Scarlet: Matthew Berman's AI video editor is now open source, and you edit by telling Claude Code or Codex what to cut
Scarlet is a desktop editor for macOS, Windows and Linux where you don't drag clips. You connect the Claude or ChatGPT (Codex) plan you already pay for, then say or type what to change, and review each cut on a timeline. It's aimed squarely at interviews, podcasts and talking-head takes, and your original footage is never touched.
Where it came from. On Nov 11 Matthew Berman (Forward Future) posted that he'd built an AI video editing app and wanted testers. That post passed 3,600 likes and 330K views. He followed up that he uses it to edit his own videos end to end, with about 10 minutes of his focused time per 20-minute video, mostly reviewing. That's his own claim about his own channel. The source repo went public under MIT on Nov 13, early access builds followed, and by this morning the stable line had reached v0.1.136.
How it works, per the README. Make a project and drop in video, separate audio and stills. In Settings, connect Claude Code or Codex using your existing subscription. Click a point on the timeline, then type or hold V and speak: 'Remove the false starts and repeated takes. Keep the pauses before each answer.' Scarlet renders only the affected section so you can hear the change in context. Edits you asked for go on the working timeline, while the agent's own suggestions wait for you to approve them. Press \ to hear the removed footage, then again to hear the join. Every version stays in the project history.
The workflows are skills. /dialogue-pass does a cut-only first edit that removes abandoned starts, redundant takes and distracting pauses. Others fit the edit to a target length, speed through installs and renders instead of cutting them, pull shorts out of a long video, frame tight on the part of a screen recording you're using, add punch-ins on a talking head, add captions, and build lower thirds and titles. There's also 'Matt's talking-head edit', his own house style built on top of the dialogue pass. You can write your own style the same way: put extends: dialogue-pass in a skill header and write only what's different. Several workflows are adapted from Alex's DaVinci editing skills.
Getting out. Export an MP4 of the whole video or a range. Advanced export writes a folder with the timeline as .fcpxml and .otio, a copy of the footage, captions as .srt and a read-me for DaVinci Resolve, Final Cut Pro and Premiere Pro, including what doesn't carry over. So you can use Scarlet for the rough dialogue cut and finish in your normal NLE.
Things to know first. Projects and renders stay on your computer, and transcription runs locally (Whisper large-v3-turbo by default, or CrisperWhisper if you install its separately licensed weights). Your AI provider does receive your request plus the project context the agent uses, which can include transcript excerpts and video frames, so check that against any client NDA. Packaged builds send anonymous usage reports, edit review decisions and export outcomes to PostHog by default. Turn this off under Settings → About, or set SCARLET_TELEMETRY_ENABLED=false. The macOS apps aren't notarized and the Windows installers are unsigned, so expect OS warnings. The project calls itself early development.
Why it matters. Most agent video tools so far build motion graphics from code. Scarlet goes after the slow, unglamorous part of real client work: cutting dialogue. And it makes you review every join instead of trusting a one-shot render. If you cut interviews or podcasts every week, this is the open-source tool from the past few days most worth a look.
Signals: the author's tester call took off on Nov 11, and tracked curator @RoundtableSpace posted it this morning (Nov 16, 3:15 AM MST). Smaller accounts are now writing it up in several languages. Those two signals are more than 72 hours apart, though, and the repo hasn't shown up on GitHub Trending, so it isn't HOT by our rule.
Skills & repos in our directory
- Scarlet · Repo, Agentic video production
More from the Lab
- ffmpeg-skill hits GitHub Trending again: what changed in the 3.x line of the local video editor for coding agents
- brief2mp4 checks its MP4 against your brief, plus three open-source agent video editors worth knowing: OpenChatCut, Pireel and FireRed-OpenStoryline
- Two Claude Code skills for editing real footage
- rea-video breaks a finished video down shot by shot, and Treg puts Seedance, Veo and Kling behind one pay-per-clip token