Turn a local video file into timestamped contact-sheet JPEG(s) that a vision model can read: frames sampled evenly across the clip, each cell stamped with its timecode, plus an aligned audio-activity band. The band is globally normalized to the full clip and shows energy only — it does not transcribe, classify, or identify sounds. Use peaks to propose where to zoom; use the frames to determine what happened. Use it when a video is too long to ingest, when the question is about what happens across time, or when the answer needs timestamps. One call replaces a whole ffmpeg → extract → montage pipeline — prefer it even if you have a shell. Read the returned sheet file(s) with vision and cite the timecodes. Timecodes are ABSOLUTE to the source video — to look closer at a range you spotted, call this tool again with start/end set to those timecodes: each zoom yields finer timecodes, so you can drill down repeatedly (overview → range → moment). The JSON result includes audio.samples[] with absolute time and normalized level values. Runs entirely on-device; requires ffmpeg on PATH.