Can Grux OS run entirely offline?

The model turn can, through Ollama. Features that are about the network, Research and Mailbox among them, still need the network, because that is what they do.

Does Grux OS work with Ollama?

Yes. Local Models reads your hardware and rates each model against it, and Chat routes to a local endpoint when you have one.

Where does Grux OS store my API keys?

In the macOS Keychain, under the service com.gruxai.grux. They are never written to disk in plaintext and go only to the provider they belong to.

How do I run an AI assistant locally on a Mac?

Grux OS 1.2.1 · last checked 2026-09-13 · generated from the shipping release

A local AI assistant on macOS means two separate things and most tools only do one of them. The first is where your data sits: Grux OS reads your mail, calendar, contacts and files directly from the accounts already on your Mac rather than from a hosted copy, and it has no server to copy them to. The second is where inference happens: point Grux at Ollama and the model runs on your own hardware, so no token leaves the machine and the running cost is $0. Use a hosted model instead and the first half still holds, but that provider sees what you send it. Meeting audio is the strict case. It is transcribed on device with WhisperKit and the recording never leaves the Mac at all, whichever model you chose.

The two halves, separately

What runs locally	With a hosted model	With Ollama
Your mail, calendar, contacts, notes, files	Local	Local
Meeting recording and transcription	Local, on device	Local, on device
Speaker separation	Local, on device	Local, on device
Focus log of your window titles	Local	Local
Shell commands and their snapshots	Local	Local
The model turn itself	Sent to your provider	Local

That table is the honest answer to the question. Most of Grux OS is local under either configuration, because there is no backend for it to be anything else. The row that changes is the model turn, and it is the one row people mean when they ask.

Running it with no key at all

Install Ollama, pull a model, and the Local Models surface picks it up automatically. Chat accepts either an Anthropic key or a local endpoint, so a clean install with Ollama running is a complete setup that has never had a key typed into it.

brew install ollama
ollama pull qwen3-coder:30b
open -a Grux        # Local Models finds the endpoint on its own

What you give up by going fully local

Research wants a Brave Search key, because searching the web means talking to the web.
Mailbox wants your IMAP details, which are your mail server's, not Grux's.
Compare exists specifically to run a prompt against a hosted model and a local one side by side, so it wants both.
Quality. A model that fits in your memory is a smaller model. Local Models tells you which ones actually fit before you download one to find out.

How to verify the claim rather than trust it

The tree is public and there is no analytics SDK, no crash reporter and no telemetry package in it. Everything the model reads from your disk is appended to a plain text audit log, refusals included, and you can read it at any time.

~/Library/Application Support/Grux/fs-audit.log

Questions

Can Grux OS run entirely offline?: The model turn can, through Ollama. Features that are about the network, Research and Mailbox among them, still need the network, because that is what they do.
Does Grux OS work with Ollama?: Yes. Local Models reads your hardware and rates each model against it, and Chat routes to a local endpoint when you have one.
Where does Grux OS store my API keys?: In the macOS Keychain, under the service com.gruxai.grux. They are never written to disk in plaintext and go only to the provider they belong to.

Download Grux OS 1.2.1 Read the source

Free, MIT licensed. macOS 14 or later, Apple silicon. 22.4 MB, signed and notarized by Apple. No account, no server, no subscription.