CookiesForAI

watch

Watch is an open-source agent skill that turns a video link or file into frames and a transcript, or hands it to Google's Gemini, so your agent can answer questions about it.

Agent skill by bradautomates · Updated

Install Watch

Inside Claude Code, add the claude-video marketplace

Terminal
/plugin marketplace add bradautomates/claude-video

Then install the watch plugin

Terminal
/plugin install watch@claude-video

Or, for other agents, run this in your computer's terminal (needs Node.js)

Terminal
npx skills add bradautomates/claude-video -g --skill watch

The publisher says it does not work in the Claude chat app or in Cowork.

Or paste this into your coding agent: Install the agent skill watch from github.com/bradautomates/claude-video. A skill can include scripts that run on your computer, so read its source first.

What Watch does

Watch is an agent skill that lets your agent answer questions about a video. Give it a video link or a file on your computer plus a question, and the agent answers from what is in the video. Links can come from YouTube or any other site yt-dlp can download, and the frames and transcript it works from carry timestamps.

Watch has two engines. The local engine downloads the video with yt-dlp, pulls frames with ffmpeg and takes the transcript from the video's own captions, falling back to WhisperX on your computer or a cloud Whisper service (Groq or OpenAI). With a free Google Gemini API key, the Gemini engine lets Google's video model watch the whole video, sound included: YouTube links are sent to Google, and other videos are uploaded to Google and deleted after the answer.

On first use of Watch, a Python setup script checks for the tools it needs and asks which engine you want. The local engine needs yt-dlp and ffmpeg installed. Settings and API keys stay in ~/.config/watch/.env on your machine, and the optional local WhisperX transcriber installs into its own environment under ~/.cache/watch.

When to use Watch

  • You want the key points of a long YouTube talk or tutorial without watching all of it.
  • You need to know what happens at a certain moment in a screen recording or bug report video.
  • You want your agent to learn from a product demo or course video and answer questions about it.

When to pick something else

  • You use Claude.ai chat, Cowork or a browser-only app: the publisher says Watch does not work there.
  • Your videos must not leave your machine: use the local engine, since the Gemini engine sends YouTube links to Google and uploads other videos.

Which agents Watch works in

bradautomates documents Watch for Claude Code (terminal, VS Code and Claude Desktop Code sessions), Codex (app, CLI and IDE extension), Cursor, GitHub Copilot, other local agents through the skills CLI. The open skills CLI also installs it into 78 agents, including Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, OpenCode (we listed it with the CLI on October 1, 2026). See where each agent looks for skills.

The SKILL.md lists Claude Code's allowed tools (Bash, Read, AskUserQuestion) but says they are examples: any host with a shell, an image viewer and a way to ask you questions works. The /watch command is Claude Code's slash command form; in Codex you ask for watch by name.

Watch license

Watch is published under MIT.

  • Algorithmic Art: Algorithmic Art is Anthropic's open-source agent skill for generative art: the agent writes a short art philosophy, then builds an interactive p5.js piece with seeds and parameter sliders. (179,326 repository stars on October 2, 2026)
  • Canvas Design: Canvas Design is Anthropic's open-source agent skill for static visual work: the agent writes a design philosophy, then makes a poster or art piece as a finished PNG or PDF file. (179,326 repository stars on October 2, 2026)
  • HyperFrames: hyperframes is HeyGen's entry-point agent skill for HyperFrames, its open-source framework that renders HTML, CSS and animations into MP4 video; it captures your brief and routes to the right video workflow. (55,348 repository stars on October 2, 2026)
  • Remotion best practices: remotion-best-practices is Remotion's own agent skill and the entry point to its other video skills: it routes an agent to the right guide for creating, previewing, captioning and rendering Remotion videos in React. (4,803 repository stars on October 2, 2026)

Lists that include watch