local · offline

Local & offline autocomplete for your Mac

MacAutocomplete runs a real language model directly on your Apple Silicon. No internet, no server, no monthly bill. On a plane, in a tunnel, or with Wi-Fi switched off, it keeps finishing your sentences exactly the same — because the model lives on your Mac, not in someone's datacenter.

How local inference actually works

On first launch, MacAutocomplete downloads a single local model — IBM Granite 4.1 3B, quantized to 4-bit — about 2GB, once. From then on it runs through Apple's MLXframework, the same machine-learning runtime built for Apple Silicon. When you pause mid-sentence, the text around your caret is read, handed to the model on your Mac's GPU and Neural Engine, and a completion comes back as greyed ghost text right where you're typing. Nothing is bundled into a cloud request, queued behind other users, or shipped off-device. The weights sit on your disk; inference happens inside the app's own process.

Works with zero internet — airplane mode and all

Because every suggestion is generated locally, there is no connection to lose. Put your Mac in Airplane Mode, board a flight, drop into a subway tunnel, or work from a cabin with no signal — the ghost completions keep appearing and ⇥ Tabstill accepts them. Try it: turn off Wi-Fi right now and start typing. Most "AI writing" tools go dark the moment you disconnect. This one doesn't notice. It's the same on-device privacy guarantee, viewed from the other side: if nothing ever leaves your Mac, nothing ever needs a network either.

No round trip means it feels instant

Cloud writing tools have to send your text to an API, wait in line behind everyone else's requests, generate, and stream the answer back over the network. Even on a fast connection that's tens to hundreds of milliseconds of latency you can feel as you type. Local inference skips all of it. There's no DNS lookup, no TLS handshake, no server queue — just the model on your own silicon. Completions surface at the caret fast enough to keep up with your typing, and they don't stutter when your connection does.

Costs nothing to run — no subscription, ever

Once the model is on your Mac, running it is free. Every completion is computed on hardware you already own, so there are no token charges, no API bills, and no usage meter ticking in the background. There's no monthly fee to keep it working and no rate limit to throttle a heavy writing day. You pay once and own it.

Curious how that lines up against a specific competitor? See MacAutocomplete vs. Cotypist, or read the longer case for keeping it on-device in local vs. cloud AI writing.

Frequently asked questions