Your footage, queryable by Claude.

Search

Three vector spaces. Picking the right one is most of the work.

text_search

What happens, who's in it, what's said — descriptions, transcripts, faces, objects, OCR

visual_search

What the frame looks like — CLIP text-to-image with sharpness and quality filters

audio_search

What it sounds like — music, ambience, sound, not words

Orient

Know what exists before you search. Never guess a story the footage can't support.

library_overview

Months, people, locations, objects — what was actually shot, per year

entity_search

One record per real person, place, sport, or device, with screen-time stats

index_stats

Index health — tells an empty index apart from a missed query

list_sources

Every indexed file — the vocabulary for scoped search

Read one video

From gist to full timeline, cheapest read first.

get_video_summary

2–4 sentence summary plus chapter map

get_video_chapters

Deterministic chapter spans, always available

get_video_scenes

The whole timeline in capture order, with joined transcript

filename_search

Resolve a half-remembered filename to its scenes

Filter by metadata

Exact ranges, no vectors — a range means a range.

camera_search

Drone / phone / action / professional families and recording modes

cinema_search

Sony MXF cine data: ISO, aperture, focal length, colour, timecode

speaker_search

Who is speaking, narrowed by topic

face_emotion_search

Whose face is on screen, optionally by emotion

person_moments

One person, on screen and speaking about a topic, in the same scene

telemetry_search

Speed, motion, altitude, GPS bounding boxes

strava_search

Heart rate, distance, grade, sport from matched activities

Look and export

Judge visually, then land on the timeline.

contact_sheet

Up to 48 thumbnails in one labelled grid — pick the best shot by looking

send_to_resolve

Append chosen clips to the open DaVinci Resolve timeline

get_resolve_timeline

Read back the timeline name, fps, and clips

Typical flows

• Documentary in an afternoon

  1. library_overview(2026)→
  2. pick a narrative the footage supports→
  3. browse_scenes by month→
  4. contact_sheet→
  5. send_to_resolve

• The exact line

  1. text_search("she explains the pricing")→
  2. get_video_scenes(file)→
  3. full transcript, no paging

• Best b-roll

  1. visual_search(…, max_per_source=1, min_sharpness=100)→
  2. contact_sheet→
  3. pick a cell