Realtime local dictation for macOS

Speak. It’s typed.
Nothing leaves your Mac.

Words appear as you talk — about a third of a second behind your voice, already punctuated. Release the key and they land in any app. And it’s the only dictation app your coding agent can talk to.

Star on GitHub Download for macOS

Free & open source · macOS 14+ · Apple Silicon · MIT

Speak. It’s typed. Nothing leaves your Mac.

hold to talk · release to insert

Realtime, not batch

Powered by Parakeet, our default engine: words stream in ~0.3 s behind your voice with punctuation, and the final text lands the instant you let go.

Never phones home

Transcription runs entirely on your machine and works offline. Your audio never leaves the Mac unless you explicitly turn on a cloud model.

Free, and yours

No price, no account, no subscription, no usage meter. It runs on hardware you already own, and the source is yours to read and change.

Agent Bridge · built-in MCP server

Give your coding agent a voice — yours

OpenWhisp doubles as a local MCP server for Claude Code, Cursor, and any MCP-aware agent. Instead of stopping mid-task to wait for typed input, your agent asks you out loud — you answer by voice and it keeps working. No other dictation app ships this.

openwhisp_dictate
Your agent pauses and asks you a question by voice. The overlay opens, you speak, and the moment you go quiet the transcript returns to the agent — no typing paragraphs into a chat box.
openwhisp_refineonly in OpenWhisp
Agents rewrite text with the same on-device model as your refine hotkey — your prompts, your machine, nothing sent anywhere.
openwhisp_history
Recent dictations — text, timestamp, target app — readable by the agent, stored only on your Mac.

# one-time setup — registers the server + agent instructions
openwhisp setup claude-code

# later, mid-task, your agent asks out loud:
● openwhisp_dictate("Deploy to staging or production?")
  ↳ overlay opens · you speak · silence sends it
● "production, and tag it v1.4"

# the CLI composes in pipelines, too
pbpaste | openwhisp refine -i "make it formal" | pbcopy

Works where built-in voice modes don’t. Any API key or Bedrock setup, over SSH, in regulated environments, in any editor — because it never phones home in the first place. It even picks up your project’s branch and file names so spoken code lands right, and a tap finishes your answer.

Off by default. Nothing listens until you enable it, and you approve each agent and each capability. It’s a private socket on your Mac — no open port, no cloud — and your hotkey always takes the mic back.

Also in the box

Everything else, still on your Mac

Writes like you meant it

Filler words gone, punctuation right, your jargon spelled correctly. Say “new paragraph” and get one.

Built-in offline AI

A small model runs on your Mac to polish text — no setup, no server. Dial it from a light touch to a full polish, and revert to your exact words anytime.

Meetings

Record a call, label who spoke (Me / Them), and get a Markdown summary — all locally.

Files & folders

Drop in audio or video, or point it at a watch folder. Exports text or SRT/VTT subtitles.

Modes & automation

Per-app tone — casual in Slack, formal in Mail, verbatim in your terminal. Drive dictation from Raycast or Alfred with the openwhisp:// URL scheme.

Rules & output targets

On finish: run a Shortcut, hit a webhook, append to a file. Fails open — your words always land.

Searchable history

Everything you’ve dictated, stored locally. Re-run any entry through a newer model.

Five engines

Parakeet by default. WhisperKit, whisper.cpp, Apple Speech, and Apple’s SpeechAnalyzer (macOS 26) are one click away — with ~40 languages auto-detected.

New · voice editing

Select text anywhere, then edit it by voice

OpenWhisp isn’t just for dictating new text. Highlight a sentence in any app, hold your dictation key, tap the Refine key — right ⌃ by default — and say what you want changed. Release, and it’s rewritten in place. No retyping, no copy-paste, no switching windows.

“Make it more formal” · “Translate to Russian” · “Tighten this up” — plain language, any language.
Your selection is read through the Accessibility API — and if an app doesn’t expose it, your clipboard is restored after the fallback copy. Secure and password fields are never read.
The rewrite can run on the built-in offline model or your own local server to stay fully private, or OpenAI if you prefer.

selected

hey so i can’t make the 3pm thing, can we push it

tap right ⌃“make it polite and professional”Refining…

replaced in place

Privacy by construction

The default is “stays on this Mac”

Audio is recorded locally and the recording is deleted after each transcription.
History, settings, and any API keys live on your machine — keys in the macOS Keychain, not in plain text.
AI cleanup can run fully on-device: the built-in offline model and your own local server both keep everything on your Mac. The only time text leaves your machine is if you choose the OpenAI provider.
Your transcript text is never written to the app’s log files.

your Mac

MicTranscribeRefine (AI)Typed text

↳ only if you pick OpenAIOpenAI cleanup

Open source · MIT

Read it, build it, change it

OpenWhisp is MIT-licensed and built from plain Swift scripts — no Xcode project required. Clone it, build the bundled runtimes, and run:

Star on GitHub All releases

The download is signed and notarized by Apple, so it opens with a double-click — no Gatekeeper warnings. Prefer to build it yourself? The full source is right here.

# clone with the whisper.cpp submodule
git clone --recursive \
  https://github.com/initcore0/openwhisp.git
cd openwhisp

# build the bundled whisper runtime
./scripts/build-whisper.sh

# compile, package, run
./build.sh && ./package.sh
open build/OpenWhisp.app

Questions, answered

Frequently asked questions

Does my audio ever leave my Mac?

No. Transcription runs entirely on your Mac and works offline. Audio is never uploaded, and the recording is deleted after each transcription. The only time any text leaves your machine is if you explicitly choose the optional OpenAI provider for AI cleanup.

How much does OpenWhisp cost?

Nothing. OpenWhisp is free and open source under the MIT license — no account, no subscription, no usage meter.

What are the system requirements?

An Apple Silicon Mac (M1 or newer) running macOS 14 Sonoma or later.

How is it different from the built-in macOS dictation?

OpenWhisp streams words in about 0.3 seconds behind your voice, already punctuated, and adds things macOS dictation doesn’t have: voice editing of selected text, AI refinement that runs offline, custom vocabulary, per-app modes, a searchable local history, and a built-in MCP server so AI coding agents can use your voice.

Can AI coding agents like Claude Code use OpenWhisp?

Yes. OpenWhisp doubles as a local MCP server, so Claude Code, Cursor, and any MCP-aware agent can ask you questions out loud and take your spoken answer, rewrite text with your on-device model, and read your recent dictations. It works with any API key or Bedrock setup, over SSH, and in regulated environments, because nothing is ever sent to the cloud. It is off until you enable it, and you approve each agent and each capability.

How do I install OpenWhisp?

Download the DMG, drag OpenWhisp to your Applications folder, and open it. The build is signed and notarized by Apple, so it launches with a double-click — no Gatekeeper prompts. You can also build it yourself from source.

Try it on your next sentence

Free, open source, and entirely yours.