I built a cyberdeck from a 45-euro phone — no root

Introduction

I built a cyberdeck, from a used Google Pixel 6a that cost 45 euros. No root, no expensive setup, no server in the basement. And yet it runs a mobile AI agent that operates multiple agents, executes scheduled tasks, manages a long-term memory, and even survives when Android tries to kill it.

That sounds like a success story. But the road here was anything but straight. It didn't start with a plan, but with an old Raspberry Pi, a vague idea, and the question of whether you can really build something like this without spending much money. In this post I tell the honest story from the beginning, including the costs, the failures, and the limits I discovered along the way.

The Beginning: A Raspberry Pi that never became a cyberdeck

The idea of a cyberdeck isn't new. It sits somewhere between maker dream and science fiction: a compact, portable device you can flip open, type on, and work with anywhere. Many people build one from a single-board computer, a display, and a keyboard. In the maker scene, the Raspberry Pi is the de-facto standard for that.

I had a Raspberry Pi 2 lying around for years. I kept picking it up, playing with it, and eventually bought a touch display for it. It was exciting to work with the thing. And in the back of my mind I always had the same thought: maybe I'll attach a keyboard and turn it into some kind of cyberdeck.

That was the original idea. It stayed an idea for a long time, because a cyberdeck made from individual parts gets expensive quickly, and because I didn't know if the effort was worth it. A finished Raspberry Pi cyberdeck with display, keyboard, battery, case, and cooling quickly costs between 300 and 500 US dollars. That was too much for an experiment I wasn't sure would be more than a toy.

The Turning Point: Try it, but without spending much

At some point I reached the point where I just wanted to try it. But I didn't want to sink a lot of money into a project whose outcome was completely open. The Raspberry Pi path was out, the parts, the case, the keyboard, it all adds up fast.

So I searched Google to see if you could do something like this with an old phone instead. The search turned up a few devices, including the Google Pixel. A used Pixel 6a was available for around 45 euros. That was a price I could live with. A complete cyberdeck from a used Pixel, a keyboard, a right-angle adapter, and a 3D-printed case comes to around 78 euros, a fraction of what a Raspberry Pi cyberdeck costs.

The First Idea: Ubuntu on the phone

My idea was simple at first: I wanted to put Ubuntu as an operating system on the smartphone and turn it into a small Linux computer. A compact cyberdeck that fits in your pocket.

It didn't take long for that idea to hit reality. It doesn't just work. You don't get real root access on a normal Android device. And without root, many things aren't possible, certain terminal commands simply don't run because the system doesn't allow them. Installing a full Ubuntu on the Pixel is practically impossible: the Google Tensor chip has almost no custom-ROM or Linux support, and even the experimental mainline kernel brings many limitations.

That could have been the end of the story. Instead, it was the beginning.

The Gap: A cyberdeck without root

Most people skip Android without root as a platform for an agent, because "no root" is seen as a dealbreaker. If you have a powerful PC, you don't need this workaround. But I had almost nothing, just a cheap phone and the curiosity to see if it works anyway.

That's exactly where the real project lies. Not "I built an expensive device," but "I found out it barely costs anything and can still do more." The unconventional, undocumented path was the only one that worked.

What It Became: A mobile AI agent

Out of the idea came a system that now runs permanently. The foundation is Termux, a Linux environment for Android, in which a PRoot-Ubuntu container runs. Inside it, a Hermes agent works, a gateway waits in tmux for scheduled tasks, and a watchdog monitors the heartbeat.

The thinking happens in the cloud, the orchestration runs locally on the device. No root, no expensive setup, and still an agent that improves itself, runs in the background, and is remotely controllable from anywhere. The architecture is deliberately lean: Termux as the base, PRoot-Ubuntu as the robust environment, tmux for the sessions, and a watchdog that restarts the gateway in about two seconds after a crash.

The Honest Costs

An important point I don't want to hide: the setup is cheap, but not free. There are two levels:

**One-time hardware costs:**
- Used Google Pixel 6a: 45 euros
- Rii keyboard: 22.90 euros
- USB-C right-angle adapter: 5 euros
- 3D-printed case: around 3 euros
- Power: around 2 euros
- **Total: around 78 euros**

**Monthly running costs:**
- ChatGPT subscription (for the coding agent): 22.90 euros
- Ollama subscription (cloud inference for Hermes): 20 euros
- Aldi-Talk SIM for a fixed internet connection: around 10 euros
- **Total: around 53 euros per month**

These numbers are honest. "Free" refers to many software components and the ability to use existing hardware, not to every possible use. Anyone who wants to rebuild this setup should know these costs.

The Honest Stability Numbers

- **Record:** 88 hours without interruption and without an OOM kill (Aug 28 to Sep 1, 2026)
- **OOM kills since hardening:** 0
- **Cron jobs:** 8 active, 0 failed
- **Verified projects:** Network Obfuscation Layer (29/29 tests) + Preiskompass (7/7 tests) = 36/36 tests

The honest limit: Android kills background processes under memory pressure, always. That's a fundamental kernel mechanism that no setting disables, not even "Unrestricted" or Wake Lock. On August 28, the gateway was killed by SIGKILL at 18:44, even though Termux was set to "Unrestricted" and running with Wake Lock. The system is therefore built to survive kills, not to prevent them. Since the hardening, memory pressure has stayed low, hence zero kills.