I’ve spent the last few weeks obsessing over one goal: having a personal, self maintaining AI assistant that costs $0and can be controlled from my phone.
It wasn't easy. I started with an AWS Ec2 with 50GB storage and t3.micro memory- minimal setup (using the free credits) and made Oracle Cloud instance ($300 free credits but just for a month so I used it for experimenting with local models) I was using Termius to SSH into everything from my phone
At first I used OpenClaw. It was cool, but I spent more time fixing it than actually using it. I almost gave up until I saw a video about Hermes Agent. And i actually found Hermes while looking for how to fix an OpenClaw error on YouTube (thanks NetworkChuck 🙌🏽)
He mentioned the exact same frustrations I was having, and that Hermes had been stable for a month. I didn't even finish the video before I pulled the repo.
The best part? It had a "migrate from OpenClaw" feature. I was up and running in minutes.
The hardest part is the rate limits. If you use cloud models especially for code, you hit a wall fast. My solution? The Fallback Chain.
Initially I was using openrouter/owl-alpha (stealth models are usually flagships in testing, like big-pickle is deepseek v4) which has 1M context window and was on multiple rankings.
Over time after I transitioned to Hermes, I wanted a bit more customization, while owl alpha was good at tasks,
It’s nothing to talk about on roleplay, it just scrapes the surface of the character I set in SOUL md file.
On my oracle instance I had been experimenting with local models (keep in mind, if you go local, you’ll be sacrificing speed but privacy. Ofc since the vms don’t have a gpu it would be slower, about 3-5 minutes for a simple response)
The one I was most impressed with is Google’s Gemma-4-31b-it
It played the role perfectly
Buuut if you know Google, you’re familiar with their aggressive rate limiting.
So I set up my agent to rotate through providers. I start with Gemma 4 for that perfect personality and roleplay via openrouter (add an ai studio api key in BYOK for longer usage). If that hits a limit, I’ve also set the same model via ollama cloud and using Google OAuth directly (basically Gemma 4 3 times lol)
And if those all hit limits, it jumps to Qwen3-coder-next (Alibaba, 1M free tokens per model. There’s like 80), then Nova (AWS bedrock), DeepSeek v4 (Azure and Opencode Zen), and Claude Haiku (GitHub). If everything fails, I have Owl Alpha; which is an absolute beast, took almost 70M tokens before I got rate limited once, that too for a few hours.
It lives in my Telegram and Discord. It manages my Spotify, handles my emails, and when I need real research done, I have it spawn three separate agents to work in parallel.
It’s been 8 days and it hasn't broken once. If you're looking to get AI without spending a fortune, I highly recommend looking into this
[link] [comments]