Reference
Privacy: what leaves your device
What may leave your device, and when
Ghira is built to work offline. This page defines "offline" precisely, so anyone
can check it against the code (all network access lives in crates/ghi-net).
The rules
| Class | What | Rule |
|---|---|---|
| Your content: audio, transcripts, notes, voiceprints, meeting titles, people | Never sent | Only two exceptions, each an explicit opt-in: cloud AI (transcript text only, after you review exactly what will be sent) and sync with your own paired devices over your local network (end-to-end encrypted, pinned keys) |
| Content-free traffic: model downloads, update check | Only to an allowlist. Model downloads: huggingface.co and its CDN (hf.co). Update check: GitHub Releases (github.com and its download host *.githubusercontent.com). | Strict offline blocks all internet traffic; sync with your paired devices on your private network stays allowed |
| Telemetry / analytics | None | Crash reports and an event log (no meeting content) are written locally (Settings → About → Diagnostics). You can open and review them, and send them yourself if you choose |
| Fonts, icons, UI assets | Bundled with the app | Nothing is loaded from the internet at runtime |
Ghira connects to no other service, and there are no integrations with other services yet.
Model downloads
The speech and notes models are not bundled. Ghira downloads them from Hugging Face, at a pinned revision, and checks each file's SHA-256 before use. Nothing is sent but the request for the file. Strict offline blocks the download; you can install the files from a copy instead (see Speech models).
Update checks
This build has no update feed, so it checks for nothing and sends nothing. Once signed releases exist, the check will ask GitHub Releases for a public manifest and nothing else: no ID, no meeting data. You can switch it off (Settings → About → Check automatically), and Strict offline blocks it.
What we promise, and how it is checked
- With Strict offline on, recording a 60-minute meeting and processing it
makes zero outbound connections. The release checks measure this with a
socket audit (
tools/release/net-audit.sh) and firewall logs (Little Snitch on macOS). - In the default mode, no network request contains your content. The release checks measure this with an intercepting-proxy capture.
Cloud AI
Cloud AI is off by default and chosen per meeting. You bring your own API key; Ghira runs no server and never sees your key or data. Before anything is sent, a preview shows the exact text, with names and other personal details you can redact. Audio is never sent, and neither are the notes you type yourself. Each request is recorded on your device (provider, model and token counts, not the text), and the meeting is marked as having used cloud AI.
Voice profiles
Recognizing a speaker by voice uses a voiceprint, which is biometric data. Ghira only creates one with explicit consent, stores it encrypted on your device, and deletes it permanently when you ask.
Calendar
Calendar access is optional and off until you connect it. Ghira reads your local calendar on your device (EventKit on macOS and iOS, or one ICS file you choose) only to name a meeting and know who attends. Events are read when needed and are not stored. Only a meeting you record keeps its event's title and attendees, encrypted with that meeting. Nothing from your calendar is sent anywhere.
Sensitive meetings
A meeting marked sensitive keeps no audio: none is written if it is sensitive from the start, and any audio already written is deleted. It is never sent to cloud AI and is never used to learn a voice. Only its transcript and notes are kept, encrypted like every other meeting.
Edit this page on GitHubLast updated