Edit video by chatting with a coding agent — drop raw takes in a folder, get a cut, graded, subtitled final.mp4 back. The model never watches the footage; it reads a word-level transcript and a rendered timeline.
The LLM never watches the video. One ElevenLabs Scribe call per source gives word-level timestamps, speaker diarization, and audio events; every take packs into a single ~12KB markdown file that is the agent's primary reading view. A timeline_view composite — filmstrip, waveform, word labels, silence gaps — is rendered only at decision points, like an ambiguous pause or a retake comparison. Because the surface is the transcript, cuts land on word boundaries. It's the same swap Browser Use makes by handing an agent a structured DOM instead of a screenshot, applied to video.
From there the pipeline is opinionated: filler words and dead space go, every segment is color graded, cuts get 30ms audio fades so there's no pop, subtitles burn in as two-word uppercase chunks, and animation overlays are generated by parallel sub-agents through HyperFrames, Remotion, Manim, or PIL. The agent then runs timeline_view over its own render at every cut boundary and fixes what it finds before showing a preview. Session memory persists in project.md.
Two things to budget for. It needs an ElevenLabs API key and a local ffmpeg, so a run costs transcription minutes rather than nothing. And it installs by cloning the repo and symlinking it into the agent's skills directory — ~/.claude/skills, ~/.codex/skills — not through a package manager, so updates are a git pull.
MIT, ~20k stars.