What it does

Claude Video, invoked as /watch, gives an agent video evidence. You point it at a URL or a local file and it hands back sampled frames and a transcript with timestamps, which the agent can use to summarise the video or answer questions about it. It offers two engines. With a Gemini API key, the whole video, picture and sound, is passed to Google's video model and the timestamped answer is relayed back. Without a key, or for private videos, the local engine extracts frames and a transcript on your machine using ffmpeg and yt-dlp. Native captions are used first, and WhisperX, Groq or OpenAI transcription can be added for videos that have none. It runs in Claude Code (desktop app, VS Code or a terminal), in Codex and in other Agent Skills hosts, but not in Claude Chat or Cowork. Python 3.10+, FFmpeg and yt-dlp are installed separately.

Capabilities

Turns a video URL or local file into timestamped frames plus a transcript for the agent
Reads native captions first, with WhisperX, Groq or OpenAI transcription as optional fallbacks
Choice of a Gemini engine or a local engine built on ffmpeg and yt-dlp
Detail levels, frame caps, start and end ranges and pinned timestamps
Installs as a Claude Code plugin, a Codex plugin or skill, or through the Skills CLI

When to use it

You want a coding agent to answer questions about a screen recording, tutorial or demo
You need a timestamped summary of a short video that covers both what is shown and what is said
You want captions and key frames pulled from a public video URL for analysis