Good to know.

Yes. Navi is free and open source under the GNU GPL v3, with no account, no subscription and no word limit.

AI is optional. Through Claude Code or Codex it counts against the Claude or ChatGPT plan you already have. An Anthropic or OpenAI API key is billed by them for what you use, and Ollama is free because it runs on your own computer. If you use Speak, ElevenLabs counts the characters it reads against your plan with them.

Windows 10 or 11 (64-bit), and macOS 14.2 or later.

On Windows, Navi uses Microsoft’s WebView2, which Windows 11 already has. On Windows 10 the installer adds it if it’s missing, which needs an internet connection. On a Mac, Apple silicon (M1 or later) runs the speech engine on its graphics chip. An Intel Mac works too, more slowly.

If you use Claude Code, copy the install prompt on the home page and paste it in. It downloads Navi on Windows, or builds it on a Mac, and opens it. Codex can do the same.

To do it yourself on Windows, run the installer from the latest release.

Navi for Mac is built on the Mac that runs it, by one script: scripts/mac-setup.sh in the repository. It checks for the tools it needs and offers to install any that are missing, builds the speech engine and Navi, and offers to copy Navi to Applications. An app you build yourself opens without macOS blocking it, and nothing in the script uses sudo.

macOS may ask for your login password once, to trust the certificate Navi is signed with, so it keeps its permissions from one version to the next. To update, run git pull in the Navi folder and the script again. The README’s Build on a Mac has the details.

That’s Windows SmartScreen. It warns about any new app from a publisher it doesn’t recognise yet, and Navi’s installer isn’t code-signed. Choose More info, then Run anyway.

Navi’s source code is public on GitHub, so anyone can check exactly what it does.

No. On Windows the installer sets Navi up for your account only, and the Mac set-up script doesn’t use sudo.

Navi checks GitHub 20 seconds after it starts and every six hours after that. If there’s a new version, a card offers it, and nothing downloads until you press Update now.

Updates are signed, and they never install in the middle of a take, while Navi is taking call notes or watching your screen. You can turn the checks off in Settings > About, and Local only stops them too.

No, but an NVIDIA card makes a big difference. With one, set-up downloads Whisper Large v3 Turbo (1.6 GB) and runs it on the card. Without one, Navi picks Whisper Base (English), about 150 MB, which runs well on the processor.

You can change the model in Settings > Speech engine any time.

No. Whisper turns your speech into text on your own machine, and your recordings stay there. There’s no cloud speech-to-text anywhere in Navi’s code.

If you switch on a cloud AI, it gets text only, never audio.

Only if you turn on Listen for voice commands, which is off to start with. Then the microphone stays open so Navi can hear “Navi”, and Windows shows its microphone icon. What it hears is checked on your computer and never kept.

Otherwise the microphone is only open while you hold the shortcut, during a hands-free take or a ramble, or while Navi is taking notes for a call.

Text only, never audio or pictures. With your words it gets a little context so the rewrite fits: the app’s name, the window title and some of the text before your cursor, plus your name, your style notes and any of your dictionary words that came up.

Password boxes are never read, and Navi only ever uses the name of apps on your Never read these apps list. Turn the context off in Settings > Context and privacy.

On your computer. On Windows, settings, history, recordings, screenshots and speech models are in %APPDATA%\com.goodlad.navi, notes go to Documents\Navi Notes, and keys are kept in Windows Credential Manager.

Recordings are deleted after 30 days unless you change that. Clearing History deletes your takes and their recordings, and your notes stay.

Yes. Turn on Local only in Settings > Context and privacy. Nothing goes to the internet except downloads you start and the Check for updates button. AI then works only with Ollama or a server on your own computer, and without one Navi still tidies your words.

No. Without AI, Navi still takes out the ums, fixes punctuation and capitals, applies your dictionary and spells things the British way, all on your computer.

If you pay for Claude or ChatGPT, Navi can borrow Claude Code or Codex with no key at all. An API key, a local Ollama model or your own server are there as alternatives.

The rewriting modes (Email, Message, Prompt and any you write), edits like “Navi, make that shorter” and “what’s wrong with this?”, notes from rambles and calls, jobs on calls, screen-watch write-ups and learning your writing style.

“Scratch that” and “change Tuesday to Wednesday” work without it.

Go to Settings > AI and press Sign in on its card. A window opens with the tool’s own sign-in, which carries on in your browser, and Navi notices when you’re done.

If Navi doesn’t see the tool at all, press Find it on the card. Navi searches your computer and connects it.

Clean mode doesn’t wait for AI, so it types as soon as Whisper has finished, which on an NVIDIA card is usually well under a second. A rewrite through Claude Code takes about 3 seconds, and about 4 or 5 through Codex. An API key or a local Ollama model is usually quicker.

Almost all of them. Windows won’t let Navi type into an app running as administrator, so there your words go on the clipboard for you to paste. For the rare app that blocks pasting, switch Settings > General > How text goes in to Type.

English to start with. Pick Detect automatically in Settings > Speech engine and the Large v3 Turbo models handle 99 languages. The Small and Base models understand English only.

Yes, in Settings > Shortcuts. It starts as Ctrl + Win on Windows and Right Option on a Mac. A single key such as Right Ctrl is the easiest to live with, and you can record any combination you like. The same page sets the ramble key, the cancel key and optional keys for each mode.

Hold your shortcut and say “Navi, scratch that”. To throw a take away before it’s typed, press Esc while you’re still talking. Every take is also kept in History, where you can copy it or re-run it in another mode.

Fix it in the text Navi typed. Navi notices, and after the same fix twice it adds the word to your dictionary. You can also add words yourself on the Dictionary page, with the ways Navi tends to mishear them.

Yes, under the GNU GPL v3: you can use, study, change and share it, as long as what you share is under the same licence with its source.

A changed version needs a different name, must say it’s based on Navi by GoodLad, and doesn’t get the Navi name, the logo or the firefly. The details are in the NOTICE file on GitHub.

GoodLad. Speech recognition is by whisper.cpp, running OpenAI’s Whisper models on your own computer.

Use the Report a bug page, or Request a feature for an idea. It opens a new issue on Navi’s GitHub page with your answers filled in; you need a free GitHub account to send it. On Windows, Navi’s log is in %LOCALAPPDATA%\com.goodlad.navi\logs, and it never contains what you said.