Hold a key. Speak. Release.
Clean text lands wherever your cursor is.

Hark is dictation and meeting transcription that runs entirely on your Mac. Speech recognition on the Neural Engine, speaker separation and search in one local database, and nothing sent anywhere. Not to us, because there is no us to send it to. 

brew install --cask tjameswilliams/tap/hark

Free and open source. macOS 15 or later on Apple Silicon. Or download the .dmg.

Download for Mac Read the source

See it work

Fourteen seconds, start to finish. Hold the key, say what you want, let go. The text lands in the agent's prompt and it gets to work.
Dictation

Hold right ⌘ and talk

Release, and the transcript is pasted into whatever field is focused, in any app. Filler words and punctuation are tidied by a cleanup model you choose, and if that model is slow or down, the raw transcript pastes instead. It never fails closed.

Meetings

Record any call, no bot in the room

Hark captures the meeting's audio straight from Core Audio, whichever app is running it, then separates the speakers and transcribes it on your Mac. The recording never goes to a cloud service and nobody joins your call.

Knowledge

Everything you said, searchable

Dictations and meetings are embedded and grouped into projects. Search them by keyword or by meaning, or ask a question and get an answer with the transcript to back it up.

Neural Engine

Fast because it is local

NVIDIA's Parakeet model runs on the Apple Neural Engine through Core ML. A ten-second utterance transcribes in about a tenth of a second, on battery, with the fans off.

Dictionary

Teach it your names

Add the people, products and jargon it keeps misspelling. Terms feed speech recognition, the cleanup model and a final spelling pass, so they come out right every time.

Your data

One SQLite file

Every transcript lives in a single database in your Library folder. Back it up, query it, delete it. There is no account and no sync, because there is no server.

Zero to dictating

From nothing installed to text on the screen, in three steps.

  1. 1

    Install Hark

    brew install --cask tjameswilliams/tap/hark

    Or download the .dmg and drag it to Applications. Signed and notarized by Apple.

  2. 2

    Grant two permissions

    Microphone, to hear you, and Accessibility, to paste where your cursor is. Hark asks for both on first launch and opens the right System Settings pane for each. Meeting capture asks for audio recording the first time you use it.

    ⚙︎
  3. 3

    Hold the key

    Right ⌘ by default; right ⌥ or right ⌃ if you prefer. A rufous pill appears while Hark is listening. Let go and the text arrives. Taps shorter than a third of a second are ignored, so the key still works in shortcuts.

    ⌘

For your AI tools

Hark ships a small MCP server, hark-mcp, that lets Claude Code, OpenCode and Codex search and read your meetings and dictations. It opens the database read-only and runs only while the tool is asking. Register it once per client from the Settings window, or from the terminal:

Claude Code

hark-mcp install --client claude-code

OpenCode

hark-mcp install --client opencode

Codex

hark-mcp install --client codex

Tools: search_knowledge, list_projects, list_sessions, get_transcript, recent_meetings. Read-only by construction.

What the subscriptions charge for

The paid dictation and meeting tools are good. Most of them are also a monthly fee for sending your voice to someone else's computer. On the things that matter, Hark matches them or wins; where it doesn't yet, the table says so.

Hark Wispr Flow superwhisper Granola
Price Free. Open source. $15 per user per month ($12 billed yearly). Free tier capped at 2,000 words a week on desktop. $849 lifetime is the headline price, or a monthly plan. Free tier keeps the small local models after 3,000 words. Free until you want notes older than 30 days, then $14 per user per month.
Where your speech is transcribed On your Mac, always. Neural Engine. In the cloud, always. Their privacy page says so in those words. On your Mac on Apple Silicon. Cloud models optional. Their servers (AWS, in the US).
Where your transcripts live One SQLite file in your Library folder. Wispr's cloud. Dictation storage is a setting; meeting transcripts are stored there for syncing. On your device. Their servers.
Account required No. Yes. Yes, for a licence. Yes.
Meetings recorded without a bot joining Yes. Core Audio tap on the Mac. Yes, on Mac. Yes. Yes.
Your AI coding agents can read your notes Yes. Local MCP server, read-only. None listed. None listed. On the Business plan, through their cloud.
Source you can read Yes. MIT or Apache-2.0. No. No. No.
iPhone app Not yet. Yes. Yes. Yes.
Windows No. Hark is built on Apple's Neural Engine and stays a Mac app. Yes. Yes. Yes.

Read from each vendor's own pricing and privacy pages on 8 September 2026: Wispr Flow (privacy), superwhisper, and Granola (privacy). They change their plans; if a cell is out of date, tell us and it gets fixed. Hark's speech model is NVIDIA Parakeet, the same family superwhisper offers in its paid tier.

What stays on your Mac

Everything, with one exception you control.

  • Audio is transcribed on the Neural Engine and discarded. Meeting recordings are kept only if processing fails, so you can retry.
  • Transcripts, embeddings and speaker profiles live in one SQLite file under your Library folder.
  • The exception: the cleanup step sends each transcript's text to an OpenAI-compatible endpoint of your choosing. Point it at LM Studio or Ollama and nothing leaves the machine. Point it at a cloud API and that text does. It is off until you turn it on.
  • No telemetry, no crash reporting, no accounts, no update pings. The site you are reading is the only server involved, and it only serves the download.

The full privacy page says the same thing at more length.