FnScribe brings open-source, offline Whisper dictation to the Mac menu bar

A menu-bar dictation tool that runs entirely on your Mac

FnScribe, a new open-source dictation utility from Algorithmic Research Group, was announced in a Show HN post that reached Hacker News's front page on 28 August 2026, linking to the project's GitHub repository. The app sits in the macOS menu bar: hold the fn/Globe key, speak, and release, and the transcription is typed into whichever app currently has focus. A hands-free mode for longer passages toggles with fn + Space, and a push-to-talk session can switch to hands-free mid-recording by tapping Space while still holding fn. Escape cancels a recording without inserting anything.

According to the project's GitHub page, FnScribe requires an Apple Silicon or 64-bit Intel Mac on macOS 13 Ventura or later, ships as separate disk images for both architectures, and currently transcribes English speech only. Recordings stop automatically after two minutes, and the menu-bar icon signals state: a microphone means ready, a square means recording, three dots mean transcription or insertion is under way, and an exclamation mark flags a permission or microphone problem.

Local Whisper, no servers

The privacy case rests on where recognition happens: on the device, using a Whisper model bundled with the app in whisper.cpp's GGML format. There is no account, no cloud service, no transcription history and no notes database; audio and transcripts exist only in memory and disappear when the app quits.

To place text into other applications, FnScribe writes a temporary, hidden clipboard entry and then restores whatever was on the clipboard beforehand, provided the user has not overwritten it in the meantime. The repository describes a "Copy Last Transcript" command as the only action that leaves a lasting clipboard entry, and it doubles as a fallback when automatic insertion fails. An "Original" option recovers the unprocessed text if smart cleanup altered the previous dictation.

Cleanup, corrections and a personal dictionary

Smart cleanup is enabled by default and strips common filler words — um, uh, erm — while interpreting spoken formatting commands such as comma, period, question mark, new line and new paragraph. Repeated phrases like "number one, number two" are turned into lists, "scratch that" discards the current clause, and corrections such as "actually make that" replace a previous value.

A personal dictionary pins exact spellings and capitalization, pairing a "write as" value with an optional "common mishearing" replacement — FnScribe versus "fn scribe", for instance. Entries are stored in a local settings file with owner-only permissions and are never transmitted, the project says. Settings also cover microphone selection and launching at login.

Permissions, installation and building from source

The app asks for two macOS permissions: microphone access to record while the shortcut is held, and Accessibility (keyboard) access to detect the shortcut and insert text into the active app. The project's documentation warns that macOS can stop recognising permissions if the app is moved, and recommends keeping it in Applications and re-granting access from System Settings.

FnScribe is free software under the GNU General Public License version 3. Building it requires Rust 1.85 or newer, CMake, Clang and the Tauri CLI; the repository provides scripts for native and Intel builds, a notarization path for Intel installers, and an environment variable for pointing development builds at a different GGML model. An opt-in regression harness exercises the production insertion path across native apps, terminals, browsers and Electron.

Why it matters

Dictation is routine on modern operating systems, but for anyone handling confidential material — legal, medical, journalistic — sending audio to someone else's servers is often unacceptable, and closed tools make their data handling hard to verify. FnScribe offers the alternative in auditable form: a GPL-licensed, menu-bar push-to-talk workflow built on open-source Whisper models, with no accounts, no servers and no stored history. It is also a useful reference for developers building local-first Mac utilities, since its clipboard-based insertion approach and permission handling document exactly the compromises that come with refusing a cloud dependency. The English-only support and two-minute recording cap mark it as an early-stage tool, but the architecture is the story: dictation that can be trusted because nothing leaves the machine.