10 Apps in a Month With AI: Code Got Cheap, Judgment Didn't

In this article you will learn:

• what I built for myself in one month with Claude Code, Codex and a Raspberry Pi from 2016
• how much it cost - in money and in hours
• where my technical background still made the difference, with real examples

Where have I been?

My last post went out on August 30. That is two months of silence, so I owe you an explanation.

When Opus 5.5 came out, I quickly became addicted to it. It shortened the distance from the idea to the implementation massively. So I started building. A lot 😅

Quick answer

• 10 projects in one month. 7 of them run 24/7 on a Raspberry Pi 3B with 1 GB of RAM.
• More than 500 commits. The fastest app went from the first prompt to running on the Pi in 51 minutes.
• $20 a month at the start, $140 now. The same usage at API prices: about $2,300.
• Code got cheap. Judgment didn't. The AI wrote nearly all of the code. The decisions that kept my data private, my bills low and my numbers correct still came from knowing how things work under the hood.

The setup

Nothing exotic here:

• PC - the workshop. Every project is a repository here, and this is where Claude Code runs.
• Raspberry Pi 3 Model B - the home of the apps. Each one is a small Bun server running as a systemd service. Everything scheduled runs on systemd timers. One command deploys an app from the PC to the Pi.
• Phone - the screen. Every app is installed on my Android as a PWA and reachable through Tailscale, so only my own devices can open it. Nothing is exposed to the internet.
• Raspberry Pi Zero W - a tiny 2-inch display next to my monitor. More on that one below.

Here is how they talk to each other:

Does an old Pi struggle with seven apps? Not at all. Together they use about 280 MB of RAM.

The AI part is simple too: Claude Code (Opus 5.5) builds, Codex reviews.

What I built

Let me show you the ones that surprised me the most. Wherever a screenshot would show my personal data (car, money, portfolio), the data is made up.

Fin Stats

The only public one. Stocks, treasury bonds, retirement accounts (IKE, IKZE, PPK) and savings on one calm screen. Everything is computed in your browser, and if you want to keep your data, you keep a copy encrypted with your password.

Reports, RPI Manager and Agent Jobs

Every weekday at 8:30, before the Warsaw Stock Exchange opens, the Pi writes me a report: news about the companies I hold, the economy, oil prices - with a link to every source. It costs about 1% of a 5-hour window on my subscription.

Speaking of windows: AI subscriptions have a 5-hour limit and a weekly one, and whatever you don't use is gone. RPI Manager shows me how much is left - on the phone, on the PC and on a 2-inch display next to my monitor.

And the spare limits? Every 30 minutes Agent Jobs checks for room. If there is some, Codex reviews one of my repositories, Claude checks each finding and fixes the confirmed ones. In its first four days, this loop ran 124 fix sessions - mostly on its own. Some of the bugs it found were in its own earlier fixes 😅

Car Stats, FPL Scout, House Builder

Car Stats replaced an Excel sheet with my car's fill-ups. I take a photo of the dashboard and the pump, a model on the Pi reads the numbers, and I tap Accept.

FPL Scout helps me with Fantasy Premier League. Before each deadline, several models read the news about my players, and the app shows where they agree.

House Builder is a floor-plan editor for designing my future house, with a 3D view and Polish building rules. The projects live on the Pi, so I can work from my phone - which, funnily enough, has a better GPU than my PC. (The house below is a demo, not mine! And yes, the floor plan is complaining: the dining room's windows are too small for Polish regulations 😅)

How fast did they come together?

Career Stats

This one is different. It is a map of what I know, built into the way I work. A skill in Claude Code judges my knowledge from what I say, build and answer in VS Code, updates the map and generates what I am missing: explanations, quizzes and flashcards. The flashcards land on my phone, so I can learn in a free moment.

Finfluencer Effect

I once heard that finance YouTubers move stock prices, so one evening I asked Claude to check it. It analysed 2,533 videos from the biggest channels in the US and in Poland. The short version: a small effect that doesn't grow after the first session. In the US it sits mostly on news days; in Poland it holds on quiet days too. The code and the results are public (opens in a new tab).

Party Flags

On October 29, at 21:35, I asked for an app that prints bunting for a party my wife is co-organizing. At 22:30 it was done: 12 flag shapes, any Google font, one A4 page per letter.

Do you see what changed? The question is no longer "is this tool worth building?". If I need it for one evening, I build it for one evening.

How cheap is it, really?

I started at $20 a month: one Claude Pro plan. On October 20 I added ChatGPT Plus for Codex (the first month was free), and a day later a second Pro plan, because one was no longer enough. Five days later two were not enough either, so I switched one of them to Max ($100). Now it is $140 a month.

RPI Manager also prices every request at API rates. Since September 25 it counted about $2,300 of usage (it is on the second screenshot of RPI Manager above). Without prompt caching it would be about $22,000 - keep that in mind for later. The Pi? I already had it.

And time? A year ago, any single one of these apps would have been a project for several weekends. Most of them I would never have started - a tool for an audience of one was never worth the effort. That is the real change: the threshold for "worth building" dropped to almost zero.

Where experience still wins

Now the part I actually wanted to write about.

The AI wrote nearly all of the code. But again and again, the decision that mattered came from me. Not from knowing how to code a feature - from knowing how the system underneath works.

Notifications in the wrong app

At some point, notifications from several of my apps started showing up under a single one. The best part? That app didn't even have push notifications.

The push code was fine. The problem was the address: all of the apps were served from the same host, just on different ports. To the phone, they all came from the same place.

So I didn't ask the AI to debug notifications. I asked how to give each app its own host in Tailscale. Problem gone - also for every app I build next. If you follow my hosting series, you already know: hosts and addresses are not details.

Numbers that are right and still wrong

Car Stats told me my car's mileage was 48,456 km. Well, the car has over 315,000 km. The number was computed correctly - it was the distance driven since my first entry, labelled "mileage".

FPL Scout showed my captain sitting on the bench. I knew why: at the end of a gameweek, FPL automatically swaps out players who didn't play, and the API returns the lineup after those swaps. Right endpoint, wrong state.

No test catches this. The code does exactly what it was told. You need to know what the number should be.

The report agent that could leak my portfolio

One step of my report pipeline searched the web, saw my private portfolio, and could fetch any URL. That is a textbook setup for prompt injection: private data, untrusted content and a way to send data out. Simon Willison calls it the lethal trifecta. A malicious page could tell the agent to put my holdings into a URL and "fetch" it.

To be fair - I didn't spot it, a Codex review did. But the AI's first idea was to drop the risky step, which would make the reports much worse. So we went through what an attacker could actually do: run a shell, encrypt my files, smuggle data out through an image or a link.

The fix: split the data. Company names go to the steps with web access. My positions and investment theses go only to the step that has no tools at all.

You don't need to write this code yourself. You need to know the threat exists, so you can ask the question.

What didn't work

• Headroom went into RPI Manager and came out two days later. Small savings, and it blocked Claude Code's Remote Control.
• Car Stats had 7 bugs in its first 24 hours.
• The FPL Scout deadline widget needed about 14 iterations to look right.

Final thoughts

• Building small software for yourself is now often cheaper (and faster) than searching for an existing tool.
• AI is excellent at answering questions. Experience is knowing which question to ask.
• Security and privacy are where not knowing costs the most - and you won't get an error message.
• Know how you are billed. Caching, limits and sessions decide whether this costs $140 or $22,000.

Could someone without a technical background build these apps? Many of them - yes. Would they notice that their portfolio could leak through a URL? That is the edge. For now.

I hope this was a meaningful read - if so, let me know in the comments. And if you want to know how I work with the agents day to day, tell me too. That one deserves its own post :)

What's next

Back to the hosting series: the next episode explains the four rendering modes - SSG, SSR, CSR and ISR. See you there!

FAQ