// the take
BaoCut is a desktop video editor (Electron + Rust native workers) built so an AI agent and a human edit the same project through a shared timeline instead of the agent just spitting out a rendered MP4. It's aimed at developers who already live in Claude Code or Codex CLI and want those agents to transcribe, subtitle, translate, dub, and cut video via CLI/MCP rather than a GUI.
The core idea is solid: edits go through a transactional command gateway with receipts and a shared undo stack, so a 'fix this one subtitle line' request only touches that line instead of re-rendering everything — a real architectural answer to the usual one-shot-render problem. Capabilities (transcribe, speak, generate image) are decoupled from providers, so you can run fully local (Whisper, MLX/Core ML on Mac, candle/whisper.cpp with CUDA or Vulkan on Windows) or swap in OpenAI/Anthropic/ElevenLabs/etc. per capability, with a usage ledger tracking cost per job. The protocol is genuinely shared across CLI, MCP, desktop and web clients rather than the CLI being a bolted-on afterthought, and the portable .baocut bundle lets you hand off a project (with version history) to another machine.
It's source-available, not open source — the BaoCut Community License forbids reselling, rebranding, hosting it as a service, or embedding its code/engines in a commercial product without a separate license, which matters if you were hoping to build on it rather than just use it. Agent features are a hard dependency on a signed-in Claude Code or Codex CLI session (other engines like Copilot, Pi, OpenCode, and ACP agents are wired in but explicitly 'not individually tested'), and ffmpeg/ffprobe aren't bundled so you're assembling a fairly tall stack just to get transcription working. It's early: Windows builds are unsigned, GPU inference is 'untested on hardware', there's no Linux installer or release validation at all, and some of the better local voice/image models (Fish Audio S2 Pro, OmniVoice, Qwen-Image) carry non-commercial licenses, which undercuts the README's claim that you can freely sell output made with it.
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →