add_meetings
ChatGPTAdd meetings to a collection the user owns or collaborates on. Everyone who can see the collection can then read these meetings, so precision matters more than recall: when unsure whether a meeting belongs, leave it out and list it for the user to decide. Always preview first 1. Call with dry_run: true. Nothing changes. You get one row per meeting with its outcome, plus a preview_token (valid 10 minutes). 2. Show the user the preview in plain words (use summary; never show outcome codes) and wait for them to confirm. 3. Call again with the same collection_id and meeting_ids, dry_run: false and the preview_token. If apply returns preview_stale or invalid_preview_token, preview again; do not retry the apply. Outcomes: will_add; already_in_collection; not_allowed (the user cannot share this meeting); not_found (no access, deleted or expired). Only will_add meetings change. Meetings the user cannot share are never added.
create_collection
ChatGPTCreate a recording collection owned by the user, optionally with its first meetings in one go. Call list_collections first. If the user already owns a collection with exactly this name, no duplicate is made: without meetings the existing one is returned with created: false; with meetings they are previewed and added to it. If several have that name, the call is refused with ambiguous_collection_name: ask the user which one, then use add_meetings. Without `meeting_ids`: creates an empty collection right away. Do not pass dry_run; there is nothing to preview. With `meeting_ids`: works like add_meetings. Find the meetings first, then: Always preview first 1. Call with dry_run: true. Nothing changes. You get one row per meeting with its outcome, plus a preview_token (valid 10 minutes). 2. Show the user the preview in plain words (use summary; never show outcome codes) and wait for them to confirm. 3. Call again with the same name and meeting_ids, dry_run: false and the preview_token. If apply returns preview_stale or invalid_preview_token, preview again; do not retry the apply. Nothing is created on the dry run. Apply creates the collection and adds the meetings together. Everyone who can see the collection can read its meetings, so leave out any meeting you are unsure about and list it for the user.
get
ChatGPTFetch full content of a specific meeting by its ID (e.g. "meeting:abc123"). Use after list_meetings once you have identified a relevant meeting ID. To fetch multiple meetings at once, use fetch. Use the optional include array to request only the sections you need: "summary" (full meeting summary text), "action_items" (complete action item list with owners), "decisions" (all key decisions), "chapters" (chapter titles + detailed summaries), "participants" (names + emails where available), "transcript" (raw transcript — large; use with chapter_index to scope to one chapter), "metadata" (platform, duration, series ID, org domain), "collections" (recording collections this meeting belongs to — id + name). Omitting include returns all sections except the transcript. To fetch a transcript for a specific chapter, pass include: ["transcript"] and chapter_index (0-based index into the chapters array).
get_live_meeting
ChatGPTLive transcript of an in-progress meeting. Use get for ended meetings. No id: returns the transcript when one meeting is active, or JSON {active_meetings[{id,title,platform,started_at,last_line}]} when multiple are active. With id (meeting/bot id): returns that meeting's transcript as Speaker: words lines prefixed by a header (Meeting ID, Title, Platform, Started at). Requires live captions enabled. Only reliable in English.
get_screenshots
ChatGPTScreenshots from a meeting recording, as time-limited image URLs. Reach for it whenever a recording's visuals carry meaning the transcript cannot: turning a recorded demo or walkthrough into a slide deck or illustrated step-by-step guide, writing onboarding and how-to docs with real screenshots, analysing the slides, dashboards or UI that were actually on screen, or feeding a design/PR review to a coding agent alongside what was said. id is the meeting id get and list_meetings return. Images cost real tokens Frames are full-size (typically 1280×720); there is no thumbnail — a smaller image invites confident guesses from an illegible one. Default 3 frames, ceiling in max_limit. Use format: 'index' to survey a meeting's frames with no images and no cap. Roles are capture reasons, not content A role says why the recording was sampled then, not what the image shows. All roles extract from the same composite video, so a hero/filmstrip frame taken while someone presented shows their screen too. • screenshare — every ~30s inside a reported share window; likeliest to show shared content. • hero — the most representative frame. • filmstrip — chapter and settled-start moments. role defaults to screen_content (screenshare + hero); any adds filmstrip. Zero `screenshare` never means nobody shared — the sampler did not run; check the other roles. Scoping Offsets run from the recording start. Pass exactly one of at, chapter_index or from_time/to_time, or none for the whole meeting. Prefer at for "what was on screen when…" — capture is irregular, so an exact window usually misses. Response frames[] carry url, timestamp, role, chapter_index, measured width/height and the max_width/max_height ceiling. URLs are JPEGs that expire (expires_at); re-sign one by calling again with at. available_by_role counts roles in the window; meeting-wide is meeting_screenshot_count. An empty result names the cause and carries nearest_frame.
list_collections
ChatGPTList recording collections the user has access to. Returns per collection: id, name, recording_count, role (owner | collaborator | viewer), created_at. Use the returned collection id with list_meetings (pass as collection_id) to filter meetings within a specific collection. Collections are curated groups of meetings organized by the user or shared with them. Only collections the user has at least viewer access to are returned.
list_meetings
ChatGPTStep 1 of 2: list and filter meetings with lightweight metadata. Returns per meeting: id, title, date, participants (names), platform, is_external, has_transcript, url, tldr, summary_snippet (≤600 chars: top action items/decisions/agenda), chapters ({title, summary, subsections[]} per chapter), action_item_count, decision_count. Does NOT return: full summary, full action/decision lists, or transcript. Use get or fetch once you have relevant IDs. Search strategies (chapter text is NOT indexed) 1 — Title (instant): title_query when the topic is likely the meeting's main subject. 2 — Chapter scan: Request fields:["id","title","date","chapters"], scan chapters[].summary across results, then fetch matches. 3 — Narrow first: Combine from_date/to_date, participant, platform, or recurring_meeting_series_id to shrink the set. Use limit:100. 4 — Collection: Use collection_id from list_collections to scope to a curated group of meetings. 5 — Code-filter: Paginate and filter chapters[].summary programmatically before fetching. If a result looks relevant, call get with its ID to retrieve full content before answering. Use fields to return only what you need (reduces tokens; id always included). Paginate via next_cursor → cursor.
remove_meetings
ChatGPTRemove meetings from a collection the user owns or collaborates on. The meetings themselves are not deleted, only their membership in this collection. A meeting the collection added by rule stays out on later rule updates. Always preview first 1. Call with dry_run: true. Nothing changes. You get one row per meeting with its outcome, plus a preview_token (valid 10 minutes). 2. Show the user the preview in plain words (use summary; never show outcome codes) and wait for them to confirm. 3. Call again with the same collection_id and meeting_ids, dry_run: false and the preview_token. If apply returns preview_stale or invalid_preview_token, preview again; do not retry the apply. Outcomes: will_remove; not_in_collection. Only will_remove meetings change.
search_meetings
ChatGPTHybrid BM25 + semantic search across indexed meeting chapters and subsections. Returns meetings ranked by Reciprocal Rank Fusion (RRF) of keyword and semantic signals. Each result includes: • id — use with get to fetch the full meeting • score — RRF relevance (keyword + semantic combined) • title, date, url • matched_chapters — up to 3 chapters/subsections that drove the match Pass include: ["transcript"] to add verbatim transcript text to each matched chapter. vector_weight (0.0–1.0, default 0.5) Balances semantic (vector) vs. keyword (BM25) scoring in RRF fusion: • 0.5 — balanced hybrid; best default for most queries • 0.7–1.0 — prefer semantic; use for intent/concept queries where the user expresses meaning rather than exact words: "meetings where the team seemed frustrated", "discussions about technical direction", "when did we talk about burnout" • 0.1–0.3 — prefer keyword; use when the query contains specific names, IDs, ticket numbers, or exact quoted phrases: "SAI-3364", "Kirill's comment about the deadline", "bun.lock conflict", "unfunny remarks" • 0.0 — keyword only (BM25); 1.0 — semantic only (vector) Metadata filters (from_date, to_date, platform, is_external) narrow the candidate set. Max 100 results (default 25). No cursor — ranked by relevance, not date. For full content, call get with a returned id.
Fetch
ClaudeSearch
ClaudeTestConnection
Claude