JayWebtech/autoshorts

Worth a look if you publish vertical clips from long interviews, talks or podcasts and want candidate moments and rough cuts produced on your own machine. Expect to supply FFmpeg, an LLM provider and an API key or local Whisper setup before the pipeline runs end to end.

The problem

Turning an hour-long recording into short vertical clips normally means scrubbing a timeline, guessing which exchanges will travel, cutting each clip and then reframing horizontal footage for a phone screen. Each step is repetitive and easy to postpone, so good moments stay buried in long files. AutoShorts collapses that chain into an import-transcribe-rank-crop flow: you pick a project, the transcript appears, ranked candidate moments come back with hooks, and selected segments are rendered as portrait H.264 clips. The result is a shortlist and draft cuts rather than a finished, styled edit.

What's different

The design centres on the ranking step rather than clip editing. Moment detection and hook analysis go to DeepSeek by default or Claude, and the project stores transcripts, candidates, custom names and rendering data in a local SQLite file. The documentation is unusually candid about its own weak spot: it says local Ollama models such as LLaMA 3.2 3B or Qwen 2.5 3B/7B are supported for offline analysis but generally not recommended, because small models often produce fragments that are too short when calculating segment timestamps. That is a useful trade-off statement: offline operation costs quality here, and the cheap cloud path is the one the author points to.

Cropping is also blunt by design. The README describes automatic center-cropping of landscape video, so footage where the speaker is off to one side will need checking rather than trusting the automatic framing. Captions depend on the installed FFmpeg build having drawtext and subtitle filters, which is why macOS users are directed to a special Homebrew formula. The app is local-first for storage and media handling, but transcription with Deepgram or analysis with DeepSeek/Claude still sends content to those services; only the Ollama plus Whisper setup keeps the whole chain on the machine. The README gives no license, which matters if you plan to modify or redistribute the code.

Who should try it

A creator or small podcast team that already pays for an LLM API and wants a fast first pass over a long recording is the clearest fit: the app will surface moments to review and render draft portrait clips, and the author's own notes say cloud analysis gives the best results at a low per-run cost. A developer comfortable installing FFmpeg and managing keys also fits, since setup expects both.

Poor fits are editors who want fine control over framing, b-roll, caption styling and transitions; this is not that kind of editor. People who need a guaranteed offline pipeline should weigh the documented quality warning about small local models before committing, and anyone in a regulated setting should note that cloud transcription and analysis send recordings to third parties.

Requirements and limitations

The hard prerequisite is FFmpeg and FFprobe on your PATH; without them audio extraction, cropping and captions fail, and the README warns that some FFmpeg builds lack the filters captions need. You then choose an analysis path: cloud keys for Deepgram, DeepSeek and Claude, or a local setup with Ollama plus openai-whisper installed through Python. The UI surfaces missing environment variables, which helps, but this is still a setup you have to assemble.

Prebuilt packages are on the releases page, and macOS builds are unsigned, requiring the Gatekeeper bypass described there. Windows packages are self-signed and trigger SmartScreen. No license is declared, and there is no stated statement of compatibility beyond macOS, Windows and Linux packages. Ongoing care means keeping FFmpeg and Python, if used, working, and monitoring API keys and provider changes.

HotRepo's take

The repository is young: created 2026-06-22, at 1,323 stars and 257 forks when HotRepo recorded it on the 2026-10-26 daily board, a single observation with +14 stars that day. That is one snapshot, not evidence of sustained momentum. The decision point is the documented quality gap between local and cloud analysis; check the project instructions on provider setup before assuming the offline path meets your standard.