Who video-use is for#
Developers editing tutorial and demo videos
video-use fits directly into a development workflow: record a terminal session or screen capture, point the agent at the folder, and describe the edit without leaving the terminal. Filler removal and subtitle burning happen automatically.
Skip if:
Skip if you need frame-accurate cuts for motion graphics or narrative editing where visual timing is critical. The text-based workflow is strongest for speech-driven content.
Solo creators producing talking-head content
Handles the repetitive editing tasks that consume the most time on talking-head content: removing filler words, cutting dead air between takes, grading color consistently across segments, and burning captions in a consistent style.
Skip if:
Skip if your edit requires multi-camera switching with timecoded syncing. video-use works best with single-camera or sequentially-shot footage organized in one folder.
Indie developers shipping a product launch video
Drop multiple takes in a folder, ask the agent to edit them into a launch video, and get a cut with color grading and subtitles ready for upload. The agent proposes a strategy and waits for approval before touching the footage.
Skip if:
Skip if your launch video requires motion design, b-roll from multiple sources, or precise sound design beyond automatic audio fades at cut boundaries.
Developers running always-on editing from a VPS
video-use can run through Browser Use Box for an always-on agent endpoint accessible from Telegram or any messaging client, suited to teams that produce regular video content and want automated editing as a persistent background service.
Skip if:
Skip if your infrastructure budget is limited. The always-on path requires a VPS with sufficient RAM for ffmpeg rendering and an ongoing ElevenLabs API budget.
The problem it solves#
Video editing for developers and solo creators involves a painful mismatch: the intent is simple (cut dead air, grade warm, add captions) but executing it requires navigating timeline UIs, plugin menus, and export dialogs that interrupt the coding workflow. Commercial tools like Descript reduce some of this friction, but they are cloud-hosted subscription services. Footage leaves your machine, pricing is per-minute or per-seat, and the edit workflow does not compose with the shell tools or agent sessions you already use.
The deeper problem is how AI models process video. Naive approaches feed raw frames into a vision model: 30,000 frames at 1,500 tokens each produces 45 million tokens of noise with no signal about when words were spoken or where silence gaps fall. Without word-level timestamps and speech-boundary data, no model can make precision cuts that sound natural. This is the gap that existing commercial tools paper over with GUIs rather than addressing at the pipeline level.
How it solves it#
Filler-word removal and dead-space cutting
Transcribes all takes with ElevenLabs Scribe to get word-level timestamps. The agent identifies umm, uh, false starts, and silence gaps, then cuts them at speech boundaries so the edit sounds intentional rather than clipped.
Auto color grading via custom ffmpeg chains
Applies a color grade to every segment through ffmpeg. Default presets include warm cinematic and neutral punch; any custom ffmpeg chain can be substituted. Color grading runs per-segment so retakes with different lighting match.
Subtitle burning with style control
Burns subtitles directly into the video in 2-word UPPERCASE chunks by default, matching the pacing style common in tutorial and social media content. Font, size, and chunking rules are fully customizable.
30ms audio fades at every cut boundary
Adds 30ms audio fades at every cut point automatically, preventing the audio pops that result from hard cuts between takes recorded at different microphone positions or room tones.
Self-evaluating render loop
Runs a visual check on the rendered output at every cut boundary to detect visual jumps, audio pops, and hidden subtitles. The agent re-renders up to three times if it finds issues; you see the preview only after the output passes.
Session memory via project.md
Persists the edit strategy, file inventory, and cut decisions in a project.md file next to the footage. Future sessions pick up where the last left off without re-transcribing or re-proposing a strategy.
Strengths and trade-offs#
Strengths
- Text-based video reasoning, no frame dumpingInstead of feeding frames to the vision model, which would cost approximately 45 million tokens on a 30-minute take, video-use represents footage as a 12KB transcript markdown file and calls visual snapshots only at decision points. This makes full-length editing feasible inside a standard context window.
- MIT license with full local executionThe entire pipeline runs on your own machine. No footage leaves your infrastructure except for the ElevenLabs transcription call. MIT license means no restriction on commercial use, modification, or redistribution of the pipeline.
- Compatible with multiple coding agentsDesigned as a Claude Code skill but also compatible with Codex, Hermes, Openclaw, and any agent with shell access. The universal setup prompt handles installation and skill registration automatically under whichever agent runtime you use.
- Content-type agnostic editing without presetsWorks for talking heads, montages, tutorials, travel footage, and interviews without content-type presets. The agent inspects the footage, proposes a strategy, and waits for your approval before making any cuts.
Trade-offs
- -ElevenLabs API required for all transcriptionEvery transcription call goes to ElevenLabs Scribe, which is not free at scale. Heavy use with long takes will accumulate API costs. There is no bundled local transcription fallback, so an active ElevenLabs account is a hard prerequisite for all editing workflows.
- -No GUI or visual timelineThe entire workflow is conversational: describe edits in text, the agent executes them, and you approve the output file. Editors who need to scrub a visual timeline or do frame-accurate cuts by eye will find the text-only interface limiting.
- -ffmpeg and animation dependencies must be installed separatelyRendering is handled by ffmpeg, which must be installed separately. Animation overlays using Remotion, Manim, or PIL each require their own dependencies on the host machine, adding setup friction for complex overlay workflows.
video-use vs alternatives#
video-use vs Descript
Descript and video-use both remove filler words and add captions to talking-head footage, but they take opposite approaches to the workflow. Descript is a cloud-hosted document editor where you edit a transcript and the video follows. video-use is a local agent skill where you describe the edit in plain language and ffmpeg executes it.
| Feature | video-use | Descript |
|---|---|---|
| License | MIT | Proprietary |
| Self-hosting | Yes, fully local | No |
| Filler removal | Yes, via ElevenLabs timestamps | Yes |
| Subtitle burning | Yes, customizable | Yes |
| GUI | None | Full visual editor |
| Pricing | Free (ElevenLabs API costs apply) | $12-24/month subscription |
video-use is the better choice when you want local data control, no subscription, and an edit pipeline that composes with shell scripts and agent workflows. Descript is the better choice when you need a visual timeline, team collaboration features, or a no-code interface for non-technical editors.
video-use vs RunwayML
RunwayML focuses on AI-generated video content: text-to-video synthesis, background removal, motion tracking, and generative inpainting. video-use focuses on editing raw footage you have already recorded. The overlap is limited; both use AI but for different stages of video production.
| Feature | video-use | RunwayML |
|---|---|---|
| License | MIT | Proprietary |
| Self-hosting | Yes, fully local | No |
| AI video generation | No | Yes |
| Editing existing footage | Yes | Limited |
| Filler removal | Yes | No |
| Pricing | Free (API costs apply) | Credit-based, from $15/month |
video-use is the better choice when your footage exists and you need to edit it through agent commands with local control. RunwayML is the better choice when you need generative video effects, text-to-video synthesis, or AI visual tools beyond timeline editing.
HeyGen and Google Gemini API (Veo)
HeyGen specializes in AI avatar generation and lip-sync dubbing for marketing and localization workflows. Google Gemini API (Veo) is a text-to-video generation service. Neither directly competes with video-use's core workflow of editing raw footage through a coding agent; they serve generative use cases where no source footage exists yet.
Install and self-host#
Setup requires cloning the repo, linking it as a coding agent skill, and installing ffmpeg.
```bash
git clone https://github.com/browser-use/video-use ~/Developer/video-use
ln -sfn ~/Developer/video-use ~/.claude/skills/video-use
cd ~/Developer/video-use
uv sync
brew install ffmpeg
cp .env.example .env
```What it's built on#
- Languages
- Python
FAQ#
Does video-use work with agents other than Claude Code?
Yes. video-use works with Claude Code, Codex, Hermes, Openclaw, and any agent with shell access. The setup prompt is universal: paste it into your agent and it handles cloning, dependency installation, and skill registration automatically. The helpers and ffmpeg pipeline are shell-agnostic.
What does video-use use the ElevenLabs API for?
ElevenLabs Scribe provides word-level timestamps, speaker diarization, and audio event labels from each raw take. These timestamps are what let the agent cut precisely at speech boundaries rather than guessing from waveform amplitude alone. The agent prompts you for your ElevenLabs API key during setup.
Can video-use run without any cloud dependencies?
Almost entirely. The ffmpeg pipeline, color grading, subtitle burning, and rendering all run locally. The one external call is the ElevenLabs transcription API. The current version ships no local transcription adapter, so an ElevenLabs account is required. All rendered output stays on your own machine.
What video content types does video-use support?
The README lists talking heads, montages, tutorials, travel footage, and interviews as supported content types. The agent makes no assumptions about content: it inspects the footage, proposes an edit strategy, and waits for your approval before cutting. Any footage that can be transcribed to word-level timestamps works.
Is video-use free to use commercially?
Yes. video-use is MIT licensed, so there are no restrictions on commercial use, modification, or redistribution of the pipeline code. ElevenLabs API costs apply separately and depend on the volume and length of footage you transcribe.
Similar open-source tools#
MoneyPrinterTurbo
AI-powered short video generation from topic to publish
VoiceStudio
Voice cloning, dubbing, and transcription that runs fully offline
stt
Local speech-to-text with multi-format output and GPU acceleration
YouDub-webui
Self-hosted AI video dubbing: YouTube to Chinese, Bilibili to English
pyvideotrans
Open source video translation with AI dubbing and voice cloning
whishper
Local speech-to-text transcription and subtitling with a web UI