Skip to main content
You talk to the connector in plain language and the client picks the tools. Each recipe below pairs the prompt you type with the chain it runs, so you know what to expect and what to fix when a step is missing.

​
Highlight reel from a long video

What runs
  1. upload_media - pulls the URL into the collection and returns the m- video ID
  2. index_spoken_words - transcribes and indexes everything spoken
  3. create_highlight_clip - matches your prompt against the index and compiles the hits into one stream
Indexing has to finish before the search returns anything. create_highlight_clip matches spoken content by default, and on an unindexed video it comes back with an empty shots list. To control the cut yourself, ask for search_video to get timestamps, then generate_clip for one video or compile_clips to stitch across several.

​
Transcript Q&A

What runs
  1. generate_transcript
  2. get_transcript - plain text, or segments timestamped per word, sentence or time window
  3. understand_video - runs the transcript analyzer over the video
  4. create_index - builds a retrieval-ready index from that analyzer’s output
  5. ask_video - answers from the index, with sources
ask_video and semantic_search read only indexes built by create_index, so a transcript on its own is not queryable. While step 3 runs, get_understanding reports the status of each analyzer.

​
Subtitled clip for social

What runs
  1. reframe_video - re-crops the 2:10-2:40 segment to vertical and saves it as a new video in the same collection
  2. index_spoken_words - on that new video, since subtitles come from its own index
  3. add_subtitles - burns them in and returns a stream URL
  4. download_stream - renders the stream to an MP4, which needs a file name
Order matters here. reframe_video and add_subtitles both take a video ID rather than a stream URL, so the reframe goes first and the new video gets its own index. download_stream is the only step that accepts a stream URL.

​
Search across a whole library

What runs
  1. list_collections - IDs, names and visibility
  2. list_videos - the videos in the collection you named
  3. index_spoken_words - once per video
  4. search_collection - matching moments grouped by video, with start/end timestamps and relevance scores
search_collection reads spoken-word and scene indexes. ask_collection answers in prose instead of returning moments, but it reads v2 indexes, so every video needs understand_video and then create_index first.

​
Live camera monitoring with alerts

What runs
  1. connect_rtstream - registers the RTSP, RTMP or HLS feed and starts ingesting it
  2. index_rtstream_scenes - describes frames on a rolling window using your prompt
  3. create_event - the condition to watch for, written in natural language
  4. create_rtstream_alert - attaches that event to the index with your callback_url
  5. search_rtstream or get_rtstream_scenes - search the indexed records, or read them straight through for a time range
Alerts POST to a callback URL you control. See Alerts and callbacks for the payload fields. Real-time timestamps are Unix seconds, and get_rtstream_stream turns a start/end window into a playable URL.

​
Record and search a meeting

What runs
  1. record_meeting - sends the bot to the Zoom, Google Meet or Microsoft Teams URL
  2. get_meeting - poll until status is done and video_id appears (initializing, joined, processing, done); the speaker timeline comes back here too
  3. index_spoken_words - on that video_id
  4. search_video - moments with timestamps
For a prose answer rather than moments, run the Transcript Q&A chain on the recorded video: understand_video, create_index, ask_video.

​
Working with long jobs

Uploads, indexing runs and renders that outlast the request budget return {"status": "processing"} and a retry_with field naming the read tool that will show the result, such as get_meeting, get_transcript, get_transcode_status or list_videos. The job keeps running. Ask the client to check with that tool; re-issuing the write starts a second job.

​
Tips

  • Name the collection you mean. default always exists and is where things land when you do not say.
  • Index before you search. An unindexed video returns nothing from search_* and ask_*, and a timeline caption on one renders with no captions and no error.
  • Every stream URL has a matching player URL that opens in a browser.
  • ID prefixes tell you what you are holding: m- video, a- audio, img- image, rts- real-time stream, c- collection.
  • Deletes are permanent. delete_video, delete_collection, delete_index and delete_scene_index have no undo. remove_video_storage drops the stored bytes but keeps the record, indexes and transcripts.

​
Next Steps

Tool Reference

Every tool, grouped by toolset, with what it does and what it needs first

Connect the MCP Server

Setup for Claude, Claude Code, Cursor, MCP Inspector and other clients

MCP Server Overview

Endpoint, auth, toolsets, and what the connector does not cover

Examples and Tutorials

Longer builds that use the same pieces outside a chat client