An offline AI writing assistant for Mac: what actually works without internet

You're on a plane, behind a locked-down corporate VPN, or you simply don't want your drafts on someone's server — and you'd still like AI help while you write. Here's the honest map of what works with no internet, and how to set up a writing assistant that shrugs at airplane mode.

The uncomfortable truth: most "AI writing tools" are cloud tools

Type "AI writing assistant" into a search engine and nearly everything you'll find — ChatGPT, Grammarly's AI features, Notion AI, Jasper, Copy.ai — sends your text to a server to be processed. Unplug the network cable and they're paperweights. That's not an accusation; big models need big hardware. But it means "offline AI writing" is a real requirement those tools structurally can't meet, whatever their settings say. If the intelligence lives in a data center, your draft has to travel there.

So the first filter is simple: an offline writing assistant must run its language model on your own machine. On Apple Silicon Macs, that's now entirely realistic — small models run fast on the M-series chips.

Local chatbots exist — but chat isn't writing assistance

The best-known way to run a model locally is a chatbot app: Ollama, LM Studio and friends let you download open models and converse with them fully offline. They're excellent at what they do, and if you want a local chat window, start there.

But notice the workflow: you leave your email, open the chat app, paste your draft, explain what you want, wait, copy the answer, paste it back. That's a consultation, not assistance. A writing assistant should work the other way around — where you type, while you type: suggesting the rest of the sentence in Mail, catching the typo in Slack, translating the message you're composing. The model being local is necessary; the model meeting you at your cursor is the point.

A truly offline writing assistant: the setup

Souffleuse is a menu-bar app for macOS built exactly for this: a small language model (Gemma 1B class, via llama.cpp and Metal) running on your Apple Silicon chip, assisting you inside almost every app through the Accessibility API. It needs the network exactly once:

1. Download Souffleuse (~7.5 MB, notarized by Apple, requires a Studio license) and drag it into Applications.
2. Grant the Accessibility permission, then let it download its model — 0.4 to 1.5 GB depending on the model you choose. This is the only connection the app ever makes. Do it at home or at the office; after that, the network is optional forever.
3. Write anywhere — Mail, Notes, Slack, your browser. The continuation of your sentence appears in grey; Tab accepts it, Esc clears it. This is system-wide autocomplete, airplane-mode-proof.
4. The other roles work offline too: translation as you type (pick a target language once per conversation), tone rewriting with //, and typo fixes in passing.

Two honest notes. Souffleuse is French-made and its interface is in French for now; the suggestions themselves follow whatever language you type, and English works well. And it's Apple Silicon only (M1 or later, macOS Sonoma 14+) — no Intel Macs.

Who actually needs offline AI writing

• Privacy-minded writers. With no server in the loop, there is nothing to leak, store or train on. Souffleuse also keeps no account and sends no telemetry; the history it uses to learn your style is encrypted on your Mac, readable by you alone. More on the model in our local LLM guide.
• Compliance-bound teams. Legal, healthcare, finance: when policy says client text must not reach third-party processors, on-device is the setting that makes the question moot — no subcontractor, no data transfer.
• Travelers. Planes, trains, hotel Wi-Fi that hates you: your assistant works identically everywhere.
• VPN-locked and air-gapped machines. Corporate Macs that can't reach consumer AI endpoints can still run a local model that never asks for one.

Frequently asked questions

Can an AI writing assistant really work with no internet?

Yes, if the language model runs on your own machine. Souffleuse downloads its model once at setup (0.4–1.5 GB) and never connects again: autocomplete, translation and tone rewriting all run on the Apple Silicon chip, in airplane mode if you like.

Aren't Ollama or LM Studio already offline AI writing tools?

They run models offline, but they are chat interfaces: you go to them, paste text, and copy the answer back. A writing assistant works the other way around — it helps where you type, suggesting the rest of your sentence in Mail, Slack or Notes as you write.

Is offline also better for privacy and compliance?

By construction. With no server involved, drafts cannot leak, be stored by a third party, or be used for training. Souffleuse adds no account and no telemetry, and macOS secure fields (passwords, card numbers) are invisible to it.