Slop TVNewsLatest
Explainers

Scarlet: an open-source AI video editor that edits from a transcript

Forward Future's MIT-licensed desktop app drives Claude Code or Codex over a cut list, renders only the changed seconds, and keeps your footage local.

Illustration: Scarlet: an open-source AI video editor that edits from a transcript
Illustration: SLOP TV News. Logos are trademarks of their owners.

Key takeaways

  • Scarlet is MIT-licensed and runs on macOS 13 or later, Windows x64, and an Ubuntu 24.04 preview, with no browser version.
  • Scarlet connects to Claude Code or Codex on a subscription the editor already pays for, so the app adds no separate AI bill.
  • Projects, renders and transcription stay on the machine, but the connected provider receives the request and context that can include transcript excerpts and video frames.
  • Packaged Scarlet builds report anonymous usage, edit review decisions and export outcomes to PostHog until you turn usage reports off.

Forward Future published Scarlet on November 13, an MIT-licensed desktop video editor that takes its editing instructions from a transcript.

Scarlet is an editor that drives an AI agent. It does not generate footage, and it does not turn a prompt into a video. You bring the files you already shot, an interview or a video podcast or a talking-head take, and the agent cuts them.

The workflow is conversational. Connect Claude Code or Codex in Settings, using a subscription you already pay for, then click a point on the timeline and type or hold V to speak. The README's example request is "Tighten the pause before 'three things.' Let me hear it both ways." Scarlet renders only the affected seconds, so you hear the change in context instead of waiting on a full re-render.

How Scarlet works

The cut list is the source of truth. Each project version in Scarlet's SQLite store describes a cut list of tracks, clips, captions and markers, according to the development guide. When the agent changes that list, Scarlet compares the versions and renders the affected ranges. Requested edits land on the working timeline, and suggestions wait for you to apply them.

Scarlet does not run the model itself. The plan document keeps the model, the harness and the editor as separate layers: the model makes editorial decisions, the harness (Claude Code, Codex, or a Gemini runner) runs the agent loop, and Scarlet supplies context, applies edits safely and gives the editor control. The repository ships 24 agent tools plus a shared system prompt over the Model Context Protocol, so an outside agent can inspect the project and propose changes.

Voice is a first-class input, and the plan is explicit about how it is handled: "Pass the editor's words as they said them. Don't summarize or rewrite a request before the model sees it."

Transcription runs locally. CrisperWhisper is the default, and until its separately licensed weights are installed Scarlet uses Whisper large-v3-turbo, which downloads on first use. The transcripts are word-timed, which is what lets the agent find where a speaker restarted a sentence and cut the abandoned first attempt.

Review is built around short playback. Open a cut and the transcript shows the words marked for removal. Press the backslash key to hear the removed footage, press it again to hear the join. Approve the cut, skip it, or ask for a revision. Every applied edit is versioned and can be undone on its own, and earlier versions stay in the project history.

Export gives you a finished file or a hand-off. Save the whole video or a selected range as MP4, or use Advanced export, which writes a folder with the timeline in .fcpxml and OpenTimelineIO format, a copy of the footage, the captions as an .srt file, and a Read me that lists what does not carry over. That folder opens in DaVinci Resolve, Final Cut Pro or Premiere Pro.

What it costs

Scarlet is MIT-licensed, so the app itself is free. It adds no separate AI bill, because the agent runs on the Claude Code or Codex subscription you already have.

That is not the same as free of AI cost. If you pay for that subscription, the editing runs against it. The local models are a one-time download: the bundled Whisper large-v3-turbo is 574 MB, and Forward Future's development guide says an English dialogue pass using bundled Whisper can pull about 1.1 GB across its steps. Later runs reuse the models.

What stays on your machine, and what does not

Scarlet stores projects on your computer and renders video there. Transcription is local. The README says so plainly.

The trade sits at the agent connection. Your connected AI provider receives your editing request and the project context the agent uses, which the README says can include transcript excerpts and video frames. An interview you have not published may be inside the transcript the agent reads to judge whether a line was repeated.

Packaged releases also report to PostHog: anonymous usage, edit review decisions, and task and export outcomes. You can turn that off with Send usage reports in Settings, under About, or by setting SCARLET_TELEMETRY_ENABLED=false before launch. Early Access versions through 0.1.3 use the environment setting.

How to try it

Scarlet is in early development. It builds on macOS 13 or later, on Apple Silicon or Intel; on Windows x64; and on Ubuntu 24.04 x86-64 as a preview. There is no browser version, and the plan document states that Scarlet runs as a desktop application because the embedded renderer needs the native Tauri bridge.

Forward Future publishes Early Access downloads for macOS and Windows and a separate Ubuntu preview. To run it from source, clone the repository, then run pnpm install --frozen-lockfile and pnpm build, with Node.js 22.x or 24.x, pnpm 12.16.1, Rust and the Tauri prerequisites installed. To look at the editor before touching your own footage, pnpm desktop:demo opens a sample interview with a simulated assistant. Import and export need FFmpeg and FFprobe, which macOS and Windows users install separately.

The changelog is the clearest sign of how early this is. Its entries begin on November 13, the day the repository went public, and the first few days of work include Ubuntu 24.04 support, a motion graphics tool, transitions and a Spanish transcription pass. Forward Future's own plan notes that the Codex and Gemini runners had not been exercised for real when it was written, and that Scarlet is single-user for now, with accounts and team features removed. The repository had more than 275 stars when SLOP TV read it on November 16, 2026.

Scarlet's bet is that cutting is the mechanical part and the model should carry it, while the editor keeps taste. Whether that holds on a messy two-hour interview is the question the project has not answered yet, and the changelog suggests the answer will arrive in public.

Download Early Access for macOS or Windows, or the Ubuntu 24.04 preview, from the Forward Future releases page, and connect Claude Code or Codex in Settings before your first edit.

Sources

  1. github.com - README: what Scarlet is, platforms, install, export, transcription, telemetry, MIT license
  2. github.com - design principles, cut list as source of truth, model/harness split, MCP tools, voice handling
  3. github.com - dated entries from November 13-16, 2026, including "Scarlet is open source"
  4. github.com - engine, FFmpeg and FFprobe, whisper-cli, transcription models, platform builds
  5. api.github.com - star count, license and creation date read November 16, 2026