Granola

Get the verbatim transcript of a Granola meeting by ID; no summaries or notes. Returns up to 40000 characters per call by default (max_chars, up to 80000), cut at an utterance boundary with total_chars, truncated and next_offset reported. A page that splits an oversized utterance has ends_mid_utterance and the next starts_mid_utterance. With meeting_ids, meetings has one entry per requested ID in request order, with the single-read fields or an error. max_chars and offset apply per meeting, so a batch can return 10 times max_chars; lower max_chars when batching, and pass offsets to continue several meetings in one call. An accessible meeting with no transcript returns transcript null (status "unavailable" for chunk reads) and transcript_unavailable with a reason: transcript_pending means retry in a few minutes; not_happened_yet means retry after the meeting; no_transcript and transcript_deleted are final. "Meeting not found" means the ID does not resolve or is not accessible. Transcript blocks separate audio source (Microphone/System audio) from speaker identity; mobile/imported transcripts may keep Speaker A/B labels. For native recordings, recording context identifies the document owner as the recorder, not necessarily the speaker; missing recorder data stays unknown. Requires access to the meeting and transcript access permitted by its workspace. For specific details or quotes, first use search_transcripts with meeting_ids (batch known meetings together). Use this tool when the excerpts are insufficient or the user needs the full conversation. Start with the most relevant meetings; a detail missing from summaries may still be in the transcript. If access is denied, report that and do not retry through another tool. Keep at most 5 get_meeting_transcript calls in flight, and only for calls you already need. On a rate_limit error, retry only the failed call after retry_after_seconds, one at a time, and not when retryable is false.

get_meeting_transcript

Read one or more Granola meetings by ID: private notes, AI summary, attendees, metadata and note URLs. Use it when you have meeting IDs from discovery results or note URLs; batch up to 10 IDs per call. Private notes appear only for meetings the user owns (status "empty": no text; "not_included": another user owns the note). meeting_at is the calendar event's scheduled start (created_at without an event), so it can differ from when the meeting was recorded. Invalid, missing and inaccessible IDs are reported separately; the other meetings are still returned. The note creator and participants are not necessarily the connected user: confirm the connected user's identity with get_account_info and match it to explicit ownership evidence before drafting in the user's voice or assigning them actions; first-person notes are not proof of the user's voice. Dated plans do not establish current progress. known_participants comes from the calendar invite, attendees added by hand, and the note creator; anyone else who spoke appears only in the notes and transcript. Does not return transcripts; notes can omit details that were discussed. If a detail is missing, use get_meeting_transcript where that workspace permits; if access is denied, report that rather than conclude it was never discussed. Shared-screen screenshots appear in <attachments> with an ID, time offset and lossy description; for anything visual (designs, slides, charts, code), view them with get_meeting_attachment. Attached photos appear as <attachment type="photo"> without a description; view them with get_meeting_attachment. Keep at most 5 get_meetings calls in flight, and only for calls you already need. On a rate_limit error, retry only the failed call after retry_after_seconds, one at a time, and not when retryable is false.

get_meetings

Remove a Granola note from a folder. Use list_meeting_folders for the folder ID and meeting discovery results for the note ID. Removing it may revoke folder-derived access for other people. The response confirms the outcome (removed or not_in_folder) and the folder's visibility and audience.

remove_note_from_folder

Find the user's meetings by keyword, participant/company, date, or a combination. Keyword searches cover meeting titles, summaries and your private notes. They do not search transcripts: an empty result does not mean a term was never spoken in a meeting. Returns metadata only (title, date, participants, note headings), not what was said: read the relevant meetings with get_meetings before answering about their content, and use get_meetings directly for a meeting you already identified. With keywords or participant, results are ranked by relevance favouring recent meetings; with neither, newest first, defaulting to the last 7 days. Pass next_cursor as cursor with the same filters to continue; pages can be short or empty, so continue while next_cursor is present. completion_status incomplete means continuation is unavailable: narrow the search. participant matches participant metadata (names, emails, companies, domains), not mentions in notes, as an AND filter with the other inputs; metadata can be incomplete. For broader questions about a person or company, also search their name plus the topic in keywords with participant omitted: internal discussions may hold the answer. Captures of the same calendar event are grouped, keeping the user's own note or else the newest; read other_ids in duplicate_recordings only for details the listed note lacks or when the user asks to compare attendees' notes. Use attended_only=true for 'my meetings'. list_meeting_folders gives folder_id values. Keyword matching is literal, so pass the user's words plus obvious synonyms. Meetings with viewable attachments carry an attachment_count attribute; get_meetings lists them. Keep at most 2 search_meetings calls in flight, and only for calls you already need. On a rate_limit error, retry only the failed call after retry_after_seconds, one at a time, and not when retryable is false.

search_meetings

Find specific words or quotes in accessible meeting transcripts in the active workspace. Returns ranked passages with approximate timestamps. Supply meeting_ids when known, batching meetings that share the same terms and filters. Matching is literal; put alternative words in one keywords array, e.g. ["pricing", "price", "cost"]. For what a person or company said, use participant (names/companies) or attendee_emails/attendee_domains (invite list). participant is an AND filter on metadata that can be incomplete, so for broader questions also search their name and the topic in keywords without it. Known meetings with no indexed match are also searched in their current transcripts. Without meeting_ids, search relies on an asynchronously updated index and groups duplicate recordings; empty results do not prove a topic was never discussed. Follow next_cursor even on empty pages, passing only cursor and optionally limit. If completion_status is incomplete, narrow the search or read known transcripts. Read surrounding dialogue with get_meeting_transcript using meeting_id, the passage's chunk_ids and transcript_version; adjacent chunk IDs continue, and omitting chunk_ids reads the whole conversation. counts is cumulative across pages; total_matching_chunks is a first-page estimate, null for direct search. Neither counts spoken mentions. Use another tool when: - The term is likely in the notes, or the user wants meetings about a topic: use search_meetings - You need the whole conversation: use get_meeting_transcript

search_transcripts

Revoke the connected Granola account's consent and refresh tokens for this MCP client; use only when the user asks to sign out, disconnect or switch accounts. It affects every connection using the same client authorization, and new requests are rejected immediately. No browser action is needed. To switch accounts, reconnect Granola in this client; open the returned sign_out_url in the browser only if it auto-selects the previous account.

sign_out

Rename a Granola folder or change its description, only when the user asks. Get the folder ID from list_meeting_folders. Pass title, description or both; anything omitted stays the same, and an empty description clears it. Requires owner or collaborator role on the folder. Returns the updated folder.

update_folder

View up to 5 meeting attachments (screenshots of shared screens, or photos the user attached) as images in request order, each followed by its metadata including pixel width and height. Pass just meeting_id (or note_id) to browse its viewable attachments: described screenshots first in capture order, then photos; continue with pagination.next_offset as offset, and limit can request fewer. Or pass attachment_ids (or attachment_id) with the meeting_id or note_id they belong to, chosen from the <attachments> section of get_meetings results; browsing by meeting ID needs no get_meetings call first. A failed attachment returns an error with its ID without discarding the others. Thumbnail-sized or likely blank images carry an image_quality_notice instead of a description; do not use them as evidence. Each successful attachment includes a signed url and url_expires_at, valid for up to 10 minutes; offer it when the user asks to see or download the image (inline display depends on the client), and fetch again for a fresh link. Anyone holding the link can open it until it expires. A response holds about 6 MB of images, often only one to three screenshots; attachments that don't fit are listed in left_out_attachment_ids with a next_step, and pagination advances past the whole batch. Use this when a description alone does not answer the question, for example to read a slide, chart or code on screen.

get_meeting_attachment