__ __
/ <` '> \
( / @ @ \ )
\( \_-_/ )/
(\ `-/ \-' /)
"==/ _ \=="
.=') [_] (`=.
` .=' `=. 'A full AI gateway running 24/7 on any old phone. Not a stripped-down chatbot rewritten in Go or Rust — the real thing. OpenClaw, unmodified, with 30+ providers, Telegram bot, web dashboard, and REST API. Starting from a $20 Moto E2 with 1 GB of RAM.
57 hacks. The first phone is running. More devices, more hacks, and a lot more coming — follow on X and GitHub.
[LOG.001]
It started on a whim. February 10th, 2 AM. Robin had a Moto E2 sitting in a drawer — a phone from 2015 that nobody wanted. $20 on the used market. 1 GB of RAM. The kind of device people throw away without thinking twice.
The idea: what if this thing could run an AI? Not a toy chatbot. A real, always-on AI gateway — OpenClaw, an open-source platform that connects to 30+ AI providers (Gemini, Groq, OpenAI, Claude, Kimi, Ollama...) with a Telegram bot, a dashboard, and an API. The same software people run on $200/month servers.
No target. No deadline. No plan. Just “let's see how far this goes.” One phone to prove the concept. Then every phone.
[LOG.002]
OpenClaw needs Node.js 22. The phone ships with Node 12. OpenClaw needs ARM64. The phone is ARM32. OpenClaw needs Android 10+. The phone runs Android 6. OpenClaw needs 3 GB of RAM. The phone has 1 GB.
Every single requirement — failed.
The first attempt used proot (a Linux emulation layer inside Termux). It worked — technically. The gateway booted at 780 MB of RAM on a phone with 898 MB usable. Constant crashes. OOM kills every few minutes. Unusable.
Most people would have stopped here. “The hardware can't do it.” But the question wasn't whether the hardware was good enough. The question was: how much of that 780 MB is actually necessary?
[LOG.003]
What followed was a week of obsessive optimization. Every day, the RAM usage dropped. Every hack unlocked the next one.
From 780 MB to 385 MB. Same hardware. Same software. 57 modifications between “impossible” and “running in production.” That was the first phone. The next ones will be easier — and documented on GitHub as they happen.
[LOG.004]
This is not a chatbot rewritten from scratch in Go or Rust to be as small as possible. Those exist — they're impressive, but they're a different product. Minimal clients that talk to one API. A different trade-off.
PocketClaw runs OpenClaw — the full, unmodified, open-source AI gateway. Node.js 22. 30+ provider support. Conversation history. File handling. Tool use. Web search. Telegram bot. REST API. Web dashboard. Module system. The same thing you'd deploy on a VPS.
It started on a Moto E2. The mission is bigger: make it run on any old phone. The $30 Galaxy J2 in your drawer. The cracked Redmi 4A your kid stopped using. Any device people consider dead.
No server bill. No vendor lock-in. Powered by whatever free or paid AI provider you choose. Yours, on hardware you already own.
[SYS.001]
4-layer native execution stack. No emulation. No WebView. No Electron on-device.
[INT.001]
30+ verified external AI providers. OpenAI-compatible protocol. Free, paid, or local — your choice.
| Interface | Classification | Allocation | Context Window |
|---|---|---|---|
| Google Gemini | Free | 1,000 req/day | 1M tokens |
| Groq | Free | 500K tokens/day | 131K |
| Cerebras | Free | ~1M tokens/day | 128K |
| Mistral | Free | 1B tokens/month | 128K |
| OpenRouter | Free | :free models | Varies |
| Anthropic | Paid | Pay-per-token | 200K |
| OpenAI | Paid | Pay-per-token | 128K |
| Kimi K2.5 | Paid | ~$19/month | 262K |
| Ollama | Local | Unlimited | Depends on model |
[TERM.001]
[UI.001]
Two display units. Same CRT protocol. One operates on-device. The other provides remote management.
Native Android application. Pure Canvas rendering protocol. No XML layouts. No WebView runtime. 4-tab interface: STATUS, LOGS, KEYS, CTRL. Floating navigation element. Replaces default Android home screen.
Electron desktop application. Pip-Boy inspired display interface. 5 CRT color themes. Real-time status monitoring, log streaming, API key management, provider control, garbage collection trigger, gateway restart. Full remote operation.
This is Phase 1 of 5. One phone proven. The goal is every phone. New devices, new hacks, new features — everything shared live on X and GitHub. The story is just getting started.
Phase 2 is coming — Operator dashboards, early access, and a way to support the mission. Drop your email to be first in line.