bradautomates/claude-video

What it does

Watch is an Agent Skill that lets a coding agent "watch" a video. Given a URL or local file, it downloads or reads the media, extracts frames at selected points, builds a transcript from native captions or a speech backend, and returns the result to the agent with source-relative timestamps. It installs into Claude Code (desktop Code session, VS Code, or terminal), Codex, Cursor/Copilot and other Agent Skills hosts; it does not work in Claude Chat or Cowork (supported surfaces).

Two engines exist. auto uses Gemini when a GEMINI_API_KEY resolves, sending the whole video to Google and relaying a timestamped answer; this handles YouTube links without local tools and uploads local files that are deleted after the answer. Otherwise the local engine extracts frames and audio on the machine with FFmpeg, ffprobe and yt-dlp (engine settings).

Install and first run

The quickest path for Codex is to paste this in the message box (Codex setup):

Use $skill-installer to install the watch skill from:
https://github.com/bradautomates/claude-video/tree/main/skills/watch

In a terminal-based Claude Code session, enter (terminal setup):

/plugin marketplace add bradautomates/claude-video
/plugin install watch@claude-video

For Claude Desktop, add the repository as a marketplace under Customize → Plugins and then install Watch; adding the marketplace alone does not install the skill (desktop steps). Other Agent Skills hosts can use npx skills add bradautomates/claude-video -g --skill watch, which needs Node.js (other agents).

Then open a folder containing a short video and ask the agent in a new session:

Use the watch skill on example.mp4 in the folder I shared.
For setup, choose balanced detail and captions only (none).
Summarize what is visible, with timestamps.

Success is a timestamped visual summary; this path needs no transcription key (first video). To use the Gemini engine, paste a free Google AI Studio key when the wizard asks, or add GEMINI_API_KEY= in ~/.config/watch/.env.

Detail, captions and transcription

Detail levels trade frames for cost: transcript (no regular frames), efficient (keyframes, cap 50), balanced (scene changes, cap 100) and token-burner (uncapped). A --start/--end range or --timestamps on a transcript focuses the agent on specific moments. Native captions are always preferred; optional speech backends are WhisperX (local, managed Python 3.12 environment), Groq whisper-large-v3 or OpenAI whisper-1 (fallback options).

Constraints

• Requires Python 3.10+, current FFmpeg/ffprobe and the latest yt-dlp; YouTube also needs a JavaScript runtime/EJS setup. The skill ships instructions and scripts, not these programs (missing tools).
• Local WhisperX needs about 3 GB free disk and 8 GB RAM; Apple Silicon macOS is the verified configuration, while Intel macOS, Linux and Windows installs are untested. CUDA is untested and MPS is not promised.
• The Gemini engine sends the video (or its URL) to Google; the local engine never contacts Google. Cloud ASR sends extracted audio to the selected provider.
• Claude Chat and Cowork are unsupported; cloud sessions have their own setup and network permissions and may need reconfiguration per task.
• Licensed MIT (LICENSE).

HotRepo's take

The repository has no recorded GitHub daily rank, but its star snapshots show a fast climb: about 3,945 stars on 2026-07-06 growing to 10,804 by 2026-07-27, and 20,184 by 2026-11-08 with 2,145 forks. That trajectory is consistent with real interest in a practical agent capability, though no discussion threads are available to show how users are applying it. The README is unusually specific about engines, data egress and failure modes, which matters more here than raw growth: whether to use it depends on accepting either the local media toolchain or sending video to Google.