BaoCut: an AI video agent where every caption, dub and cut stays editable

BaoCut is a desktop app (macOS 14+ on Apple Silicon, Windows x64 in CPU, CUDA and Vulkan builds) from Jim Liu, known on X as 宝玉 (@dotey). The repo started in July, but it's moving fast: three releases in four days, ending with 3.9.0 on Oct 19, which adds versioned project videos that don't depend on Git. @GitHub_Daily featured it this morning. It had 1,122 stars when we checked. Its core idea is the one we keep saying matters most for agent editing. Fix one subtitle line, swap one shot or redo one dubbed sentence, and the agent changes only that, tells you exactly what it changed, and re-renders nothing else.

What it does with footage: transcription with local models (Whisper large-v3/v3-turbo, Qwen3-ASR, MOSS with speaker labels) or online APIs, with word-level timing. Subtitles in the original language, translated, or both. Dubbing that separates the voice from the background with HTDemucs, then synthesizes each sentence with a preset or cloned voice. Talking-head cleanup that removes filler words, long pauses and retakes without touching the original file. Chapters, summaries and shorts written from the timed transcript. From a brief, it has 24 scene templates (explainers, launches, trailers, data stories, kinetic type). Motion graphics are code: the agent writes a self-contained HTML piece, BaoCut bakes the frames and drops it on the timeline. It accepts its own baocut/1 format and also hyperframes/1, so HyperFrames work carries over.

The agent side is open. The app drives Claude Code or Codex CLI (both validated end to end), and Copilot CLI, OpenCode and any Agent Client Protocol agent can plug in. If you'd rather use your own agent, baocut skill install and baocut mcp install expose the same tool catalog, and there's a JSON-emitting CLI for scripts. baocut web open gives a remote agent a one-time link to show you the editor in a browser. The README's 'bounded autonomy' rules are worth copying anywhere: the agent works within a scope and budget, asks before destructive actions, treats downloaded pages and transcripts as data rather than instructions, and reports a failed check as a failure.

Read the fine print before using it for client work. BaoCut uses its own source-available Community License, not an open-source one. You can use it for free to make and sell videos, including paid editing services, but selling, hosting or embedding BaoCut itself needs a separate licence. Some local model weights it offers (OmniVoice, Fish Audio S2 Pro, Qwen-Image) are non-commercial, and the app warns you before you pick them. FFmpeg and ffprobe aren't bundled, Windows builds are unsigned, and GPU inference on Windows hasn't been tested on real hardware yet. Only clone voices you have permission to use. Spotted, not tested. Credit Jim Liu.