DeftNib makes two deliberate choices that most dictation apps make the other way: everything runs on your Mac, and you pay once. Two more follow from them: the app is native, and a good part of the cleanup is deterministic. Here are the four, in the order they were decided.

Because your voice is the most unguarded thing you type with

When you dictate you are not sending a finished sentence, you are sending a half formed idea with the hesitations still in it. Cloud dictation apps upload that to a server you do not control, together with the context of whatever you were writing. DeftNib cannot do any of that, not because it promises not to, but because it does not have the code to.

Because renting makes no sense here

A subscription earns its keep when a server is running on your behalf every month. Here there is none: the work is done by your own Mac, on an engine that already shipped with it. Paying monthly for software that runs on your hardware is paying twice.

Because an app should weigh what an app weighs

The category standard is a desktop app built on an embedded browser, with the resource use that implies. DeftNib is native, sits still until you press the key, and uses the system speech engine instead of downloading four gigabytes of its own model. When Apple improves that engine, you improve for free.

Because the same dictation should give the same text

A good part of the work is done by deterministic engines, not by a language model deciding fresh every time. Clean sentences never reach the model at all. That means no lottery: dictate the same thing twice and you get the same thing twice, which is the least you can ask of a tool for writing.

What that adds up to

  • Your audio and your text never leave the Mac. The app makes four kinds of calls to our servers, none carrying either, listed with what each carries on the private dictation page.
  • One payment of US$ 49 for 3 Macs, every 1.x version included, and a free plan of 4,000 words a week with every feature. What buying once means, exactly.
  • 9.7 MB on disk, under 30 mb of RAM at rest, 0.1% CPU while idle, 1 third-party dependency.
  • 15 rule engines above Apple's engine, and the local model only for what the rules cannot prove. The engine, compared with Whisper.

Mac with Apple Silicon and macOS 26 or later. Free up to 4,000 words a week, no account.

By downloading you agree to the Terms and the Privacy Policy.

Questions

Is any of this cheaper for you than the cloud?

Yes, and that is not the reason. A server that never sees your audio is a server that cannot leak it, and an app that runs on your hardware is an app that can be sold once. The cost follows from the architecture; the architecture was chosen for the other reasons.

What happens if Apple changes the engine?

The app follows macOS. It is built on the speech engine and the language model Apple ships with the system, so it inherits their improvements and, if they ever change shape, it is the app's job to adapt. That is the trade of building on the platform instead of shipping a model of your own.

Where is the code that would send my audio?

There is none. The app has no code path that can transmit audio or text; its four kinds of calls to our servers are about the license, updates and the free plan's counter, each listed with what it carries on the private dictation page.