Claude Eyes: let your coding agent watch video and measure an edit's 'DNA'

Most agent video pipelines are blind after render. Claude Eyes (MIT, created Oct 15, still being updated today) closes that gap. It turns any video into evidence an agent can read: a transcript with timestamps, on-screen text, and contact sheets. It also gives the agent commands to move around inside the video: eyes index, eyes search, and eyes crop at a timestamp and box to read text that was never spoken. Everything runs locally; the README says nothing is uploaded.

The ReelStack angle is 'editing DNA'. Point it at a reference edit and it measures cuts per minute, shot lengths, caption style, loudness and palette. Then it checks your edit against the reference, metric by metric (match, close or off). The README's own demo is a good example of honest QA: Eyes flagged narration at 222 words a minute and two empty moments in the project's video, and admitted it misses soft dissolves, which is now on the roadmap.

Install is agent-first. Paste 'Install Eyes from https://github.com/moritzkremb/claude-eyes and set it up for yourself' into Claude Code, Codex or Cursor, or run uv tool install, eyes setup --fix (ffmpeg, whisper.cpp and a speech model), and eyes install-skill. You can also add just the skill with npx skills add moritzkremb/claude-eyes. It works on macOS, Linux and Windows with Python 3.10+.

Our take: pair it with any HyperFrames, Remotion or claude-motion pipeline as the 'watch it back' step. Use it to study a hook, pacing and caption style before you brief the agent, then confirm the render actually matches. Spotted, not tested. Credit moritzkremb.