Minutes

Minutes is a local-first meeting recorder and voice memo tool for people who want full ownership of their conversation history. Audio never leaves your machine. Transcripts land as plain markdown files on your disk, readable by any tool, searchable by any agent.

It's built for three overlapping audiences: developers who want inspectable, grep-friendly output; knowledge workers who need to capture decisions and action items without a SaaS subscription; and AI-heavy workflows where agents need structured context from past conversations.

How the pipeline works:

• Local transcription via whisper.cpp or the multilingual Parakeet backend, with GPU acceleration and streaming partial results as you speak
• Speaker diarization using pyannote, so multi-person meetings show who said what with timestamps
• Structured extraction pulls action items, decisions, and commitments into queryable markdown automatically
• Relationship memory tracks people, projects, and unresolved commitments across meetings, not just within one
• Voice memo pipeline watches for iPhone Voice Memos, transcribes them on your Mac, and folds them into the same memory layer
• Dictation mode sends spoken text to your clipboard and daily note without switching apps
• 31 MCP tools expose your transcript archive to Claude Desktop, Codex, Gemini CLI, and similar agents

Summarization is optional. Claude or Ollama-backed local models can work conversationally over your transcripts when you ask, using your existing subscription. No API keys are required for transcription.

The file format is a genuine design choice. Every transcript is a markdown file with YAML frontmatter. Timestamps, speakers, and action items stay visible in the raw file, so the source is readable before any assistant touches it. That also means the workflow survives outside Minutes entirely: git, grep, and any note-taking tool that reads plain files all work without modification.

Compared to tools like Meetily or cloud recorders like Otter.ai, Minutes adds cross-meeting intelligence and a broader agent surface. It's MIT licensed.