video-use is an MIT-licensed, self-hostable tool that lets Claude Code or another coding agent cut filler words, color grade, and subtitle raw footage, the same job a Descript subscription or a freelance editor is usually hired to do.
A marketing team that ships a weekly talking-head video, a founder demo, a customer testimonial cut, a LinkedIn clip, has historically paid for the editing pass one of two ways: a Descript seat running $24 to $65 a month per editor depending on the plan, or a freelance video editor billing by the project. Browser Use just open sourced video-use, an MIT-licensed tool that hands that same job, cutting filler words, color grading, burning in subtitles, to whatever coding agent you already have open. Drop raw footage in a folder, tell Claude Code to edit it, get a finished mp4 back.
The repo is sitting at 22.8k stars and 2.8k forks, built by the same team behind browser-use, the agent framework that lets an LLM drive a web browser through a structured page representation instead of a screenshot. video-use applies the identical trick to video. The model never watches the footage. Every source clip gets one pass through ElevenLabs Scribe for a word-level transcript with speaker labels and audio events like laughter or applause, and the whole shoot packs into a single file small enough, around 12 kilobytes, for an LLM to read start to finish. That transcript is the primary interface. A visual filmstrip only gets pulled up when the agent needs to sanity-check an ambiguous cut.
From there the agent proposes an edit strategy, waits for a yes, then works through a pipeline the README lays out as transcribe, pack, reason, generate an edit list, render, and self-evaluate. It cuts filler words and dead air between takes, applies a color grade (there's a warm cinematic preset, a neutral punch preset, or your own custom ffmpeg chain), fades audio 30 milliseconds on either side of every cut so splices don't pop, and burns in subtitles in short two-word bursts by default, restyleable if you want something quieter. It can also generate animation overlays through Manim, Remotion, or PIL, each spawned in its own sub-agent. Before it shows you anything, it re-renders and checks its own cut boundaries for visual jumps or hidden captions, retrying up to three times if something looks off. The license itself is a clean, unmodified MIT grant, permissive with no strings attached.
This is a fit for whoever already has an agent-driven workflow running: a marketing ops person comfortable pasting commands into Claude Code, a solo founder who edits their own demo videos, an in-house content team at a B2B company producing recurring talking-head or tutorial footage. It is not a fit for someone whose only relationship with video editing software is dragging a clip onto a timeline. You need a terminal, ffmpeg installed, and a coding agent already set up before this does anything for you.
Self-hosting here is not free, it's just billed differently. You still need your own ElevenLabs API key for the transcription step, an ongoing usage bill stacked on top of whatever you already pay for agent access, and the README doesn't publish a rate card for that piece. There's no tagged release on the repository yet, worth noting for a tool that touches your raw footage and an API key before you point it at anything you can't afford to re-shoot. It also has no answer for the collaboration layer that comes bundled with a Descript seat: shared brand templates, a review link a client can comment on, translation and dubbing into dozens of languages. It edits alone, reasonably well, but it doesn't manage a team. And worth flagging plainly: the README's own quickstart is written as a block of instructions addressed directly to a coding agent, telling it to install dependencies, register the tool, and ask you for an API key before touching anything. That's a deliberate installation pattern for a tool meant to be set up by an agent, not a request aimed at whoever wrote this article, but it's exactly the kind of paste-into-an-agent text worth reading closely before you hand it to something with shell access.
The notable thing about video-use isn't that a language model can now cut a sizzle reel without watching a single frame of it, even though it can. It's that the fastest path to installing the tool is no longer a package manager command, it's a paragraph of English handed to whatever agent is already open on your machine.