JimLiu/baocut

BaoCut is a desktop app (macOS + Windows) where an AI agent can transcribe, subtitle, translate, dub, cut, and animate videos — but instead of handing you a finished MP4, every result is written into a real timeline you can keep editing. Errors are fixed locally (change one subtitle line, re-dub one sentence) without re-rendering everything else.

Core features:

• From footage: local or cloud transcription (Whisper large-v3/v3-turbo, Qwen3-ASR), subtitling with translation and per-project styles, dubbing into another language (voice separation + cloned/preset voices), cutting talking-head recordings to remove filler words and pauses, and generating chapters/summaries/shorts from timed transcripts.
• From a brief: generate a video from a prompt using 24 built-in scene templates, script-first narration production, and "motion graphics as code" — the agent writes self-contained HTML compositions that get baked into frames and placed on the timeline.
• Delivery: MP4 export with burned-in or sidecar subtitles, SRT/WebVTT, transcripts, audio stems, and a portable .baocut project bundle.

Architecture highlights:

• Four interfaces talk to one local Runtime: an Electron desktop app, a CLI with JSON output, an MCP service (so your own agent — Claude Code, Codex, Cursor, Gemini CLI — can drive BaoCut), and a web client for remote viewing.
• Bring-your-own models: local inference on Apple Silicon (MLX/Core ML) or Windows (whisper.cpp + candle, CUDA/Vulkan variants), or you can point capabilities at OpenAI, Anthropic, ElevenLabs, and many other endpoints. A usage ledger tracks cost per job.
• Design philosophy: the video is the source of truth; the agent works within explicit scope/budget; destructive actions require confirmation; it reports what it actually did and what still needs human review.