Vision — Cue

Cue is the local-first video timeline evidence tool for agents.

• Identity: package @sylphx/cue, bin cue, MCP io.github.SylphxAI/cue, site https://sylphxai.github.io/cue/.
• User: an agent that must find a quote, a chapter, a code demo, or a visual moment in a local video.
• Job: the default read is a searchable timeline of streams, chapters, and embedded subtitles. Scene boundaries are a separate request. A frame, crop, or OCR is a video_evidence follow-up.
• Promise: every claim points at a timestamp, stream index, subtitle range, or source hash. A frame index appears only after a frame follow-up. Missing subtitles are reported as gaps. The shipped server does not run speech recognition.
• Defaults: omitted profile is fast (container metadata, streams, chapters, embedded subtitles). quality adds ffmpeg scene detection. It does not extract frames or run speech recognition.
• Boundaries: Cue owns the video timeline. It does not own cloud video APIs, frame-by-frame VLM summaries, or generic media-library management.