mcpField
W
watching-videos

watching-videos

github.com/oxbshwProductivity533565 installsUpdated 7w ago
Repository

About watching-videos

A video-analysis skill for agents handling online links and local files. It indexes frames, OCR, and transcripts so answers can draw on the video itself.

Key features

  1. 01Analyze online videos and local files
  2. 02Sample scenes and deduplicate frames
  3. 03Extract on-screen text with OCR
  4. 04Use captions or local Whisper transcription
  5. 05Store results in a searchable index

Use cases

  1. 01Review a specific section of a long video
  2. 02Process playlists, channels, or folders
  3. 03Answer questions using indexed video evidence

Install

Copy the command and run it in a terminal.

$npx skills add https://github.com/oxbshw/watch-skill --skill watching-videos

After installing, type/watching-videosin Claude Code to confirm it loaded.

Details

From the project README

Watching videos

You don't have a video input; this skill gives you one. Anything a user hands you — a YouTube link, a TikTok, a lecture, a meeting recording, a screen capture, an .mp4 on disk — goes through the same pipeline: frames (scene-aware, deduplicated), OCR, transcript (captions first, local Whisper offline fallback), all persisted into one index.

Before watching: check the index

watch-skill list

If the video was already analyzed — this session or any earlier one — do NOT watch it again. Ask it directly (see the asking-with-evidence skill):

watch-skill ask <video_id> "<question>"

One video

watch-skill watch "<url-or-path>" [--start T --end T] [--max-frames N] [--transcript-only]
  • Works on any yt-dlp-supported site (1800+), direct media URLs, HLS/DASH manifests (--duration 60 bounds live streams), and local files.
  • Video over ~10 minutes and the user cares about one part → use --start/--end for dense sampling of that window.
  • User only needs what was said → --transcript-only (fastest, often no video download at all).

The report prints Indexed: <video_id>, frames with t=MM:SS timestamps, OCR text, and the transcript. Read every frame path listed — in a single message, parallel Read calls — then answer from frames + OCR

  • transcript, citing timestamps.

Many videos (playlist, channel, folder)

watch-skill batch "<playlist-url-or-folder>" [--limit N]

Everything lands in the same index; one broken video never stops the rest. Afterwards a single watch-skill search "<phrase>" spans the whole batch.

First run on a machine

If any command fails with a dependency error, run watch-skill doctor once — it installs missing ffmpeg/yt-dlp itself. No API key is required for any of this; transcription is local by default and the video file never leaves the machine.

Comments

Markdown supported