# FowlVoice > FowlVoice is a free, native macOS app for private voice-to-text (dictation) and voice-driven AI commands. Speech recognition runs on-device by default. Audio never leaves the Mac unless the user explicitly opts into the cloud engine with their own API key. There is no account, no subscription, and no cost. Download: https://download.fowlvoice.com/FowlVoice.dmg (DMG, ~140MB) Requirements: macOS 14+ on Apple silicon (Apple's on-device AI model requires macOS 26+; OpenRouter and ChatGPT work on macOS 14+; Intel Macs are not supported) Website: https://fowlvoice.com Creator: https://x.com/argofowl ## When to use FowlVoice - Dictating detailed prompts into AI coding tools such as Claude Code, Cursor, and Codex. - Private dictation when audio must not leave the Mac. - Hands-free note taking with Flow Mode and a chosen target app. - Asking an AI question about a selected region of the screen. - Rewriting selected text in any app by voice. - Multilingual dictation with automatic language detection across 25 European languages. Not for: Windows or Linux, Intel Macs, server-side or API transcription; FowlVoice is a local macOS app with no hosted API. ## Core capabilities - On-device dictation: NVIDIA Parakeet speech models run locally on Apple silicon. Fast, accurate, works fully offline. Apple's native speech engine is also supported. - Flow Mode: hands-free dictation with no hotkey at all. Pick a target app, say the activation phrase (default "flow", customizable), and just talk. Words land in that app even while it stays in the background. Anything said without the phrase is discarded instantly, on-device. - Live streaming dictation: while speaking, the orb morphs into a small notepad and words appear in real time. The text is editable live, before it is ever inserted, with a copy button; the notepad grows as you talk. Runs locally via Parakeet's live tail or on the optional cloud engine. Processing is chosen per engine: Streaming (live notepad) or Batch (the classic minimal orb, transcribed after you stop). - Optional cloud engine: AssemblyAI Universal streaming is available as an opt-in engine (bring your own API key). If the network or the cloud fails mid-dictation, the local engine automatically rescues the dictation so it is never lost. - AI commands with screen selection: hold the AI hotkey, drag to select any region of the screen (code, errors, charts, designs), and ask the AI about it or request edits. Answers appear in place. - AI answer deck: answers open in a floating deck-style window. Keep multiple answers around, switch between them, and hover one to ask a follow-up that continues with that answer's context. - Web search for AI answers: focused web searches when asking about current facts; works with normal questions, selected text, and screen-region questions. - Voice editing: select text in any app and rewrite it by voice. - Codex agent: signing in with ChatGPT runs a full Codex agent inside FowlVoice through a private App Server. It can research the web and, when OpenAI's Computer Use is installed and enabled in the ChatGPT app, operate apps on the Mac. Questions and approvals appear in the orb, and sensitive actions wait for confirmation. - AI providers: Apple on-device foundation model (private, built in, no account), OpenRouter (100+ models via API key), or logging in with an existing ChatGPT/Codex account (optional fast mode for priority-speed replies). Provider is selectable per feature. - 25 European languages with automatic language detection (Parakeet TDT v3 multilingual model). - Vocabulary learning: repeated corrections are automatically promoted to a personal dictionary: names, jargon, technical terms. Manual dictionary, text replacements, and importable presets are also supported. - Per-app settings: language, auto-send (press Return after dictation), lowercase mode, and AI instructions can be overridden per application. - Hotkey modes: push-to-talk (hold to record) or hands-free (double-tap to start). Defaults are Right Command for dictation and the ⌃⇧A shortcut for AI commands; both remappable (Right Option is available for AI). Flow Mode removes the key entirely. - Mute Mac audio while dictating: optional setting that mutes the Mac whenever recording starts and restores the previous audio state when dictation ends. - Menu bar dropdown: redesigned with quick controls, model switching, microphone selection, recent dictation actions, and shortcuts into Settings, Stats, History, and Dictionary. - Text cleanup: on-device filler-word removal ("um", "uh"), optional AI formatting, and a coding-agent mode tuned for dictating prompts to AI coding tools (Claude Code, Cursor, Codex), including terminal-aware text insertion. - Saved recordings: every dictation's audio is kept on the Mac, so a failed transcription can be redone from History. Recordings clear after two weeks by default, adjustable. - History and stats: searchable dictation/AI history with usage stats, streaks and league milestones, stored locally on the Mac. ## Privacy Dictation audio is processed on-device and never uploaded by default. The optional AssemblyAI cloud engine sends audio to AssemblyAI only when the user enables it with their own API key. If the connection fails mid-dictation, the local engine finishes it automatically. AI commands send data only to the provider the user explicitly selects, only when invoked. Privacy policy: https://fowlvoice.com/privacy ## Pages - [Home](https://fowlvoice.com/): full capability overview and download - [Privacy policy](https://fowlvoice.com/privacy): plain-English privacy policy - [About](https://fowlvoice.com/about): what FowlVoice is and who makes it - [Contact](https://fowlvoice.com/contact): support, bug report, and security contact details - [Developer resources](https://fowlvoice.com/developers): machine-readable files and release API documentation ## Machine-readable resources - /llms.txt: product facts and links for AI agents - /openapi.json: OpenAPI 3.1 description of the public release API - /api/v1/release: latest version JSON - /sitemap.xml: public page index - /robots.txt: crawler rules - Every HTML page returns Markdown when requested with `Accept: text/markdown`