The ChatCut agent plugin is the public npm package @chatcut/skill (currently v0.2.1), which installs a Claude Code skill named “chatcut” plus a command-line chatcut tool that creates AI-edited videos from a text prompt and a list of asset URLs or local file paths.

How do I install it?

Run npm install -g @chatcut/skill (pnpm add -g and yarn global add work the same way), then restart Claude Code. A global install is required — the skill invokes the bare chatcut binary on your PATH, so npx @chatcut/skill submit does not work. The install’s postinstall step copies SKILL.md into ~/.claude/skills/chatcut/SKILL.md so Claude Code picks up the skill automatically; that registration is best-effort and never fails the npm install, so if Claude Code doesn’t see the skill afterward, run chatcut install to register it manually.

What do I need to run it?

The chatcut CLI requires Node.js 18 or newer (engines.node >=18 in package.json).

What can the CLI do?

The chatcut command exposes nine subcommands: install (with a --link option), update, login, logout, pick, submit, status <jobId>, watch <jobId>, and help. The core one is submit, which takes a required --prompt <text>, one or more repeatable --asset <source:type> flags (type is audio, gif, image, or video), and optional --name <string>, --resolution <480p|720p|1080p>, --format <audio|video>, --codec <h264|mp3|prores|vp8>, --max-turns <n>, and --model <name> flags. If you don’t already have asset paths on hand, chatcut pick opens a native file-selection dialog and prints path:type lines you can feed straight into submit --asset.

How does `chatcut submit` create a video?

Point submit at a prompt and your media, and it runs the whole pipeline for you: you describe the edit in plain language and supply assets as either public URLs or absolute local file paths — both are accepted identically — the CLI transcodes them locally and uploads them, ChatCut’s agent assembles the timeline server-side, the video renders in the cloud, and you get back a signed download URL. If you don’t set --resolution or --format, the CLI automatically appends an export instruction to your prompt that defaults to 1080p video. Submitting a job is a paid operation, so if the agent altered your wording before submitting, it’s required to stop and show you the exact --prompt text for approval first.

How long does it take, and where do I see progress?

chatcut submit blocks until the job is completely done — transcode, upload, the agent’s edit loop, and the cloud render all run before it returns — streaming human-readable progress to stderr the whole time. Behind the scenes it polls job status every 3 seconds and gives up with a timeout error if the job hasn’t finished within 40 minutes.

Where do the results end up?

Every chatcut submit run creates a new ChatCut project — name it with --name, or accept the default — and prints back that project’s projectId together with the job’s jobId and a signed outputUrl for the rendered file, as a single JSON line on stdout once the run finishes.

Do I need to run `chatcut login` first?

No — if submit doesn’t find a saved API key, it opens your browser to sign in automatically and continues with the job once you’re done, so a separate chatcut login step isn’t required before your first submit. That sign-in opens a browser window pointed at ChatCut’s sign-in page; after you approve it, the CLI exchanges the result for a long-lived API key behind the scenes, and the whole flow times out after 5 minutes if left incomplete.

Is this the same as the “Codex plugin”?

Not as shipped today — @chatcut/skill is exclusively a Claude Code skill: the CLI registers under ~/.claude/skills/chatcut, and nothing in the package integrates with OpenAI Codex.

Install the plugin globally with npm install -g @chatcut/skill — pnpm add -g and yarn global add work the same way — then restart Claude Code. It requires Node.js 18 or newer. Installing it also registers SKILL.md at ~/.claude/skills/chatcut/ so Claude Code picks up the skill automatically; if that registration doesn’t happen (for example with --ignore-scripts), run chatcut install to do it manually. A global install is required — npx @chatcut/skill won’t work from inside Claude Code.

You don’t need a separate sign-in step: the first chatcut submit opens your browser to a PKCE-based authorization exchange, and the flow times out after 5 minutes if you don’t complete it. To sign in ahead of time instead, run chatcut login.

What can the ChatCut agent plugin do, and what are its limits?

The ChatCut agent plugin — the chatcut CLI, a Claude Code skill — creates a project, uploads your media, runs a natural-language AI edit job, and returns a signed download URL in one command. chatcut submit takes a prompt plus assets (URLs or local file paths); it transcodes video locally only when the codec isn’t browser-compatible, frame rate exceeds 30fps, or bitrate exceeds 8Mbps, clamping output to 1920px on the longest side — otherwise it uploads the file untouched. It then blocks until ChatCut’s server-side agent finishes editing and cloud rendering, streaming progress to stderr and a final jobId/projectId/outputUrl JSON line to stdout.

Limits: submit polls every 3 seconds, timing out after 40 minutes. chatcut status <jobId> checks a job once; chatcut watch <jobId> polls up to 20 minutes. The CLI checks for a newer version every 24 hours (cached) and only nudges — never blocks a command.

Agent Plugin