“Next three consultant signups get monthly. Annual stays the enterprise default.”
The product, in motion
The interface is the evidence.
Minutes captures locally, writes one inspectable Markdown file, and lets your agents recall the decision with a source. No proprietary workspace stands between you and your own history.
Transcript Output
2026-04-08-strategy-sync.md
2 speakers42 min3 actions
09:02matWe should switch consultants to monthly billing instead of annual.09:04danaTest it on the next three signups first and compare retention.09:11matMinutes, capture that as a pricing experiment and link it to Q2 planning.
Action Items
☐Test monthly billing with the next three consultant signups
☐Compare retention and payback against annual billing
☐Review experiment results in next week's pricing sync
Minutes keeps the raw transcript visible. The structure is the interface: timestamps, speakers, action items, and decisions stay readable even before an assistant touches them.
v0.24.0 adds minutes resummarize to re-run the AI pass on an edited transcript, refreshes derived views after an applied pass, and fixes a dictation double-paste. Release notes · Feed
Download, install, done. First launch downloads a speech model. Runminutes setup --demoto try the pipeline on five bundled fixture meetings.
Newer Homebrew distrusts third-party taps by default; if brew warns about silverstein/tap, run brew trust silverstein/tap once.
Reading on your phone? Minutes installs on Mac and Windows. Star the repo so it is waiting when you are back at your desk.
Works with any MCP client
Claude Code/Codex/Gemini CLI/Claude Desktop/Cowork
01Dictation
The fastest way in is your voice.
Hold the hotkey, speak, release. The text lands at your cursor and in your daily note — no app to open, no tool to switch. It's the habit most people start with, and it runs on the same local engine as every meeting and memo.
Other dictation tools hand you text and forget it. Minutes keeps every word — transcribed on your machine, part of the same owned memory your AI can search later. Your voice never touches a server.
Dictation
daily-note.md · 2026-07-10
⌥Spacehold to talk
08:14Remind me to send the Q3 numbers to the board before Friday.
11:02Onboarding idea: defer the model download until the first recording.
15:47Dana owns the pricing experiment — review the results next week.
Transcribed on-device · appended locally · searchable by your AI later
02On-device
Your conversation never leaves your machine.
Every other tool sends your audio somewhere to be understood — a cloud transcriber, a hosted AI, a server that keeps the result. Minutes doesn't. Transcription runs on your Mac and the record is markdown on your own disk. There is no server to trust, breach, or subpoena.
Cloud notetakers
Leaves your device- Capture from your micon-device
device audio, on your Mac
- Transcribe☁ cloud
cloud providers
- Enhance notes☁ cloud
hosted AI
- Store transcripts + notes☁ cloud
their servers
Even the privacy-conscious ones capture locally, then stream your audio to the cloud to transcribe and store the result on their servers.
Minutes
Stays on device- Capture from your micon-device
device audio, on your Mac
- Transcribeon-device
on-device — sealed local whisper.cpp
- Store transcripts + noteson-device
your disk — markdown in ~/meetings
Audio, transcript, and notes never leave your machine. You hold the only copy — nothing to upload, sell, or lose in an acquisition.
03Proof
See it work before you believe it.
Try it in two minutes
One command installs five sample meetings and connects them to your AI. Ask Claude what was decided, without recording anything.
How we test it
2,706 automated test declarations and a public eval harness, with the limits stated plainly. Read the receipts before you trust it.
Build on the files
Working Mem0 and Graphiti examples show how other tools read the same markdown. Your data, everyone's integration.
04Pipeline
How it works
Audio -> Transcribe -> Diarize -> Summarize -> Markdown -> Policy-safe Search
(local) (local) (your LLM) (decisions, (live authorized,
whisper.cpp pyannote Claude / action items) bounded reads)
OllamaTranscription runs locally through sealed whisper.cpp. Existing Parakeet and Apple Speech preferences resolve to Whisper until their secure byte transports pass their release gates. Summarization is optional — Claude can do it conversationally when you ask, using your existing subscription. No API keys are required to get useful output.
05Audience
Capture it anywhere. Find it everywhere.
From meetings to memos to agents
For agents
Local audio context
34 MCP tools, live transcript reads, and structured markdown let Claude, Codex, Gemini CLI, and Cowork work from what was actually said.
For developers
Local and inspectable
sealed local whisper.cpp transcription, diarized markdown, YAML frontmatter, and a plain-files workflow that still works with grep and git.
For meetings
Capture what matters
One-click recording, streaming transcription, speaker separation, decisions, and action items without shipping your audio to a SaaS vendor.
For voice memos
Phone to desktop
Minutes watches for iPhone Voice Memos, transcribes them on your Mac, and makes them available to the same memory layer.
For daily work
Dictation that stays useful
Hold the hotkey, speak, release. Minutes sends the text to the clipboard and your daily note without changing tools.
For recall
Answers from raw output
Competitors hide the transcript. Minutes keeps timestamps, speakers, and action items visible so the source stays readable.
06Features
What you get
Capture
>
Local transcription. whisper.cpp with GPU acceleration. Your audio stays on your machine.
>
Streaming results. Text appears as you speak, with partial updates every few seconds.
>
Speaker diarization. pyannote separates who said what in multi-person meetings.
>
Dictation mode. Clipboard + daily note flow for short-form thoughts and commands.
Intelligence
>
Structured extraction. Action items, decisions, and commitments become queryable markdown.
>
Policy-bound people context. Build bounded policy-safe person profiles, topic research, relationship rankings, and commitments.
>
Cross-meeting search. Search policy-authorized history or ask your assistant to pull the thread for you.
>
Voice memo pipeline. iPhone recordings arrive on Mac and join the same policy-safe search corpus.
Integration
>
Desktop app. Tauri menu bar app with recording, dictation hotkey, and meeting prompts.
>
Claude-native. 34 MCP tools for Claude Desktop, Cowork, Dispatch, and Claude Code.
>
Any LLM. Use Ollama, OpenAI-compatible gateways, local servers, or skip summarization entirely.
>
Markdown is truth. YAML frontmatter, plain files, and a workflow that works outside Minutes.
07Comparison
How it compares
| Granola | Otter.ai | Hyprnote | minutes | |
|---|---|---|---|---|
| Local transcription | No (cloud) | No (cloud) | Yes | Yes |
| Open source | — | — | MIT | MIT |
| Free | Freemium | Freemium | Free | Free |
| Agent surface | Hosted MCP | Hosted integrations | Local app | Files + 34 MCP tools |
| Cross-meeting intelligence | Cloud chat | Cloud chat | — | Policy-safe search |
| Consent provenance | — | — | — | In every file |
| Dictation mode | — | — | — | Yes |
| Voice memos | — | — | — | iPhone pipeline |
| People memory | — | — | — | Bounded profiles |
| Data ownership | Their servers | Their servers | Local | Local |
| Data format | Cloud DB | Cloud DB | Local files | Markdown + YAML |
| Agent-agnostic | — | — | Partially | Yes |
08Governance
Built in, not retrofitted.
If you take notes on client conversations for a living — legal, clinical, financial — a cloud recorder isn't a preference, it's a compliance problem. Minutes keeps both the audio and the record on your own machine, and puts governance in the record itself: every file states the consent it was captured under, because its primary reader is now an agent.
Shipped
Every recording stamps its consent basis into the file's frontmatter. Sensitive meetings capture no audio but keep structured notes, and Require mode blocks every desktop and CLI entry point until consent is confirmed.
Next
Retention rules the corpus enforces on its own audio, and enforcement of the sensitivity contract across every agent surface, not just the debrief path.
The point
Sensitivity metadata your agents are required to respect: a restricted meeting is excluded from default agent search, bounded profiles, and anything an agent assembles.
A disclosure aid, not legal advice; make sure everyone present has agreed where required. The design is public: why we built it this way and the phase 2 plan.