Paste a YouTube link into your agent. Get a report you read beside the video.
video-lens is a skill for Claude Code, Codex, Cursor and other agents that read SKILL.md. It writes a summary, the takeaway, numbered key points and a timestamped outline, and opens them in your browser with the video playing beside the text. Click any line to jump there.
video-lens.org has the rest: an example report, every feature, privacy, and all the ways to install.
npx skills add kar2phi/video-lensPick all three skills when asked. On the first video, video-lens checks what is missing and asks
before installing anything. Its Python packages go into its own folder,
~/.local/share/video-lens/venv, not your system Python.
Manual install
The same setup by hand, with the script the skill runs:
python3 ~/.agents/skills/video-lens/scripts/setup.py transcript-api yt-dlp
# Optional, Apple Silicon: transcribe videos that have no captions
python3 ~/.agents/skills/video-lens/scripts/setup.py mlx-whisper && brew install ffmpegAdjust the path to where your agent keeps skills (~/.claude/skills/… for Claude Code). Add
--upgrade when YouTube breaks an older release. If you'd rather use your own environment,
install requirements.txt into the python3 your agent runs; the skill uses those packages when
its own folder doesn't have them.
Works in any agent that reads SKILL.md. To install from a clone or as a Raycast command, see docs/guide.md.
Once a day at most, video-lens asks GitHub whether a newer version is out and tells you; it sends nothing about you and never updates without asking (what it checks, and how to turn it off).
Requirements
| Needed for | What |
|---|---|
| Running the skill | An agent that reads SKILL.md, Python 3, macOS or Linux (Windows is best effort) |
| Fetching captions | youtube-transcript-api |
| Metadata, chapters, audio download | yt-dlp 2026.8.19 or newer (older builds cannot download audio) |
| Optional: videos without captions | mlx-whisper and ffmpeg, Apple Silicon only |
| Optional: yt-dlp edge cases | Deno, only if yt-dlp fails on a video |
| Optional: install from a clone | Task (brew install go-task) |
| Optional: a global hotkey | Raycast, macOS |
The skill installs the Python packages on first use (see Install). The first local transcription downloads a Whisper model. YouTube Shorts are not supported.
video-lens reads YouTube's own captions first: fast, and nothing is downloaded. When a video has
no captions or they are blocked, it asks before it falls back to local transcription: yt-dlp
downloads the audio and Whisper transcribes it on your machine, then the report is written the
same way. This is slower than captions and needs the optional packages above.
/video-lens https://www.youtube.com/watch?v=... # a report, opened in your browser
/video-lens <url> in Spanish, focus on the pricing part # language or focus in your own words
/video-lens-gallery # every report you made, searchable
/video-lens-recs what's new in Rust this week # what to watch next, with reasons
Reports are saved under ~/Downloads/video-lens/reports/. After a report, say "share" for an
unlisted link on video-lens.org and "unshare" to delete it. Nothing is uploaded unless you ask.
Sharing, the gallery, recommendations and their privacy, Raycast and troubleshooting: docs/guide.md. What changed: CHANGELOG.md.
task test # fast suite, no networkTests and the repo layout: docs/development.md. PRs welcome; the
sources of truth are skills/*/SKILL.md and viewer/.
Found a bug or have an idea? Open an issue from the bug or feature template, or send a pull request.
MIT
