freecut is an open-source video editor forked from browser-use/video-use, replacing its paid ElevenLabs dependency with free, pluggable transcription backends like local Whisper. It enables LLM-driven video editing via chat, automatically removing filler words and dead space, with zero required API keys for the default setup.

How It Works

freecut enables LLM-driven video editing by processing video through a text-centric pipeline. The LLM receives word-level audio transcripts and on-demand visual summaries (filmstrips, waveforms) generated only at critical decision points, avoiding raw frame analysis. This approach, combined with automated transcription, packing, reasoning, EDL generation, rendering, and self-evaluation, allows precise cuts based on speech and silence. It replaces proprietary ElevenLabs Scribe with pluggable, free transcription backends, defaulting to local Whisper.

Quick Start & Requirements

Installation involves cloning the repo, linking it to your LLM agent's skills directory, and installing core dependencies (uv sync or pip install -e .). A local Whisper backend (mlx-whisper for Apple Silicon, faster-whisper otherwise) is required, along with ffmpeg system dependency. No API keys are needed for the default Whisper setup. Optional VibeVoice-ASR diarization requires CUDA and an endpoint URL.

Highlighted Details

• Automated removal of filler words and dead space.
• Automatic color grading for video segments.
• Customizable subtitle generation.
• Parallel sub-agent spawning for animation overlays.
• Self-evaluation loop at each cut boundary.
• Persistent session memory for continuity.

Maintenance & Community

The provided README does not detail specific contributors, sponsorships, community channels, or a public roadmap.

Licensing & Compatibility

The project inherits the MIT license from its base tool (browser-use/video-use) and uses the MIT-licensed microsoft/VibeVoice-ASR for optional diarization. Default local Whisper backends are permissively licensed. The core functionality is highly compatible with private and offline workflows.

Limitations & Caveats

The VibeVoice-ASR backend for multi-speaker diarization is CUDA-only. The default Whisper backend supports only single-speaker identification. The LLM's understanding is derived solely from processed text and generated visual summaries, not direct frame analysis.