No PC. No cloud IDE. Just Android, Termux, and a lot of stubbornness.

RenanBP — Thu, 05 Mar 2026 15:08:18 +0000

My name is Renan. I'm not a professional developer — I know the basics of Python and Linux, and I'm completely obsessed with technology, automation, and AI. That combination turned out to be enough to build something I'm genuinely proud of.

This is the origin story of ClawLite — a Linux/Termux-first AI agent runtime built for operational use, not just interactive chat.

It Started With a Problem: No PC

A few months ago I was away from home and had no access to my computer. I'd been following the AI agent space closely, and I'd recently discovered OpenClaw — it blew my mind. I needed to understand how it worked from the inside.

The question was: could I run it on my phone?

I started digging on GitHub and found openclaw-termux by @mithun50. Huge credit to him — without that repo, none of this would have started. Installed it, and against all odds, it worked. It doesn't install natively into Termux; it runs Ubuntu through proot-distro, which gives you a real Linux environment on ARM64 Android.

So there I was: editing code on a phone screen, inside a terminal emulator, inside a fake Ubuntu, inside Android. Perfectly normal.

The Idea: Port It to Python, Make It Compact

I also discovered nanobot by @HKUDS around the same time — a lightweight, compact bot runtime. Big thanks to that team too. That's when the vision clicked: I wanted something like OpenClaw, but leaner, more portable, and fully understandable from the inside out.

I asked OpenClaw (running on Codex 5.3) to analyze its own system and port it to Python. It did — over and over — but the output was a mess. Code without structure. Files without purpose.

My mistake early on was trying to build everything at once: the runtime, a website, documentation, a skills marketplace. Classic scope creep. I should have shipped the core first.

Rebuilding With Better Tools

I started using claude.ai to craft better prompts for the AI coding sessions. That made a real difference in output quality — the difference between "it generated code" and "it generated code that actually does what I asked."

Then I switched from OpenClaw to OpenAI Codex for the heavy lifting. Codex helped clean up the architecture: took what worked, threw out what didn't, reorganized everything into a proper structure. Progress.

But bugs. So many bugs. I'd hit a wall — Codex would try, retry, and go in circles. That's when I found Kilo.ai. I wasn't expecting much. Then I saw what it could do running Codex 5.3 in high-power mode. I've been using it ever since, and it's been excellent.

All of this, still on my phone. Still in Termux. Still ARM64.

What ClawLite Actually Is

After all those iterations, ClawLite became something with a real identity:

HTTP + WebSocket gateway as a control plane
Telegram-first delivery with retry/backoff, deduplication, polling and webhook support
Cron scheduler with lease and idempotency semantics
Persistent memory runtime with monitoring
Provider failover with a registry-based reliability layer
Tools and skills registry for extensibility
Runs entirely from a single clawlite CLI command

It's designed for headless, continuous operation — not just "chat with an AI." You start it, it runs, it handles messages, executes scheduled tasks, and persists state. On a server, a Raspberry Pi, or an Android phone.

What I Learned

You don't need to be a senior engineer to build something real. I used AI as a collaborator, not a shortcut. Every output had to be understood, reviewed, and steered. The AI wrote the code; I held the vision.

Start with the core. Ship everything else later. Trying to build the runtime, the docs site, and the skills marketplace in parallel nearly killed the project.

Constraints are features. Building on ARM64 Termux forced simplicity. If it didn't run on the phone, it didn't make the cut.

What's Next

ClawLite is shipping now. The architecture is solid. I'm going to keep improving it — better docs, more reliability, eventually a proper dashboard.

If any of this resonates with you — if you've ever built something sideways, with the wrong tools, in the wrong place, and made it work anyway — I'd love for you to check it out.

Contributions, feedback, and stars are all welcome. 😀

🔗 github.com/eobarretooo/ClawLite

Thanks to @mithun50, @HKUDS, Anthropic, and everyone whose open source work made this possible.