Claude Video Setup: Let Claude Watch Any Video

Claude Video is a free, open source skill that gives Claude Code, Cursor, Codex and other coding agents the one thing they cannot do out of the box: actually watch a video. You paste a YouTube link or a local file, ask a question, and it fetches captions, extracts frames and hands both to Claude so the model answers from what is really on screen instead of guessing from a title. It is one of the fastest-climbing repositories on GitHub right now, adding roughly 700 stars in a day on top of more than 6,800 total, and setup takes about five minutes.

• Claude Video installs as a /watch skill, not a separate app: one command in Claude Code, or one npx command for 50+ other agent hosts.
• It pulls free captions first and downloads video only when it has to, so most public clips cost nothing beyond your normal token spend.
• A Whisper API key is optional, needed only when a video has no caption track (local files, some TikToks); captions cover the majority of public videos.
• It is MIT licensed and built on yt-dlp and ffmpeg, with detail modes that trade frame count against token cost.

What is Claude Video and why is it trending?

Claude Video, published by Brad Bonanno under the /watch name, closes a real gap in how coding agents work. An agent can read a webpage, run a script or browse a repo, but hand it a YouTube URL and it either guesses from the title or pulls a transcript that misses most of what is on screen. /watch fixes that end to end: it checks for captions, downloads only what the run needs, uses ffmpeg to extract scene-aware frames, pulls a timestamped transcript, and then Claude reads each frame as an image. By the time it answers it has seen the frames and heard the audio. The repository is MIT licensed, sits above 6,800 stars, and shipped v0.2.0 on July 1, 2026. It is trending because agent users keep hitting the same wall the moment they try to summarize a talk, break down a competitor's video, or debug from a screen recording, and this is the first tool that makes that a one-line command.

RelatedPhotoGIMP Setup: Make GIMP Feel Like Photoshop

How do you install Claude Video?

The recommended path on Claude Code is the plugin marketplace, which auto-updates through the marketplace:

# Claude Code (auto-updates via marketplace)
/plugin marketplace add bradautomates/claude-video
/plugin install watch@claude-video

For Codex, Cursor, Copilot, Gemini CLI and 50+ other Agent Skills hosts, one command installs the skill into every agent it detects:

# Codex, Cursor, Copilot, Gemini CLI, and 50+ more
npx skills add bradautomates/claude-video -g

The -g flag installs globally for your user; drop it to scope the skill to the current project only. On claude.ai you download the watch.skill bundle from the latest GitHub release, add it under Settings, Capabilities, Skills, and enable "Code execution and file creation" first, since the skill shells out to ffmpeg and yt-dlp. Prefer to run from a clone? Symlink the self-contained skill folder into your host's skills directory so it stays in sync with your working tree:

# manual / developer install
git clone https://github.com/bradautomates/claude-video.git
ln -s "$(pwd)/claude-video/skills/watch" ~/.claude/skills/watch

What do you need on the first run?

On the first /watch call the skill runs a preflight check and, if anything is missing, walks you through fixing it. On macOS it auto-installs the two dependencies with Homebrew; on Linux and Windows it prints the exact package-manager commands instead of guessing:

# macOS: auto-run on first launch
brew install ffmpeg yt-dlp

If a video has no captions, the skill scaffolds a config file at ~/.config/watch/.env (mode 0600) with commented placeholders for a Groq or OpenAI key. Groq's whisper-large-v3 is preferred because it is cheaper and faster; OpenAI's whisper-1 is the alternative. You can also skip transcription entirely with --no-whisper for a frames-only pass. After setup the preflight is a sub-100ms lookup, so it does not slow later runs down.

How do you watch your first video?

Usage is a URL or a path followed by a plain-language question. Claude fetches what it needs and answers from the frames and transcript:

/watch https://youtu.be/dQw4w9WgXcQ what happens at the 30 second mark?
/watch https://www.tiktok.com/@user/video/123 summarize this
/watch ~/Movies/screen-recording.mp4 when does the UI break?

To keep token cost down on a long video, name the moment you care about and pass --start and --end. Focused mode gets a denser per-second frame budget than a sparse pass over the whole clip:

RelatedOfficeCLI Setup: Word, Excel and PowerPoint for AI Agents

/watch https://youtu.be/abc --start 2:15 --end 2:45

Two dials matter most. --detail ranges from transcript (captions only, zero frames) through efficient (fast keyframes, 50-frame cap) and the default balanced (scene-aware, 100-frame cap) to token-burner (uncapped). And --resolution 1024 bumps frame width when Claude needs to read on-screen text like slides, terminals or code.

How does /watch compare to the alternatives?

The point is not that transcripts are useless; for a talking-head podcast a pasted transcript is fine. /watch wins the moment the picture carries information the words do not: a UI bug in a screen recording, the on-screen code in a tutorial, the hook framing in an ad. It is the only option here that both sees and hears, works on local files, and lives inside the agent as a single command rather than a manual copy-paste chore.

What are the gotchas before you rely on it?

Four things worth knowing. First, long-video accuracy depends on the detail mode: on the capped modes coverage thins out past roughly ten minutes, so the script prints a "sparse scan" warning and you are better re-running focused with --start and --end, or using token-burner to keep every scene-change frame. Second, frames dominate token cost because every frame is an image; the tool's auto-fps logic and a default dedup pass, which drops near-identical held-slide frames, exist precisely to stop a thirty-minute clip from eating your context budget. Third, captions are free but the Whisper fallback is a paid API call, so caption-less local files and some TikToks cost a little unless you pass --no-whisper. Fourth, on claude.ai you must enable code execution first or the skill cannot shell out to its tools. None of these are dealbreakers; they are the knobs that separate a cheap, accurate run from an expensive, sparse one.

Our take

Claude Video is a small idea executed with the right taste. It does not try to be a video platform; it is a thin, honest bridge that turns a link into frames and a transcript an agent can reason over, and it defaults to the cheap path (free captions, deduped frames, capped budgets) so you do not accidentally spend a fortune summarizing a webinar. The install story is the best part: because it ships as an Agent Skill, one command wires it into Claude Code, Cursor, Codex and dozens of other hosts without a bespoke integration each time. The caveats are all about cost control on long videos, and the tool hands you the dials to manage them. If your work involves pulling signal out of videos, competitor teardowns, bug repros, course notes, this is worth the five minutes it takes to install today.

Original analysis by GenZTech. Tool documentation: bradautomates/claude-video on GitHub.