Own your conversations instead of renting them back

What it does Minutes records audio from your microphone or imported voice memos, transcribes and diarizes it entirely on your machine, then structures the output into markdown files tracking decisions, action items, commitments, and participants. A local SQLite relationship graph indexes people and topics across your entire meeting history, letting you or any connected AI agent ask cross-meeting questions like what you promised a specific colleague.

The interesting bit The project treats your conversation history as infrastructure rather than app data. By writing everything to plain markdown in ~/meetings/, it sidesteps the usual SaaS trap—no API keys needed for retrieval, and standard Unix tools still work in a decade. The MCP server and live transcript streams turn this static archive into real-time context for agents like Claude Code or Codex.

Key highlights

• Audio never leaves the machine unless you opt into cloud LLM summarization; transcription and speaker diarization run locally via Whisper and pyannote-rs.
• Exposes meetings through multiple surfaces: CLI, desktop app, MCP server, Claude Code plugin, and OpenCode integration.
• Relationship intelligence tracks commitments, flags stale promises, detects duplicate contacts, and surfaces contradictions across meetings.
• Live transcription streams to JSONL in real time, allowing agents to read and coach mid-session.
• Dictation mode converts speech to text progressively, with optional Apple Speech or Parakeet backends on supported macOS builds.

Caveats

• macOS cannot capture system audio directly for video calls; recording both sides of a Zoom or Meet call requires routing audio through a third-party loopback driver.
• The Apple Speech backend is currently experimental and limited to standalone live transcripts and dictation finalization, falling back to other engines for batch transcription.
• Windows builds require disabling default features and installing LLVM, suggesting the experience is less polished than on macOS or Linux.

Verdict Worth exploring if you want your AI assistants to actually remember your meetings without feeding another SaaS your voice data. Probably overkill if you just need occasional manual notes and do not already work with coding agents.