Browser Use drops video-use to let Claude Code edit raw footage

video-use is an open-source skill from the Browser Use team that lets terminal-based AI agents like Claude Code or Codex edit raw video footage directly.

You simply drop your raw video takes into a folder, launch your agent, and tell it what you want. The agent analyzes the footage, outlines an editing plan, waits for your approval, and renders out a final cut alongside your original files. It handles cutting dead space between takes, trimming false starts and filler words, auto color grading, and burning subtitles.

How it edits without watching video

Feeding raw video frames into an LLM is brutally expensive. A typical clip can easily burn 45 million tokens across tens of thousands of frames with little to show for it. video-use takes the same philosophy Browser Use applied to web agents: instead of full visual dumps, give the model lightweight structured text.

The system runs on two layers. The primary layer is a word-level audio transcript generated via a single ElevenLabs Scribe call. This captures precise timestamps for every word, speaker diarization, and non-verbal cues like laughter or applause, packing the entire shoot into a compact markdown file of roughly 12KB.

The second layer is visual inspection on demand. The agent only generates a visual timeline composite—showing filmstrip frames, audio waveforms, and word labels—at critical decision points, like ambiguous pauses or cut-boundary sanity checks.

Once the cut list is rendered via ffmpeg, a built-in evaluation loop inspects the cut points to catch visual glitches, subtitle overlaps, and pop noises. Every cut point also includes a 30-millisecond audio fade to prevent pops.

Setup and requirements

The codebase is 100% open source on GitHub. You need a local environment with Python, ffmpeg, and an ElevenLabs API key, with yt-dlp as an optional download helper. Browser Use is also preparing a cloud-hosted waitlist at video-use.com.

To use it locally, clone the repository, symlink it into your Claude Code skills directory, and run your agent inside your footage folder. Beyond simple cuts, the tool can spawn sub-agents to render programmatic overlay graphics through libraries like Remotion, Manim, HyperFrames, or PIL.

Why it matters

We think this is one of the smartest agent workflows built recently because it avoids the multi-modal brute-force trap. Treating video editing primarily as an audio transcript problem keeps token usage near zero while achieving word-boundary precision.

That said, industry observers have rightly pointed out its boundaries: it is fundamentally anchored to spoken dialogue. It is a massive time-saver for talking heads, tutorials, product updates, and interviews where rough cuts take hours of tedious trimming. But for highly visual montages or narrative shorts that rely on camera pacing rather than speech, you will still need a human timeline editor.