Most of the surface reads from collections, so what you get back depends on what has been loaded and processed there; describe_video and extract_video_entities both check for existing results before producing new ones. Source format matters in exactly one place: camera shot segmentation does not accept YouTube URLs, while chapter segmentation does. Long material arrives paged — 5-minute segments for descriptions, 25 videos per page for listings.
A server in front of Cloudglue video collections, where videos are indexed by their content — speech, on-screen text and visual description — rather than by filename.
- Semantic search at two grains: search_video_moments returns timestamped segments matched across speech, on-screen text and visual descriptions, and search_video_summaries ranks whole videos by topic and content similarity.
- Video descriptions through describe_video, which accepts YouTube, Cloudglue and direct HTTP URLs and returns results in 5-minute segments.
- Structured entity extraction through extract_video_entities, either pulling entities already stored in a collection or extracting new ones against a prompt.
- Bulk summaries and titles for an entire collection through retrieve_summaries, as a cheaper first pass than full descriptions.
- Automatic segmentation into camera shots or chapters, each returned with timestamps; segment_video_chapters also accepts Dropbox, Google Drive and Zoom connector URLs.
- Collection and video browsing with list_collections and list_videos, plus technical specifications — duration, resolution, file size and processing status — from get_video_metadata.
A Cloudglue API key, and videos reachable as Cloudglue, YouTube or direct HTTP URLs. Collection-scoped tools also need a collection that already holds the videos you want to query.