Son Piaz

watch-cli

Gives a coding agent eyes and ears for any social video.

Live · · site · repo · Updated

The problem #

An AI agent cannot watch a video. A conference talk, a product demo, a twenty minute tutorial: all of it is invisible to the model reading your terminal. The workaround is to send the whole file to a multimodal API, which is slow and costs real money per minute.

The approach #

Pull the video, cut it into timestamped frames, transcribe the audio, and hand the agent both. The agent sees the diagram on the slide and reads the sentence that explains it, so it can cite the moment ("at 04:32") instead of guessing. The trade is fidelity: you lose motion, you keep meaning, and the cost drops by roughly fifty times.

Where it is now #

Shipped and maintained. v0.3.4 as of 2026-09-04, installed with one curl command or as a Claude Code skill. Every watch is cached locally, so a repeat call on the same URL returns in 0.07 seconds and costs nothing.

What happens if this works #

Video is the largest body of engineering knowledge that agents still cannot read. When watching is one line in a shell script, the default way to learn a new tool stops being the docs and starts being the talk somebody already gave.

Evidence #