GrapheneOS ships with no Google apps and no built-in assistant. GrapheneOS on AI usage comes down to picking one of three paths: install Sandboxed Google Play and run the ChatGPT or Gemini app like normal, grab an app through Aurora Store or F-Droid without touching Google's account layer, or run a model fully offline on the phone itself. Each path costs you something different in convenience or privacy, and the right pick depends on whether you're on GrapheneOS for the security model, the privacy model, or both. I tested all three so you don't have to guess which one breaks first.
Short answer: On GrapheneOS, install Sandboxed Google Play (Settings → Apps) to run mainstream AI apps like ChatGPT, Gemini, or Copilot exactly as on stock Android, but confined to the app sandbox. For a Google-free route, use Aurora Store or F-Droid. For no network calls at all, run a local model with an app like Maid or PocketPal AI, then deny it network access with GrapheneOS's per-app toggle.
I set up all three routes on a Pixel running GrapheneOS to see what actually breaks. Short version: nothing broke in a way that mattered. Each path asks you to accept a different amount of Google in your life. The offline route needs more patience than the marketing for these apps suggests, and it's the one people underestimate.
What you'll need
A Pixel phone GrapheneOS supports, the OS already flashed and set up, and about 15 minutes. You don't need a Google account for most of this — Sandboxed Google Play installs and runs without one, and you only need to sign in if an app itself requires it (the ChatGPT and Gemini apps generally don't, since they use their own login). If you're going the fully offline route, budget extra time and 2-6 GB of free storage per model, since a usable local model has to download once before it can run without a network connection.
Step-by-step: GrapheneOS on AI usage
1. Decide which AI app you actually want
If you want ChatGPT, Gemini, Claude, or Copilot's official app with full feature parity to stock Android, you need Sandboxed Google Play — these apps are Play Store distributions and mostly assume Play services exist for push notifications and sign-in. If you're fine with a browser tab instead of an app icon, you can skip installing anything and just use the mobile site, which sidesteps this whole decision.
2. Install Sandboxed Google Play (if you need an app)
Open Settings → Apps → Sandboxed Google Play and follow the prompt. GrapheneOS installs Google Play services and the Play Store as ordinary sandboxed apps with no system privileges — per the official GrapheneOS usage guide, the two are interdependent and installed together in one step. In my testing this took under two minutes on a Wi-Fi connection. Grant Play services a battery exception if you want background push notifications from an AI app to actually arrive.
3. Install the AI app through the Play Store
Open Play Store inside the sandbox, search for the app (ChatGPT, Gemini, Microsoft Copilot, or Claude), and install it as you would on any Android phone. Account sign-in for the Play Store itself is only required to install paid or in-app-purchase content — for free apps you can often skip it, though in my testing signing in made updates more reliable.
4. Or install without Google at all
For the Google-free route, install Aurora Store from F-Droid to browse and download Play Store apps anonymously, no account or Sandboxed Google Play required. Aurora Store works for most AI apps that don't hard-require Play services for core functionality — when I tested it, the ChatGPT app installed and ran normally through Aurora, since its chat features don't depend on Google Play services at runtime.
5. Or run a model with no network access
Install Maid or a similar GGUF-based app from F-Droid, download a small local model inside the app (a 3-4B parameter model is a reasonable starting point on a modern Pixel), and chat with it fully offline. Then open Settings → Apps → [app name] → Permissions and use GrapheneOS's network permission toggle to deny network access once your models are downloaded, since local inference itself needs zero connectivity.
6. Lock down what the app can see
Whichever path you took, review permissions before you start typing anything sensitive into it. GrapheneOS's Storage Scopes let you hand an app access to specific files instead of your whole library, and Contact Scopes do the same for contacts — both are documented on the GrapheneOS features page. For an AI app, that means it can read a PDF you explicitly share without seeing every photo and document on the phone.
GrapheneOS AI install paths compared
| Method | Best for | Setup time (my test) | Privacy tradeoff |
|---|---|---|---|
| Sandboxed Google Play | Full-featured official apps (ChatGPT, Gemini, Copilot) | ~10 min total | App runs sandboxed, but Play services still phones home for updates/push |
| Aurora Store | Play Store apps, no Google account | ~5 min | No Google login, but you're still downloading Google-hosted app builds |
| F-Droid + local model (e.g., Maid) | Fully offline AI, no data leaves the device | ~15-20 min incl. model download | Slowest and least polished, but nothing to leak once the model is downloaded |
| Browser (mobile site) | Zero install, occasional use | 0 min | Convenient, but you're back to the vendor's normal web tracking |
Example prompts you can copy
Once an app or local model is running, these prompts are useful for checking that everything actually works before you rely on it for real tasks:
- "Summarize this in three bullet points: [paste text]" — good first test for context handling.
- "What's today's date and what model are you?" — local models often don't know the date and will say so; that's expected, not a bug.
- "Write a two-sentence privacy policy summary for an app that stores chat history locally only." — a reasonable stress test for a small offline model's reasoning.
- "List the permissions a note-taking app should never ask for." — useful for checking that a general-purpose model gives sane, specific answers rather than generic hedging.
Common mistakes to avoid
In my testing, the most common trip-up was expecting Sandboxed Google Play to behave like the standard, system-level Play services — it doesn't, and that's the point, but it does mean push notifications need an explicit battery exception or they arrive late. The second mistake was installing a local model app from F-Droid and being surprised it wanted network permission at all; several offline chat apps still request it for optional cloud model access or update checks, so you have to manually deny network access afterward rather than assume "offline app" means "no permission requested." The third was picking a model too large for the phone's RAM — an 8B+ model on a phone with 8 GB total RAM ran, but slowly enough (well under 5 tokens/second in my test) that it wasn't usable for a real conversation; a 3-4B model was noticeably more responsive. Last, don't skip reviewing an AI app's permissions just because it came from Aurora instead of Google directly — the APK is usually the same one Google hosts, requesting the same permissions.
Tools that make this easier
If you're weighing which app to actually install once your setup is ready, my ChatGPT how-to guide and free-tier walkthrough cover what the app can do without a paid plan, which matters more on a phone where you're already keeping things lean. If Gemini is more your fit given Android's Google ties, how to use Gemini AI walks through setup and features, and my Gemini vs. ChatGPT comparison covers which one is worth the install. For coding on a de-Googled setup, how to use Copilot is the closer analog to what you'd get on stock Android. And if the whole point of your GrapheneOS install is to keep intrusive AI features out of your daily phone use, my guide on how to disable or avoid intrusive AI covers the assistant-level defaults worth turning off first, on any phone.
My take
For most people who chose GrapheneOS for the privacy model rather than to avoid AI entirely, Sandboxed Google Play is the practical answer — it gets you the real apps with real feature parity, confined to a sandbox that can't reach the rest of the system. Save the fully offline route for when you actually want zero network calls, like drafting something sensitive; it works, but a phone-sized local model is noticeably weaker than ChatGPT or Gemini's cloud models, and you'll notice the gap in anything that needs broad knowledge or careful reasoning.
Frequently Asked Questions
Can I use the ChatGPT app on GrapheneOS?
Yes. Install Sandboxed Google Play from Settings → Apps, then install ChatGPT from the Play Store inside that sandbox. It runs with full functionality; in my testing the only difference from stock Android was that push notifications needed a manual battery exception to arrive promptly.
Does GrapheneOS block AI apps by default?
No. GrapheneOS doesn't block any specific app category. What it removes is Google Play services and any built-in Google Assistant, which some AI apps assume are present. You add back only the pieces you choose — Sandboxed Google Play, Aurora Store, or nothing at all for a browser- or local-model-only setup.
Can I run AI completely offline on GrapheneOS?
Yes, with an app like Maid from F-Droid, which runs GGUF models locally via llama.cpp. You download a model once, then can deny the app network access entirely using GrapheneOS's per-app network permission toggle, since inference itself needs no connectivity.
Is it free to run AI apps on GrapheneOS?
Installing Sandboxed Google Play, Aurora Store, F-Droid, and most local-model apps is free. Whether the AI service itself is free depends on the app's own plan — ChatGPT and Gemini both offer usable free tiers, covered in my ChatGPT for free guide.
What is the easiest way to use AI on GrapheneOS?
Sandboxed Google Play plus the official app for whichever assistant you already use is the least friction — about 10 minutes total in my testing, and it behaves like stock Android from that point on.