Docs · Reference

Privacy & data

What stays on your Mac, exactly what is sent to your AI provider, and where Sift keeps its files.

Sift runs on your Mac. Your footage is never uploaded. Some content does leave your Mac: to label moments and learn styles, Sift sends still frames and text to the AI provider you chose, under your own API key. This page lists exactly what.

Stays on your Mac

  • Your footage and audio. Sift reads them and doesn’t change them.
  • Transcription. Whisper runs on your Mac.
  • Sync, scene detection, take grouping and frame sampling.
  • Sound cues (applause, cheering), from a model that runs on your Mac.
  • The project database, cache, rough cuts and exports.
  • Your API key, in the macOS Keychain. It is never written to Sift’s settings, database, logs or diagnostics.

Sent to your AI provider

Only while Allow API calls is on, and only to the provider in Settings → Model provider (Anthropic or OpenRouter):

When What is sent
Labeling an event Sampled still frames, scaled to 512 px on the long edge. With each frame: its device label, clock time and the transcript heard at that moment. The prompt also has the event name, the cast names you entered, and example frames from your style profile.
Finding transcript moments Timestamped transcript text, in chunks.
Learning a style profile One keyframe per shot of the videos you picked, with their transcript.
Writing style notes The profile’s measured statistics and example frames.
Reading shot lists The PDFs you added, if any.

Audio is never sent. Neither are whole video files.

With OpenRouter, Sift sets data_collection: deny, which asks OpenRouter to route only to providers that don’t store or train on your data. Each provider’s own policy applies to what it receives: see Anthropic’s and OpenRouter’s.

Sent to other services

  • YouTube, through yt-dlp, when you make a style profile from a channel: Sift lists the channel’s videos with their thumbnails and downloads the ones you pick, at 720p or less. The downloads stay on your Mac.
  • Hugging Face, once, to download the Whisper speech model during setup.
  • GitHub, when you press Update yt-dlp in Settings.

Sift has no analytics or telemetry. We don’t receive your footage, transcripts, events or key.

Turn the API off

Switch Allow API calls off in Settings and nothing is sent to an AI provider. Scanning, sync, scenes, transcription, takes and frame sampling keep running. Transcript moments, vision labels and style profile analysis pause.

Where Sift keeps things

Everything is in ~/Library/Application Support/Sift:

Folder Holds
pgdata The database: events, devices, placements, transcripts, moments, rough cuts, activity
backups Automatic database backups
bin The tools Sift unpacks on launch
models The Whisper model
profiles/<id> Each profile’s downloaded videos, keyframes and analysis
logs server.log

Exports go to the output folder you chose. Delete event… removes an event’s data (placements, transcripts, moments, review items, rough cuts and cached previews) but never your footage or exports.

Diagnostics

Settings → Diagnostics → Copy diagnostics saves a zip of logs and tool versions to send with a bug report. It never includes media or your API key. Nothing is sent unless you send it.