Minutes — Meeting Memory for AI

Overview

Minutes is conversation memory for AI. Record, transcribe, search, and query every meeting and voice memo, all processed locally on your machine. Works with any MCP client: Claude Desktop, Claude Code, Codex, Gemini CLI, and more. 24 tools • 7 resources • Interactive MCP App dashboard • 6 prompt templates **Setup:** `npx minutes-mcp` is all you need to start. Search, browsing, and the dashboard work immediately with zero setup. For recording and transcription, the server auto-installs the CLI binary on first use (or install manually via `brew install silverstein/tap/minutes` or `cargo install minutes-cli`). Then run `minutes setup --model tiny` (75MB speech model) and optionally `minutes setup --diarization` (~34MB speaker ID models). **Record & Transcribe:** • Start/stop live meeting recordings • Process audio files (WAV, M4A, MP3, OGG) • Local transcription. Nothing leaves your machine. • Speaker diarization, native, no Python required • Language selection for similar-sounding languages (e.g. Urdu vs Hindi) • Dictation mode: speak to clipboard and daily notes • Live transcript with delta reads for mid-meeting AI coaching • Safety guard auto-stops forgotten recordings after prolonged silence, time cap, or low disk **Search & Query:** • Full-text search across all meetings • Find open action items by assignee • Track decisions and commitments • Cross-meeting topic research **Relationship Intelligence:** • Person profiles across all meetings • Relationship map with losing-touch alerts • Commitment tracking by person • Consistency reports flagging conflicting decisions **Speaker Identification:** • Voice enrollment and speaker profiles • Confirm or correct speaker attribution in transcripts **Interactive Dashboard (MCP App):** • Visual meeting browser, detail viewer, and people dashboard • Renders inline in Claude conversations **Meeting Insights:** • Structured insight events: decisions, commitments, questions with confidence levels • Query insights by type, confidence, participant, and time range • Only high-confidence insights trigger downstream actions **Agent-Friendly:** • MCP resources for stable context (recent meetings, action items, events) • Subscribable insight events for workflow automation • Structured YAML frontmatter on all output **Privacy:** All audio processing happens locally via whisper.cpp and pyannote-rs. No audio or transcripts are sent anywhere. Summarization is optional and uses your own LLM connection.

Links