Privacy
What leaves your machine when you self-host OpenShorts
Self-hosting takes OpenShorts out of the loop, but three optional AI stages still call out to model providers. This is exactly what each one sends, and how to turn it off.
The short answer
Self-hosting removes OpenShorts from the loop. There is no OpenShorts server on that path, no account, and no upload to anyone's bucket: the source file, the transcript, the frames and the finished MP4s live on the machine you started. What self-hosting does not do is make the pipeline offline. Three optional AI stages call out to model providers, and which of them fire depends on what you switch on.
What stays on your machine
- The source video. A YouTube link is downloaded with yt-dlp onto your own disk; files you upload never leave it.
- Transcription. faster-whisper (or Parakeet on a GPU box) runs locally. The audio is not sent anywhere to be transcribed — which is also why a first run is slow on CPU.
- Cutting, reframing and rendering. MediaPipe and YOLOv8 choose the layout, FFmpeg and the Remotion render service produce the MP4. All local, all in
./output. - The dashboard and the job queue. They talk to the backend on your own port 8000.
What can leave it, and when
| Stage | What is sent | When it fires |
|---|---|---|
| Moment picking | The transcript, plus the scene frames it samples | Every clip job — this is the core AI step |
| Layout analysis | The frames it samples for scene and layout decisions | During layout picking; this stage can include the full source video |
| Silent-video handling | The source video | Only for videos with no usable audio |
| Dubbing | The audio track | Only if you enable ElevenLabs and bring a key |
| AI Shorts (UGC) | The actor photo and the prompt | Only if you use that tool; it goes to an image model (fal.ai) |
| Backups | The finished clips | Only if you configure S3-compatible storage |
Everything in that table is opt-in or switchable except moment picking, which is the whole point of the tool. If a video cannot be sent to a model provider at all, the pipeline has nothing to score and will not produce clips.
Making it fully local
The moment picker speaks to any OpenAI-compatible endpoint. Point LLM_BASE_URL at Ollama, LM Studio or vLLM and no Google key is needed; the transcript never leaves the machine.
Be precise about what that buys. A local model replaces the scoring step, not the frame-based stages: layout analysis and silent-video handling still call Gemini unless you leave those features off. So "local model" = your transcript stays home and your bill drops to zero for that stage, while the visual pipeline still needs a key. The API keys page walks through the exact variables.
What self-hosting does change
Three things, and they are the reason people choose it:
- No account, no plan, no quota. The
/mcpendpoint on your own box is not authenticated and has no minute cap, because it is your machine. - No copy on someone else's disk. A hosted tool stores your source to process it. Self-hosted, the only copy is the file you already had, plus what the pipeline writes into
./output. - Deletion means deletion. Remove the folder and it is gone; there is no retention policy to ask about. That cuts both ways — nothing is backed up unless you back it up, and nothing is pruned unless you prune it.
What it does not change
- The model provider still sees what you send it. Self-hosting shifts who holds the file, not what Google or ElevenLabs receives when a stage calls them.
- Your keys are still your keys. Anyone with access to
.envcan spend on your account, and a compromised machine leaks the files themselves. - A hosted tool has a privacy policy; you have a terminal. There is no one to ask about retention except yourself, and no audit trail unless you keep logs.
If you need nothing to leave
Run the moment picker on a local model and turn off the frame-based layout stages. You lose automatic layout choice and silent-video handling, and you keep every frame on your disk. For most interviews and talking-head videos that is the right trade; for screen recordings and multi-speaker footage, layout picking is exactly what you would miss. The install guide covers the variables that switch each stage, and troubleshooting covers what breaks when a stage cannot reach its provider.