My AI Agent Costs $2/Month. Here's the Honest Breakdown.

I work at Cloudflare. I built a personal AI agent that actually does things - browses the web, fills forms, remembers everything. Here is what it cost and how I got there.

I’m a Customer Solutions Engineer at Cloudflare. I work with enterprise customers through the full cycle — pre-sales, POCs, onboarding, post-sales enablement. My job is to understand what a company actually needs, prove Cloudflare can do it, and make sure they get there. And for a while, I was using AI the same way most people do - typing into a chat box, copying the answer, closing the tab, and doing it all again tomorrow.

That bothered me. These tools are supposed to be intelligent. Why was I doing all the remembering?

So I built something different. I call him Agu.

What Agu actually does

Agu lives on my MacBook Air. I text him on Telegram - voice notes, images, whatever I have time to send - and he does things.

Not “answers questions” things. Actual things.

He browses websites and reports back. He fills out forms. He reads my email. He runs research in the background while I’m in meetings. He remembers that my car is a Toyota, that I prefer direct answers, that I’m working on building a side business - because I told him once, and he wrote it down, and now he knows.

This week I texted him at midnight asking him to register visitor parking for a car coming the next morning. He navigated to the building management portal, filled in the details, submitted the form, and sent me a screenshot confirming it was done.

That’s the version I wanted. That’s what took a few weeks to build.

The cost question

People assume this kind of thing costs a lot. It doesn’t have to.

My total monthly bill is $2-5. Some months it’s zero outside of the time I spend on it.

The secret is model selection. Most AI tools charge you premium rates for everything, even simple tasks. I built a hierarchy: the free stuff handles the day-to-day, and the powerful (expensive) models only get called when the problem actually needs them.

Google’s Gemini 2.5 Flash handles most of Agu’s conversations - it’s fast, genuinely capable, and free for up to 1,500 requests per day when you call it directly. I only reach for Claude when I need serious reasoning. That might happen a few times a week.

Everything else - background tasks, research, summarization - runs on free compute from Cloudflare’s platform.

The result: I have a personal AI agent running 24 hours a day, and I’m spending less per month than a single ChatGPT session would cost me in API calls.

What actually went wrong building this

I want to be honest about this because most write-ups aren’t.

I started with Docker. Docker broke. Specifically, the browser automation kept dropping connections mid-task with no error message - just silence. Agu would receive a task and nothing would come back. I spent more time debugging the container than building the agent.

I ended up moving to a proper macOS virtual machine running on my laptop. Real operating system, real network stack, real browser. Once I did that, the browser automation stopped being a problem.

I also had to rewrite how Agu interacts with web pages. The built-in tools weren’t reliable enough - certain types of forms wouldn’t respond, buttons wouldn’t click, and typing into fields produced nothing. I ended up going a level deeper, talking directly to the browser’s internals. That’s a technical fix but the point is: when something broke, I found the root cause and fixed it properly instead of working around it.

The memory problem

The hardest part of building a personal agent isn’t the AI itself. It’s memory.

Every session would reset. Agu had no idea what we talked about yesterday. Every conversation started from zero. You’d spend the first five minutes re-explaining context that should already be there.

The fix was building a knowledge base - a searchable store of everything Agu has learned about me, my projects, my preferences, my life. Before he responds to anything, he searches that store. After a meaningful conversation, he writes down what he learned.

Now he compounds. Each week he knows me a little better. He doesn’t forget.

That’s the difference between a chatbot and an agent.

Why I’m writing about this

I’m a Cloudflare SA. I work with technology infrastructure every day, and I have a pretty clear view of what’s actually possible with AI right now versus what’s being sold as possible.

Most people don’t have the time or background to figure out how these tools actually work - which ones are worth it, how to connect them, how to make them reliable. That gap is where I think I can be useful.

I’m starting to take on a small number of projects outside of my day job - helping businesses (starting with small ones, the kind where automation would genuinely change how much time the owner spends on repetitive work) set up AI systems that actually fit how they operate.

This blog is where I document what I’m learning. No polish, no hype - just what worked, what didn’t, and what it cost.

If any of this sounds like something your business needs, my contact is below.

Built with: OpenClaw · Cloudflare Workers AI · Google AI Studio · Telegram · macOS VM

Want to explore what this could look like for your business? Get in touch.