What it needs

Cutwright runs on your own computer, and your footage never leaves it. The director, Claude on your account, reads the index and the small frames it asks for; nothing else is sent anywhere.

Computers

Cutwright is in development and is not offered for download yet. Its builds are made and tested for:

System	Needs
Linux, x86-64	glibc 2.35 or later (Ubuntu 22.04 and newer), the Vulkan loader, VA-API for hardware video.
macOS, Apple silicon	macOS 14 or later.
macOS, Intel	macOS 13.5 or later. Speech runs on the processor.
Windows, x86-64	Windows 10 or later.

Hearing the words runs on the GPU where there is one (Vulkan, or Metal on Apple silicon) and on the processor otherwise. Video is encoded by the system's own encoder: VA-API, VideoToolbox or Media Foundation.

The director

The director is Claude. Use it in one of two ways:

Claude Code, on your Claude plan. cutwright claude connect sets it up for a project; then run claude in the project's folder. What it costs is what your plan covers.
The Studio's director pane, on your Anthropic API key. Each turn's cost is shown as you go, under a cap you set.

The director sees the index (words, speakers, shots, what moves) and small frames it asks for; it never receives your footage files.

Models

Indexing runs these on your computer. Each is fetched once into Cutwright's cache, checked against its published hash, and used under its own licence.

Hears or sees	Model	Size
Words	Parakeet TDT 0.6B v3 (CC BY 4.0)	1.3 GB, or 0.7 GB for machines short of memory
Word timing	wav2vec2 (Apache 2.0)	0.4 GB
Laughter, applause, music	YAMNet (Apache 2.0)	15 MB
Voices on a shared mic	WeSpeaker (CC BY 4.0)	27 MB
Faces and eyes	YuNet (MIT), face landmarks (Apache 2.0)	5 MB
What a frame shows, for finding B-roll	Qwen3-VL 4B (Apache 2.0)	3 GB, only when you turn it on (`analysis.captionFrames`)

The Studio in a browser

Chrome or Edge on a desktop. Footage dropped into the page is indexed in the page, with the speech model fetched once (about 2.5 GB) and run on WebGPU. Your files are copied into the browser's own storage for that site and never uploaded.

Is it all there?

cutwright doctor

Checks that Cutwright is whole: its resources, its own LGPL FFmpeg and ffprobe, the speech model, and where each was found. cutwright version prints the version and cutwright help lists every command.

Where things are kept

A project is a folder you own, with its own git history of every approved change:

my-show/
  cutwright.json        the project: its name, rate, size and sources
  sources/              one file per original: where it is, its hash, how it is synced
  edits/                your cuts, changed only by approved proposals
  packages/             graphics packages
  artwork/              your logos and cards, and their licences
  inserts/              Reelwright inserts
  .cutwright/           what Cutwright made and can make again: proxies, the index,
                        timelines, renders, exports

Your originals stay where they are and are never changed. What Cutwright makes is named by its content, with a note of how it was made, and can always be made again.

For cartoons

Inserts from Reelwright need Reelwright itself. cutwright reelwright fetch downloads the release Cutwright is built against, checks its hash, and keeps it in Cutwright's cache; it is never bundled. See Cartoons from Reelwright.