Claude Video is an open-source Claude Code skill and plugin that lets Claude inspect videos instead of guessing from a title or partial transcript. Users call the `/watch` command with a YouTube, TikTok, Vimeo, Loom, X, Instagram, local MP4, MOV, MKV, or WebM file and ask a question. The skill then collects captions when available, extracts frames with ffmpeg, optionally transcribes audio through Whisper, and hands the resulting images plus timestamped transcript back to Claude.
The value is practical: Claude can review screen recordings, bug reports, launch videos, tutorials, ads, podcasts, and course clips with visual context. The README shows use cases such as finding what happens at a specific timestamp, diagnosing a UI bug from a recording, summarizing a long video, turning a playlist into notes, or stripping hype from a product update. That makes it useful for builders who need Claude to reason over visual evidence, not just text.
Claude Video is careful about cost and context. It checks captions first, so many public videos can be handled without downloading full media or using paid transcription. Frame modes let users pick between transcript-only, fast keyframe extraction, balanced scene-aware frames, and an uncapped high-detail mode. The default keeps frames at a smaller width and caps long videos so image tokens do not swamp the model context. A deduplication pass drops near-identical frames from static slides or screen recordings.
Installation is easiest through the Claude Code plugin marketplace with `/plugin marketplace add bradautomates/claude-video` and `/plugin install watch@claude-video`. Other hosts can use `npx skills add bradautomates/claude-video -g`, and claude.ai users can download a `.skill` bundle. The tool itself is MIT licensed and free; users may need yt-dlp, ffmpeg, and optional Groq or OpenAI Whisper keys when a video has no captions.
For best results, ask focused questions. Long videos can be scanned, but the README notes that capped frame modes become sparse after roughly ten minutes. If you need detail, rerun the command with a start and end time around the section you care about. For private or local recordings, confirm that you have permission to process the file and understand whether Whisper fallback will send audio to Groq or OpenAI. Used carefully, Claude Video turns video into inspectable evidence for debugging, research, and content review.
This is an early-stage builder tool, so the right adoption pattern is cautious. Read the README, install in a disposable workspace first, and verify the commands before pointing it at valuable repositories or private media. The project is valuable because it exposes a concrete workflow that developers can inspect and adapt. It is not a replacement for human code review, source control hygiene, or security checks. Teams should keep logs, review generated artifacts, and document which agent or plugin performed each important action.