Record in one take, badly. Get a publishable video in a minute. Nothing leaves your machine.
A local-first screen recorder that edits itself. Flub a line? Just say it again - onetake detects the retake and keeps your last attempt, removes filler words, fast-forwards dead air, and makes every cut look intentional. Everything runs on your own hardware.
- Desktop app (Tauri 2 + Svelte 5): record a chosen screen with webcam self-view bubble and mic, or open an existing video; review every proposed cut with a veto toggle; render and A/B the result.
- Auto-edit pipeline (Python CLI): fillers, long pauses, repeated-take detection; punch-zoom transitions, 8x speed-ramped silences, audio micro-fades, burned-in captions.
- Recording is X11 (Linux) for now. Wayland and macOS are on the roadmap.
- Python >= 3.11, uv, ffmpeg (with ffplay)
- GUI: Rust + Node, plus webkit2gtk dev packages on Linux
- Optional: NVIDIA GPU (CUDA) for ~10x faster transcription; falls back to CPU
# CLI
uv sync --extra gpu # or plain `uv sync` for CPU-only
uv run onetake process demo.mp4 --captions
# GUI
cd app && npm install && npm run tauri devprocess writes demo.edited.mp4 plus demo.edl.json, the edit decision
list. Every cut carries a reason (filler / pause / retake) and an
enabled flag; flip flags to veto cuts, then re-render without re-analyzing:
uv run onetake apply demo.mp4 demo.edl.json --captionsUseful flags: --dry-run, --max-pause 2.0, --pause-style cut,
--no-retakes, --model small, --json (machine-readable progress).
- Transcribe - faster-whisper (
large-v3-turbo, int8), word-level timestamps - Detect - fillers per-word; pauses from word gaps; retakes via sentence embeddings (model2vec): a near-duplicate sentence within the next 3 means the earlier one was a flub, cut it, keep the last attempt
- EDL - merged, human-editable JSON; the single source of truth
- Render - ffmpeg: punch-zoom across cuts, 8x fast-forward through pauses, 20 ms audio fades, optional caption burn-in