Retake

Cut your video by deleting words.
Drop in a recording, get a transcript, strike out the takes you fumbled, and export the video with them gone. Everything runs on your machine.

Install · Edit it with Claude · How it works

_{A 38-second recording with three false starts. Click the words, and it's an 18-second video.}

---

Why

I record improvised, off a teleprompter, and I say most lines two or three times before one comes out right. Cutting that on a timeline means scrubbing for every restart. In Retake you read the transcript, click the words you don't want, and the cut follows: the breath before a deleted sentence and the pause after it go with it, and a word you kept is never clipped.

What it does

• Transcribes locally with faster-whisper (the full large-v3, on the GPU when there is one), with word-level timestamps. It checks its own transcript against the audio and goes back for speech it skipped. Nothing is uploaded anywhere.
• Shows you the retakes. Words and lines you said more than once are colour-coded, and repeated attempts at a line are grouped as takes, so the restarts stand out before you've played anything.
• Cuts by word. Click a word to cut it, Alt+click or double-click to jump there. Play the final cut before you export.
• Cuts by instruction. Paste an edit list in English or Arabic ("cut 'so last week'", "keep the last take", a timestamp) and Retake resolves it to exact words for you to check. Anything it can't match, it tells you instead of guessing.
• Exports a real video. H.264/AAC MP4, NVIDIA encoding when available, checked packet by packet before it's handed over. Optional voice cleanup levels speech to −16 LUFS. You can also export EDL, SRT, CSV or plain text.
• Takes files from your phone. Open the LAN link Retake prints, pick a video, and the upload resumes on its own if the connection drops.

Install

You need Python 3.11 and about 3.5 GB for the models.

git clone https://github.com/AhmedHeshamEG/Retake.git
cd Retake
python -m venv .venv
.venv\Scripts\activate         # macOS/Linux: source .venv/bin/activate
pip install -r requirements.txt

Retake never downloads anything while it runs, so fetch the models into models/ once:

python -c "from faster_whisper.utils import download_model as d; d('large-v3', cache_dir='models')"
python -c "from sentence_transformers import SentenceTransformer as S; S('all-MiniLM-L6-v2', cache_folder='models')"

Then:

python retake.py

It opens http://localhost:8710. On Windows, Start Retake.cmd does the same with the .venv.

Docker works too: docker compose up --build, with -f compose.yaml -f compose.gpu.yaml for an NVIDIA GPU. Your models/ and projects/ folders are mounted in, so they survive rebuilds.

Precise word timing (WhisperX) and voice cleanup (DeepFilterNet3) are optional extras with their own runtimes. Setup for both is in the guide.

Edit it with Claude

Retake ships an MCP server, so Claude (or any MCP client) can edit for you: "cut the bit where I fumbled the sponsor read". The model works out what you meant; Retake decides which real words that is, and answers with token IDs and timestamps, or says it couldn't find them. A client can't delete a raw time range or invent a word that isn't in the transcript. Every edit previews first and can be undone.

{
  "mcpServers": {
    "retake": { "command": "python", "args": ["/path/to/Retake/retake_mcp.py"] }
  }
}

The running editor also serves the same tools over HTTP at http://localhost:8710/mcp. The retake-brain skill turns a raw recording into a cut list ("last complete attempt wins") and applies it through the server. The full tool list is in docs/MCP.md.

Projects and tests

Each project is a folder under projects/ with the original media, project.json and its exports. Opening a project again never transcribes it again.

python test_retake.py

Licence

MIT