Scarlet goes open source with Claude Code and Codex editing interview footage

Forward Future published the source code for Scarlet, a desktop video editor, on Nov. 13. The app connects to Claude Code or Codex and lets an editor describe a cut by voice or text. The agent can then change the timeline itself. An editor working on an interview can ask it to remove repeated answers, then open a changed section and listen to how the lines on either side run together.

An early-access build arrived on Nov. 13. By Nov. 15, installers for Scarlet 0.1.100 were available for Apple Silicon and Intel Macs and for Windows x64. A separate preview package arrived on Nov. 14 for Ubuntu 24.04 on x86-64. The project uses the MIT license and connects to the editor’s existing Claude Code or Codex service.

Reviewing a cut on the timeline

Scarlet can bring video, separate audio and still images into a project. An editor can attach a request to a point on the timeline or ask for a pass through the whole conversation. Consider an interview in which a speaker starts the same answer several times. The request can identify the part to clean up and tell the agent which delivery to preserve.

The developers’ dialogue guide treats pauses and repetitions according to what they do in the recording. A false start can go when the next take completes the same thought. A pause that gives the other speaker time to respond may belong in the cut. The guide calls for staging a proposal for human review when a change could lose a reaction or alter what the speaker meant.

A reviewer can listen to the marked words and hear the edit with those words removed. If a proposed cut makes a reply sound abrupt, the editor can ask for a revision, skip the proposal or approve it. Applied edits can also be revised or undone, and earlier versions stay in the project history. This brings the transcript and the sound of the join into the same review.

Where the footage and editing requests go

Scarlet stores project data and renders video on the computer. Transcription also runs on the device. The transcription guide describes output that pairs words with offsets within the audio, which lets the editor find a restarted sentence on the timeline. CrisperWhisper is the default option. Scarlet falls back to Whisper large-v3-turbo until the separately licensed CrisperWhisper weights are installed.

The AI service connected for editing receives the request and the project context the agent uses. Forward Future says that context can include excerpts from the transcript and video frames. If an agent looks for a duplicated answer in an unpublished interview, for instance, the transcript it reads may contain material that has yet to be released.

The installers also send usage reports as a separate stream. Scarlet’s documentation lists items such as the number of projects and the outcome of a review or export. The app’s About settings offer a switch for those reports. That switch covers telemetry, while the connected AI service still handles the requests an editor submits to it.

Editors can export the full video or a selected range as an MP4. Scarlet also writes an OpenTimelineIO file for moving the timeline into DaVinci Resolve. The first option produces a video to share. The second lets editors carry their cut into another application for further work. Scarlet requires FFmpeg and FFprobe to import and export footage.

The release notes say the early-access and stable builds use separate project libraries. The developers direct people with existing early-access projects to open them in that build. Apple has not notarized the Mac app, and the Windows installers are unsigned, so the operating system may show a publisher warning. The Ubuntu version is an initial preview. The developers reported testing import, transcription, timeline playback and MP4 export under X11. Microphone capture on physical hardware and native Wayland operation were still listed for further testing.