first launch, and the feedback endpoint when you explicitly press "Send".

No account, no sync

There is no InFlow account. Your profiles, hotkey bindings, and API keys live in your Mac's Keychain and local app container. Reinstalling on another Mac requires re-entering your API key โ€” there is no cloud copy.

Signed & notarized

InFlow is signed and notarized by Apple, with the hardened runtime enabled and no debug entitlements. macOS verifies the signature before it ever runs.

You can self-host

Point InFlow at your own OpenAI-compatible endpoint โ€” Ollama, vLLM, llama.cpp server. Your text never leaves your network. The app doesn't care which model is on the other end.

In detail

The long read.

Everything we send, store, and ask permission for โ€” point by point.

1

What gets sent to the LLM provider

When you trigger a translation, reply, or dictation cleanup, the following payload is sent to the LLM you've configured:

  • The text you selected (Select & Translate), the transcribed dictation (Voice โ†’ Translate / Dictate), or the source message you replied to (Auto / Guided Reply).
  • The system prompt of the active profile โ€” that's what tells the model "translate to French, formal tone, no preamble".
  • For replies: the intent you spoke (Guided Reply only).
  • For Screen Context (off by default): nearby on-screen text from the focused window, used to give the model context. Toggle off in Settings โ†’ Privacy.

Nothing else. No device fingerprint, no IP allow-list ping, no session cookie, no advertising id.

2

Provider retention

InFlow does not retain anything on its own servers โ€” it doesn't have any. Provider-side retention:

  • Mistral. Zero-retention available on enterprise plans; default 30 days, not used for training.
  • OpenAI. 30 days, not used for training (API default).
  • Anthropic. 30 days, not used for training.
  • Groq. Per their terms, not used for training; retention per their public policy.
  • Self-hosted. Whatever you set.
3

Feedback inbox

When you press "Send feedback" from the Help menu, the modal sends: your free-text message, the app version, the macOS version, and an opaque per-install random identifier. No selection contents, no transcripts, no API keys. The endpoint is on InFlow's own server at inflow.xavier-kain.fr/api/feedback.php.

4

Keychain & local storage

API keys are stored in your Mac's Keychain under the com.inflow.app service. Profiles, hotkeys, and preferences are stored in ~/Library/Containers/com.inflow.app/Data/. Dictation history (off by default) is kept in a local SQLite via SwiftData โ€” never uploaded.

5

Permissions InFlow asks for, and why

  • Accessibility. Required to read selected text from the focused window and to paste the result back. macOS gates this; you grant it once in System Settings.
  • Microphone. Required for voice mode. Recording stops the instant the pill closes; audio buffers are released.
  • Screen Recording. Optional โ€” only requested if you enable Screen Context. Used to read on-screen text near the focused field. Disabled by default for fresh installs.
6

What the web demo does

This website's live demo runs on Mistral end-to-end: the voice tabs transcribe with Mistral Voxtral and translate with Mistral Small โ€” both EU-hosted, with audio discarded immediately after transcription. The macOS app goes one step further: it transcribes with Parakeet on-device, so your audio never leaves your Mac at all.

Demo: EU end-to-end. App: audio never leaves the machine.

7

Auditable on the wire

You don't have to take our word for it. Run Little Snitch or a network sniffer while you use InFlow โ€” the only outbound calls you'll see are:

  • Your chosen LLM provider, only when you translate.
  • A one-time on-device model download on first launch.
  • The feedback endpoint, only when you press "Send".

No analytics, no crash reporter, no telemetry, no phone-home. InFlow is signed and notarized by Apple with the hardened runtime enabled, so the binary you run is the one we shipped.

Download and run it.

Signed, notarized, ad-free, ~30 MB. macOS 14.4 or later, Apple Silicon.