WhisperShortcut

Voice-first AI for your Mac. Press a shortcut anywhere, speak, and the text appears at your cursor. Speak an instruction to rewrite what you selected, have text read aloud, record and summarize meetings, or ask an AI chat that knows your calendar and mail.

Bring your own API key (GPT, Gemini, Grok, Claude, OpenRouter or your own server) or run entirely on your Mac with no key at all. No account, no subscription, no backend of ours. Open source under AGPL-3.0.

▶️ Watch the full demo on YouTube · whispershortcut.com · User Guide · Made by Magnus Gödde

Download

⬇️ Signed .dmg from GitHub Releases (recommended)	Open it and drag WhisperShortcut to Applications. The full app: text is pasted at the cursor, the 🌐 Fn key starts dictation, Dictate Live, update notices.
Homebrew	brew tap mgsgde/tap && brew trust mgsgde/tap && brew install --cask whisper-shortcut installs the same signed build.
Mac App Store (free)	Sandboxed: results go to the clipboard instead of being pasted, and a few features need right-click → Services.

The welcome tour on first launch walks you through permissions, picking a provider key or the offline path, and a first dictation.

What it does

Dictation

• Dictate: speak, and the transcription is pasted at the cursor and copied to the clipboard. Names and jargon go in a Glossary so they come out right. Settings → Dictate offers one-click starter lists of German specialist terms (orthopedics, psychiatry/psychotherapy, expert reports).
• Dictate Live: text appears at the cursor after every short pause while you keep talking (on-device, Apple silicon).
• Offline dictation: Parakeet Ultra transcribes 25 European languages on your Mac in a fraction of a second; Whisper covers the rest.

Writing with your voice

• Dictate Prompt: select text and say "make this shorter" or "translate to English". With nothing selected, it writes what you ask for, in your own writing style if you turn that on.
• Templates: dictate into a fixed format you set up once, such as a clinical note with Anamnese, Befund and Therapie.
• Read Aloud: select any text and hear it, with cloud voices or on-device ones, from a small player with pause, skip and speed.
• Read Aloud Prompt: select text and say what you want to hear, such as "summarize this in two sentences" or "read it in English". The answer is read aloud instead of pasted. With nothing selected, it reads out the answer to your question.

Chat and meetings

• Chat: a chat window (⌥Space) with every supported model, screenshots, files, web search and YouTube videos. It works with Google Calendar, Tasks and Gmail, Trello, IMAP mailboxes and folders you share, including your AGENTS.md / CLAUDE.md context files.
• Live Meeting: records your meeting (with headphones too), transcribes it as it goes, writes live notes into the chat and answers questions about anything said so far.
• Claude, Claude Code and Codex over MCP: connect them with one click. They read your meetings, even the one running right now, and tune your dictation from your logs, one glossary term or rule at a time that you approve. Examples at whispershortcut.com/mcp.

Privacy

• Offline Mode: one switch that keeps everything on your Mac. Dictation, Dictate Prompt, Chat and Read Aloud run on-device, network requests the app builds are blocked, and nothing is written to the usage log. Built for regulated dictation such as patient findings.
• Keys live in your Keychain, data in the app's folder on your Mac. Usage statistics are opt-in and never contain text or audio. Details in privacy.md.

Every feature, setting and model is described in the User Guide.

Your own instructions

Dictate and Dictate Prompt have built-in instructions that always apply, because they define how the mode works: Dictate writes down what you said, Dictate Prompt applies what you say to the selected text, or writes it as finished text when nothing is selected. In Settings you add additional instructions on top, for you or for an AI app that tunes WhisperShortcut over MCP:

• Belongs there: who you are and your field ("I work in hospital IT"), tone per app ("Teams messages short and informal"), spelling and list style ("no em dashes").
• Does not belong there: a whole prompt. Don't start with "You are…" or "Your task…", and don't paste a dictation prompt into Dictate Prompt. The mode stays the same; only the result changes.
• Single names and terms go in the Glossary.
• A different output format for one job, such as a clinical note, goes in a Template. A template replaces the instructions for that one recording.

Settings, Chat and connected AI apps show a warning when the additions read like a whole prompt. Full prompts from earlier versions that replace the built-in instructions keep working and show a button to convert them.

Shortcuts

Action	Default shortcut
Dictate	⌃⌥1
Dictate Prompt	⌃⌥2
Template	not set (Settings → Templates)
Screenshot	⌃⌥3
Read Aloud	⌃⌥4
Read Aloud Prompt	⌃⌥⇧4
Chat	⌥Space
Settings	⌃⌥0

All of them can be changed in Settings → General; F1–F20 can be bound without a modifier, for a keyboard with a dictation key. In the direct-download build, pressing 🌐 Fn once starts dictation and pressing it again stops. Installs from before these defaults keep ⌘1–⌘4 and ⌘0. Dictate Live and Add Selection to Glossary have no default shortcut.

Models and keys

Provider	Used for
None (on-device)	Parakeet Ultra and Whisper dictation; Dictate Prompt and Chat on MLX models (Qwen3, Gemma); macOS and Qwen3 voices
OpenAI	GPT Transcribe (the default dictation model), GPT chat, Dictate Prompt, Read Aloud
Google Gemini	Dictation, Dictate Prompt, Chat, Read Aloud, meetings, YouTube videos
xAI	Grok Speech-to-Text, Grok chat with X search, Grok voices
Anthropic	Claude chat
OpenRouter	Any audio or chat model there; connect in one click, no key to copy
Your own server	Any OpenAI-compatible endpoint: LiteLLM, Azure OpenAI, Vertex AI
ChatGPT plan	Chat on your ChatGPT subscription, no API key (direct download)

Which dictation model to pick, with measured latency and cost: User Guide → Dictation.

Build from source

git clone https://github.com/mgsgde/whisper-shortcut.git
cd whisper-shortcut
bash install.sh

For development, bash scripts/rebuild-and-restart.sh builds Debug and restarts the app.

• Metal Toolchain: mlx-swift compiles Metal shaders, so Xcode needs the Metal Toolchain component (about 690 MB). Without it the build fails early with cannot execute tool 'metal'. Install it once with xcodebuild -downloadComponent MetalToolchain.
• Package plugin validation: mlx-swift ships a build plugin, so a plain xcodebuild needs -skipPackagePluginValidation. The scripts already pass it.
• Tests need macOS 26.5: the test target's deployment target is macOS 26.5, while the app runs on 15.5 and later. On an older macOS, xcodebuild build works but scripts/run-tests.sh does not.
• Requires Xcode 16.0 or later.

Useful scripts: scripts/logs.sh (filter app logs), scripts/run-tests.sh, scripts/create-release.sh, and scripts/test-{gemini,grok,openai}-models.sh to check provider model availability.

Project structure

• WhisperShortcut/: Swift source. Start with AppState.swift (state machine), MenuBarController.swift (orchestrator), SpeechService.swift (Dictate, Dictate Prompt, Read Aloud), ChatView.swift and ChatTools.swift (Chat and its tools), TranscriptionModels.swift, Settings/.
• WhisperShortcut/Docs/: docs bundled into the app so its Chat can answer questions about itself (this README and docs/guide.md are copied there on rebuild).
• server/telemetry/: the server that receives opt-in usage statistics.
• scripts/, plans/, .github/workflows/release.yml (signed, notarized releases).
• .agents/skills/ and .cursor/: rules and skills for coding agents.

Feedback and contributing

Found a bug or missing something? Open an issue, or use Send Feedback in the menu bar, which can attach your last dictation. Before you open a pull request, read CONTRIBUTING.md.

License

GNU Affero General Public License v3.0 (AGPL-3.0). See LICENSE.

Offline speech models are downloaded at runtime, not shipped in this repository: Parakeet Ultra by Moondream, a retraining of NVIDIA Parakeet TDT 0.6B v3, both CC BY 4.0, plus NVIDIA Parakeet CTC 110M for Glossary terms (CC BY 4.0), all downloaded as FluidInference's Core ML conversions and run through FluidAudio (Apache 2.0); Whisper models by OpenAI (MIT), run through WhisperKit (MIT). Qwen3 ASR 1.7B and the Qwen3 TTS CustomVoice 0.6B and VoiceDesign 1.7B models by Alibaba's Qwen team (Apache 2.0), downloaded as mlx-community's 8-bit MLX conversions and run through mlx-audio-swift (MIT).