Privacy
Esta página aún no está traducida; aquí tienes la versión en inglés.
Reelfold runs on your Mac, where it edits and renders your footage. Transcription and free narration can also run locally. A configured cloud AI step receives the inputs it needs: text and, for visual planning, reference frames or contact sheets. If you select hosted transcription, the audio track is sent to that transcription service. Hosted generation receives its requested script and reference inputs.
Why it matters
Section titled “Why it matters”Your recordings hold faces, voices, names, client material and things said off the cuff. Uploading hours of raw footage to a cloud editor is a real cost, especially for client work. Keeping the heavy work local also means no upload wait and no per-minute render fees.
What stays on your Mac
Section titled “What stays on your Mac”- Your media. Source files are read in place for local editing and rendering. Cloud services receive the inputs described below when those services are used. Intake only reads the source; nothing is written next to your files.
- Transcription. Whisper runs locally by default (on Apple Silicon). A hosted transcription API is used only if you choose it.
- Rendering. Cuts, captions, effects, covers and exports are all made on your machine.
- Keys. In the Mac app, API keys go into the macOS keychain; the interface only shows whether a key is set. In the skill, configs name an environment variable instead of holding the key, and a client config refuses key or token fields.
- Your persona.
persona.local.yaml(your speeds, colours, tags, term fixes, AI routes) is git-ignored, so it doesn’t end up in a commit.
The Mac app’s engine listens on no network port: the app reaches it over a Unix domain socket in the app’s own folder, with a per-launch token (on Windows: 127.0.0.1, a random port and the same token).
What is sent, and to whom
Section titled “What is sent, and to whom”Only to the provider you route each task to:
| Step | Sent | Not sent |
|---|---|---|
| Intake | Your request, a summary of the materials (file names, durations, short excerpts, document headings); a transcript when you ask to pick content from a recording | Audio, video, full documents |
| Segment planning | The transcript with timestamps | Audio, video, file paths |
| Visual story and performance planning | Contact sheets or reference frames, source descriptions and any relevant transcript | The full source video |
| Caption proofreading and glossary | Caption text and transcript text, your term list | Audio, video |
| Output edits | Your instruction, the clip’s captions or transcript, its edit state | Audio, video |
| Project and clip chat, autopilot decisions | Your message or the question, the clip’s transcript and settings, the options | Audio, video |
| Explainer and course scripts and scenes | Your notes or document text, the script, the scene timings | Audio, video |
| Post copy | The facts: place, month, your words and title, the music’s name | Audio, video |
| Hosted transcription (only if chosen) | The audio track | Video |
| Hosted voice (only if chosen) | The narration text | Anything else |
| Video generation, such as an AI presenter (only with your key, after you approve the spend) | What that shot needs: its prompt, and the picture or voice it animates | Your other footage |
To keep AI processing local, configure local models for every task and its fallback, use local transcription and voices, and leave hosted generation off. A custom or OpenAI-compatible endpoint is local only if the server itself runs locally; a remote endpoint still receives the task inputs. Downloads, update checks and publishing are separate network operations. Transcripts can contain names and private details of the people recorded; for client work, prefer a local model or a provider whose data terms your client accepts.
Logs on your Mac
Section titled “Logs on your Mac”The live view of a run (the Claude Code or Codex stream and the engine log) is kept on your Mac, with API keys, tokens and passwords removed. Copy diagnostic report (Settings › Advanced) holds versions, engine state and settings, never keys, passwords or video, and you decide where to paste it. Report this problem shows the whole report first (paths shortened to ~, no file names, transcripts, prompts or keys) and opens it on GitHub for you to post yourself. Sending crash reports automatically is off by default.
Downloads
Section titled “Downloads”The app downloads what it needs from public sources the first time: fonts, face models and the speech model, and, when a video first needs it, the free voice (about 330 MB, from Hugging Face). Nothing about you or your videos is sent with these downloads. See Settings › Downloads.
Analytics and updates
Section titled “Analytics and updates”The Mac app has one optional, opt-in feature that sends anonymous usage counts (such as “a batch of 12 clips finished”). It is off unless you turn it on; every field it sends, and how to delete it, is listed in Privacy: what Reelfold sends. The engine and the Claude Code skill send no usage data. Installed builds do check the public GitHub releases for updates (every few hours) and download fonts and models on first run. Setting DESK_DISABLE_UPDATES=1 turns the update check off.
Guests and other people in the frame
Section titled “Guests and other people in the frame”For calls, interviews and podcasts, you choose whose face to hide. A face-tracking sticker covers each chosen guest, with a coverage check that flags incomplete tracking, and the call app’s name labels are blurred by default. The call-clips recipe also stops at a consent checkpoint, confirming the people shown agreed, which is never answered automatically. Review the rendered masks before sharing a clip.
Reelfold never deletes your own recordings. Source cleanup for client batches is off by default, lists exact files first, and needs a confirm code.
Open source, so you can check
Section titled “Open source, so you can check”Reelfold is MIT-licensed. Everything on this page can be verified in the code: the repository, especially lib/vstudio/llm and the providers reference, which lists what each step sends.