Inspired by this post by Geoffrey Huntley, titled "How to build a coding agent", I decided to set off on creating my own Claude Code-like, tool-using agent, designed for investigating, experimenting, and writing Python applications.

I made a very simple prototype, and was amazed at how well it worked. Surprisingly it was able to do a lot of what these code agents can do with just 5 simple tools:

  • read_file - read the contents of any file
  • list_dir - list the files and directories of a given file path
  • python_tool - execute any Python code, or a Python script
  • edit_file - create or edit the contents of a file
  • code_search - search for a regex pattern across all files under a directory

Here it is on Github. You can try it for yourself: link

Anyhow, I thought it was so awesome that I could make a mini version of Claude Code CLI (or Codex CLI, etc) with just a few simple python functions - the entire system being just 220 lines of simple python code!

I decided to build upon this, and create a nicer UI around it.

srAgentCloud

Link to Github: here

Terrible name, I know. But the idea was to host an app on a small device (on my network or in the cloud), and then communicate with it using a nice chat UI, to get it to build apps, or investigate / research things for me.

Given that:

I thought it would be a good idea to deploy it on a spare Raspberry Pi Zero 2W that I had laying around. The device only has 512MB of RAM, but it's plenty for this kind of app. The only time it feels slow is when you create a new project (which creates a new venv) - which takes about 25 seconds on this Pi

Other factors:

  • I do not currently have a Claude or OpenAI monthly subscription (prefering OpenRouter + OpenwebUI), but I have used GitHub Copilot CLI quite a bit
  • I really wanted a way to work on developing an app from my phone, while sitting on the couch or while away from my computer (Copilot CLI doesnt have this)
  • I wanted to be able to quickly skim what the agent is doing for every tool call (to make sure nothing looks nefarious, and to give it pointers if necessary)
  • I wanted to see what this lean harness would cost me in LLM API credits
  • I also wanted to see what different LLM models were capable of with this very simple custom built harness

App ideas sometimes come into my head when I'm not at my computer. I wanted a way to quickly and easily convert that idea into a working app as frictionlessly as possible.

Here are some examples of the things I've made with it while sitting on the couch:

  • Weekly Check-in App: Allows me and my family to make simple "posts" in a private "social media". Complete with logins, "followers" (to see each others posts), and importantly, discord notifications from a bot to ensure we complete a weekly check-in post on Sundays - I thought it would be nice for me (or family) to look back on the posts years from now and see that: "these are the interesting things I did this year"
  • Some webscraping bots to check for price changes on specific items at a retailer periodically
  • A 'Bugs & Features' dashboard to keep track of things I want to fix for any of my projects
  • Little mini-apps to: convert emails EML files to their HTML, Extract URLs from that HTML, and others to webscrape the contents of those URLs

The system is currently generalizable enough that you can use it like a research tool as well, instead of just for simple app development: Ex. Investigate this. Make a script to automate this. Find out how to best webscrape this webpage, and then make a script that does it for all of these URLs

If you've got a spare Raspberry Pi Zero lying around, and you have an account with OpenRouter, you can feel free to give this a try: Available here on my Github. Installation and usage instructions, as well as a project description is in the readme there.

Feel free to reach out if you have any questions, or need help setting it up.

What I learned

  • Models: 'gemini-3-flash-preview' gives a good balance between capability and cost. You probably don't need a Claude Sonnet 4.6 (5x more expensive) to build a simple Flask web app. You can even try the newer 'gemini-3.1-flash-lite-preview', which runs at half the cost!
  • Costs: This depends on the model. With the Gemini 3 Flash, it costs me about 10-20 cents USD to convert an idea into a concrete, working application that I can access and start playing with on my phone.
  • Coding Agents Are Simple: At least at their core. You only need to give the agent a few simple tools to enable it to build you cool apps.
  • Raspberry Pi Zeros Are Great!: For API-heavy tasks like these, you really don't need much processing power - and at $15 USD, they're the perfect cheap little system for this kind of project. Worst case, if the agent messes up the system, you just reformat the SD card and you're back in business.

Conclusion

2026 seems to be shaping up as the year of Coding Agents - building customized software on demand, using plain english is a lot of fun. It helps turns your app ideas into something concrete very quickly.

Building a Coding Agent yourself is surprisingly very easy. I'd definitely recommend trying to build a simple one yourself - it's fun, and you'll get a better understanding of how these systems work, which can help you use them more effectively.