Vocci

Use for oversized, preparation, digest, or briefing-style requests over the authorized user's Vocci/ring-captured memory (ring recordings, Vocci meetings, captured conversations, transcripts, summaries, notes or files in the Vocci Files tab). Builds a compact, capped, source-attributed context pack by composing search and fetch. Supports the same mode/depth, sourceTypes, startTime/endTime, and minDurationMinutes/maxDurationMinutes controls as search (a four-hour session is 240 minutes). Use sourceTypes note/file for notes and files only; use depth deep only when body-only transcript or note matches are worth extra latency.

context_pack

Use after Vocci search, or when the user provides a stable ID or vocci:// URI. Fetches one read-only Vocci session, transcript, summary, note, or file with optional capped text content. For action-item or decision extraction, read transcript text - not a snippet or the (often empty) summary - since items are routinely buried mid-conversation. fetch returns only one capped page (up to 200 events) and marks longer transcripts truncated; session_get with includeEvents:true and projection:"transcript_text" is the better transcript-reading view, but it is also capped by eventsLimit, so treat truncated results as incomplete. Speaker labels (e.g. 'Speaker 1') are per-recording diarization labels, NOT stable identities - the account owner is the ring wearer and one of the speakers, but which label is the owner is unresolved. Do not attribute action items to a specific person from the label alone; mark the owner UNKNOWN when context doesn't disambiguate.

fetch

Create a short-lived signed URL for one owned Vocci audio file. The tool does not inline raw audio bytes in the MCP response.

audio_download_url

Create short-lived signed URLs for 1 to 100 explicitly identified owned Vocci audio files in one bounded call. Results preserve input order and report per-file failures; raw audio bytes are never inlined.

audio_download_urls

Read metadata for a file in the Vocci Files tab and, for small text documents, capped content.

base_file_get

Read one Vocci session together with its notes. A session's summary and notes are attached files, returned by default on first-page reads (set includeFiles:false to skip; follow-up transcript pages with eventsCursor skip notes unless includeFiles:true): small text notes - like the summary - come back with their text inline, so this single call gives you the content (larger or binary notes return as metadata; fetch those with base_file_get). The session's own summary field is usually empty in current sessions - the notes hold the real summary. For action-item or decision extraction, read transcript text with includeEvents:true and projection:'transcript_text', then check transcript/events truncated flags; events are capped by eventsLimit (max 200) per page. To read a long transcript to the END, repeat session_get with eventsCursor set to the previous response's eventsNextCursor, until eventsNextCursor is absent - that retrieves every turn without overflowing a single response. truncated:false and an absent eventsNextCursor mean there are no further STORED turns - NOT that the transcript covers the whole recording: speech the transcriber did not pick up leaves no turn behind at all. To judge coverage, compare the last returned turn's timestamp, measured from session.createTime, against session.durationMillis; a large shortfall means thin transcription, not a page you failed to fetch. That comparison is only valid when session.durationComplete is true - when it is false, session.durationMillis is a LOWER BOUND on the recording (segments are still being processed) and cannot establish coverage at all. Speaker labels (e.g. 'Speaker 1') are per-recording diarization labels, NOT stable identities - the account owner is the ring wearer and one of the speakers, but which label is the owner is unresolved. Do not attribute action items to a specific person from the label alone; mark the owner UNKNOWN when context doesn't disambiguate.

session_get

List the user's files and notes from the Vocci Files tab, most-recently-updated first. Use this to browse or enumerate notes/files; optional lastSyncTime/limit narrow the window. For a topical lookup, use search instead.

base_files_list

List a bounded, non-pageable snapshot of audio files linked to one Vocci session. Check audioFiles.isComplete before treating it as complete; use audio_download_url for a short-lived raw audio URL.

session_audio_list

List the user's Vocci sessions, most-recently-updated first (each with title, status, and summary). Use this for 'recent/latest/last meeting' or to browse; optional lastSyncTime/limit narrow the window, and optional sessionType ('recording' or 'chat') filters the page before the limit. For a topical lookup, use search instead.

sessions_list

Use when the user asks about Vocci or the authorized user's Vocci/ring-captured memory, including ring recordings, Vocci meetings, captured conversations, voice notes, transcripts, summaries, notes or files in the Vocci Files tab, or something they said or heard in Vocci. Vocci is an AI smart ring and companion app that records the wearer's conversations, meetings, and voice notes, then turns each recording into a session with a transcript and an AI summary. Search this read-only user memory first for broad or topical questions. Searches scoped sessions, transcripts, summaries, notes, indexed text file content, and file metadata, then returns snippets and stable IDs for follow-up fetch. Use sourceTypes note/file for notes and files only, or session/summary/transcript for recording-only retrieval. Use one mode argument when needed: semantic for Pinecone-only indexed transcripts and text file content, lexical for exact scans, or hybrid to combine them. Mode defaults to hybrid at every depth; depth only controls how broadly the selected mode searches. Use depth fast for a cheap first pass; retry with standard or deep only when the desired session, transcript, note, or file is not found. Deep hybrid can scan capped transcript and note bodies. Optional startTime/endTime bound the window, and minDurationMinutes/maxDurationMinutes filter sessions, summaries, and transcripts by recording length. Snippets are lossy previews for locating content - do not extract tasks, decisions, or action items from a snippet. Once you locate the right session, read transcript text with session_get (includeEvents:true, projection:'transcript_text') before concluding what was said or what to do, and check the returned truncated flags - session_get and fetch can return capped transcript pages.

search