Claude Eyes CLI
The eyes CLI and skill for agents: a video index, search by words and on-screen text, close-up frames and editing metrics
- 76 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- The tool downloads clips by link and installs ffmpeg, whisper.cpp and speech models on the computer
- Cloud transcription by flag sends audio to the third-party ElevenLabs service
Install
- Works in
Manual install
uv tool install git+https://github.com/moritzkremb/claude-eyesThen eyes setup --fix, eyes install-skill and eyes setup. pipx install or pip install --user also work.
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The tool downloads clips by link and installs ffmpeg, whisper.cpp and speech models on the computer
- Cloud transcription by flag sends audio to the third-party ElevenLabs service
This is third-party code. Review the repository files before installing.
What it does
Agents read text and look at images, but video is out of reach for them. Eyes turns a clip into material an agent can read: the `eyes index` command takes a link or file and builds a cache in `~/.cache/eyes/` with five tracks. They are frames sampled at 0.5 to 4 per second on contact sheets with timestamps, cuts, words said, on-screen words with position and time, plus loudness, silences, tempo and palette. The agent opens one sheet instead of hundreds of screenshots. Then the agent navigates as a person would. `eyes search` finds a word in one clip or in the whole library and shows whether it was said or shown. `eyes crop` cuts a region of a frame at full resolution to read small text or commands, `eyes burst` gives eight frames in a second and `eyes frames` takes exact moments and ranges. By default a local whisper.cpp makes the transcript with word-level timestamps, and creator-uploaded captions are used first. `eyes dna` measures the editing of a reference: cuts per minute, shot length, captions, loudness and palette. Your own edit is compared with it metric by metric, and the same metrics help study competitors and find moments for shorts. Everything runs on your computer and nothing is uploaded, except optional cloud transcription with ElevenLabs Scribe through the `--transcriber elevenlabs` flag.
Who it is for. Video editors and short-form creators who work with Claude Code, Codex or Cursor, and anyone who wants an agent to break down tutorials and screen recordings.
Good fit when
- A 20-minute tutorial must become a step-by-step guide with exact commands read off the screen
- You need every moment in a library of clips where a word was said or appeared on screen
- You want to check a clip edit against a reference by cuts per minute, shot length and captions
- A long video should yield moments for shorts with exact in and out points
Not a fit when
- You need to edit or cut a video: Eyes only reads and measures a clip
- You need to generate video: the tool creates nothing
- You need soft dissolves detected: the author admits the tool misses them
Example request
Watch this 20-minute tutorial and give me every setup step with the exact commands.Limitations
Python 3.10 or newer is required, and macOS, Linux and Windows are supported. The `eyes setup --fix` command installs ffmpeg, whisper.cpp and a speech model where it can, and downloads components. Indexing a one-minute clip takes about half a minute and a 20-minute video takes a few minutes. On-screen text recognition runs through Apple Vision on macOS and RapidOCR elsewhere.
Cut detection finds about 96 percent on test edits (88 of 92) but misses soft dissolves, which are on the author roadmap. Cloud transcription through ElevenLabs needs `ELEVENLABS_API_KEY` and sends audio out, so it is switched on only by a flag.
How to disable. Remove the tool with `uv tool uninstall` and the package name (or `pipx uninstall`), then delete the skill from your agent skills directory. The index cache lives in `~/.cache/eyes/` and can be deleted by hand.
FAQ
How do I install Claude Eyes?
Paste to your agent: "Install Eyes from https://github.com/moritzkremb/claude-eyes and set it up for yourself". The agent runs the four commands from AGENTS.md: `uv tool install`, `eyes setup --fix`, `eyes install-skill` and `eyes setup`.
Can I install only the skill?
Yes, with `npx skills add moritzkremb/claude-eyes`. The skill installs the CLI itself the first time it is used.
Are my videos uploaded anywhere?
No, indexing is local and whisper.cpp makes the transcript. Audio goes to ElevenLabs Scribe only if you add `--transcriber elevenlabs`.
How do I find a word or code on screen in a video?
Use `eyes search word`: results mark hits as said or shown with a timestamp. Then `eyes crop ID 12:02.4 --box 0,0,0.5,0.5` cuts a region of the frame at full size.
What is editing DNA?
It is the measured editing profile of a reference from the `eyes dna` command: cuts per minute, median shot length, words per caption, loudness and palette. Your own clip is compared against it metric by metric.
Discussion
No questions or reviews yet.
Related
CLI toolsA video editing skill for Claude Code and Codex: the agent cuts pauses and filler words, grades color, burns subtitles and adds animated overlays
Claude Code directs a video: transcription, script, HyperFrames scenes, subtitles and Remotion assembly with a localhost dashboard
Skills, a CLI and an MCP server that let Claude Code or Codex make video from a brief: editing, graphics, voiceover, subtitles
Claude Real Video
claude-real-video
A skill that lets Claude watch a video: it pulls scene-aware, deduplicated frames and a speech transcript through whisper from a link or file
A local video studio for your coding agent: describe a video in one sentence, the agent directs, your machine renders with no cloud
A Claude Code skill that turns a talking-head video into a short motion-graphics edit and checks its own cut
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Repository stars | Main difference |
|---|---|---|---|
| Claude Eyes (this card) | no | 76 | |
| AI Video Transcriber | no | 3,411 | transcribes clips into text, while this one adds frames, on-screen text, library search and editing metrics |
| YouTube AI Video Editing System | no | 145 | edits video with an agent, while this one only reads and measures other clips and references |
README in short
The description starts with the problem that an agent cannot watch video and explains three steps: index, navigate and measure. It then details the index tracks, cut detection with test figures, word-level timestamps and transcription options.