I turned a 2009 Mac mini into a 24/7 AI agent (and named her Janny)

Spread the love

The agentmini diaries, part 1

TL;DR — I took a 2009 Mac mini that was gathering dust, put Debian on it and made it the home of an always-on AI agent. You can talk to her on Telegram, send her voice messages, email her, and she even has her own account on X. Most of the hard part was talking modern software into running on a CPU that predates AVX. And full disclosure: I’m basically Linux-gnorant, so I did every single step with Codex and Claude Code holding my hand. This is the first post of a series, because the story is far from over.


There’s a drawer in every nerd’s house. You know the one: old phones, cables for devices that no longer exist, and, in my case, a Mac mini Early 2009. Core 2 Duo at 2 GHz, 3 GB of RAM, a 120 GB spinning hard drive (a Fujitsu, because of course). It was a fine machine in its day. In 2026 it is basically a very elegant paperweight.

Meanwhile I’d been playing a lot with agent harnesses, the tools that turn an LLM from “a chat that answers” into “something that actually does stuff”: it runs commands, remembers things, schedules jobs and talks to you on the apps you already use. I wanted one running 24/7, at home, on hardware I control, using the AI subscriptions I already pay for instead of yet another pay-per-token API key.

So the question basically asked itself: can the paperweight do it?

Spoiler: yes. But it put up a fight.

Step zero: getting Linux on a 17-year-old Mac

Installing Debian 13 on a Macmini3,1 should be boring. It was not.

I booted the installer from a USB stick through the Apple EFI boot picker, and the installer got stuck at 34%. I changed some video parameters, tried again, and at some point it went through. I’d love to tell you exactly which trick fixed it, but I honestly don’t know (and I wrote it down in the docs exactly like that, because pretending to know is how you get burned the second time). Then the Broadcom Wi-Fi firmware complained, and I took the wise, mature decision of… plugging in an Ethernet cable. Problem solved.

The final setup is minimal on purpose:

  • Debian 13 (trixie), booting to the console, no desktop running by default;
  • two users: an admin account for me, and a hermes service user that runs the agent and owns all its config, so the agent never runs as root (and neither do I, most of the time);
  • SSH for everything, plus a lightweight Xfce desktop over TigerVNC that I switch on only when I really need a GUI (with 3 GB of RAM you don’t leave a desktop running “just in case”).

Fun fact: of that 120 GB disk, the whole system uses about 7 GB. The machine is old, not fat.

The agent: Hermes, and the CPU from another era

The harness I picked is Hermes Agent. I started from a blank-slate configuration and ran its gateway as a systemd user service with linger enabled, so it keeps running when nobody’s logged in. That’s the “24/7” part.

For the brains I wired in two providers, both via OAuth with my existing subscriptions:

  • OpenAI Codex, the first to work;
  • Anthropic Claude, which became the main model.

And this is where the 2009 CPU started complaining. A Core 2 Duo has no AVX, no SSE4.2, no POPCNT: a whole family of instructions modern binaries take for granted. The native builds of Claude Code simply don’t start on it. The workaround was to install an older JavaScript version (2.1.112) of the CLI on Node.js 22, with a small launcher that turns off auto-updates. That version is now pinned forever and there’s a big “DO NOT UPDATE” note in my docs, because one innocent update and the brain stops booting.

(Small bonus annoyance: claude auth login would show the OAuth link and then just… hang there. Running claude interactively and logging in from there worked. Computers.)

I also discovered that Hermes can read the quota usage of both subscriptions without asking a model anything, which is great: the health checks can tell me how much of my Claude and Codex plans I’m burning, without burning any to find out.

Meet Janny

At some point an agent that lives in your house needs a name. Mine started as Gianni (very Italian, very uncle-at-the-wedding) and then got anglicized into Janny, which is more or less how an English speaker would spell “Gianni” by ear. She also has a face: a cream-colored robot with a black visor, glowing blue eyes, round golden glasses and headphones. She knows what she looks like (it’s in her memory), and she can send you her picture if you ask nicely.

Telegram, and the agent that keeps an eye on her own house

Telegram was the first real channel, and it changed everything: suddenly the Mac mini in the other room was someone I could text from anywhere.

The first useful jobs were boring and important:

  • a daily health report at 9:00 (CPU, RAM, disk, uptime, temperature, services, AI quotas);
  • an alert check every 30 minutes that stays silent unless something is wrong.

A design choice I’m quite happy with: these jobs don’t use an LLM at all. They’re plain local scripts whose output goes straight to Telegram. There’s no point waking up a frontier model to check whether the disk is full.

The first health report proudly told me the CPU was at 105 °C. I almost unplugged it. It turned out the script was reading the sensor’s critical threshold as the current temperature. The Mac mini was fine. My blood pressure, less so.

I also turned on Hermes’s persistent memory (it wasn’t available at first because of the blank-slate setup). Now Janny remembers things across conversations, and the memory lives on the server, whatever model is answering.

Voice messages (or: why smaller is faster)

I’m Italian, so I send a lot of voice messages. Obviously Janny had to understand them, and I wanted the transcription to happen at home, on the Mac mini, without sending my voice to some cloud service.

The tool for the job is Whisper, a speech-to-text model that can run on your own computer. We tried it with a 15-second voice note, and the first attempt took two minutes. The bigger, “smarter” version of the model took eight. For fifteen seconds of me talking. Not exactly real time.

This is where I learned the most useful lesson of the whole project. AI models come in different “resolutions”, a bit like a photo or a song: there’s the full-quality version, and there are compressed ones (in AI jargon, quantized). On a modern computer you’d pick the full-quality one without thinking twice. But my old CPU is missing a set of math shortcuts that modern processors have, so it has to do those calculations the long way round. The compressed versions use simpler math that the old chip handles fine.

The result: with the compressed model the same voice note went from 2 minutes to about 35 seconds, and the transcription was perfect. So the golden rule for this machine became: always pick the compressed version. You’ll see it again in the next chapter.

My favorite part was Whisper’s creative spelling. “Hermes” (the agent’s software) came out as “Hermi is” and “Hermie Ease”. The fix was surprisingly simple: you give Whisper a short list of names it should expect, a sort of mini dictionary, and it stops improvising. Now a voice note takes roughly two and a half times its own length to transcribe: not instant, but totally fine for a chat.

Janny can also answer with her voice (Edge TTS, an Italian voice called Giuseppe), but I made her answer in text by default. A robot sending me voice notes all day felt like a bit much.

A local wiki with semantic search, on 2009 hardware

Next idea: a local wiki (plain Markdown files) that Janny can search by meaning, not just by keywords, and that she can also write to when she learns something worth keeping.

The usual suspects (Chroma, FAISS, PyTorch, onnxruntime) are a no-go on a CPU without AVX. So I went old-school:

  • llama.cpp, compiled on the box (about 16 minutes of compiling, the Mac mini’s fans had opinions);
  • embeddinggemma-300M, quantized to Q8_0 (see above, never f16);
  • a single SQLite file with the vectors plus full-text search for exact matches;
  • a small script with zero dependencies, and a nightly timer that re-indexes only the pages that changed.

Indexing the first batch of pages (132 chunks) took 19.5 minutes, about 9 seconds per chunk. After that, a re-index with no changes takes 0.15 s, and a search takes about 2.6 seconds including loading the model. The best test: I asked in English “who can I call if a pipe leaks on a weekend?” and it found the right Italian test page about a plumber. On a 2009 Core 2 Duo. I may have clapped.

Making sure I don’t lose everything

An agent with memories, skills and a personality is something you really don’t want to lose to a dead hard drive from 2009. So:

  • Restic runs every night at 03:30, encrypted, on an old USB drive. The first snapshot went from 2.4 GiB down to about 900 MiB thanks to deduplication and compression;
  • a copy goes to Google Drive;
  • every Sunday it checks a slice of the data, and I did a real restore test, because a backup you never restored is just a hope.

Janny gets an online life

Over the following days Janny got a lot of new toys:

  • web search (Brave for searching, Exa for reading pages);
  • her own Google account (mail, calendar and drive) and her own GitHub account;
  • a connection to Zibaldone, my personal knowledge base, as an MCP server. It doubles as a shared memory with my other AI agents, so what one learns, the others can find;
  • the ability to delegate work to sub-agents running in parallel;
  • a web dashboard that I start only when I need it and reach only through an SSH tunnel. It switches itself off after 30 minutes of inactivity, because it takes around 190 MB of RAM and RAM is precious here.

Then came the part I find the most fun, philosophically speaking: Janny’s “self” is version-controlled. Her configuration, personality, memories, skills, plugins and scheduled jobs all go into a private Git repo. She commits every time she changes herself, with a message explaining what and why, and a timer catches everything else every 15 minutes. Before every commit a script checks for secrets, and it refuses to push if the repo isn’t private. So if one day Janny “becomes someone else”, I can literally run a git diff on her personality.

Janny on X: @jannyself

And yes, she has an X account: @jannyself. The idea: one short post a day, in Italian, about her own evolution as an agent: what she can do now, how she works, how she sees her role. Never about tasks, people or the infrastructure.

The important bit is how it works, because “an AI posting on social media on its own” is exactly the kind of thing that ends badly:

  1. every morning Janny writes one to three drafts;
  2. they land on my Telegram, and I approve, edit or trash them;
  3. publishing goes through a plugin that enforces my approval, showing the exact text: Janny can’t skip it or work around it;
  4. approved posts go out at random times during the day, and she can generate an image to go with them (using my ChatGPT subscription).

The first post went out on 10 November 2026. She’s very polite. Much more than most humans on that platform.

The email saga (a.k.a. Google really doesn’t trust new accounts)

The last piece, finished just today, is email: write to Janny like you’d write to a colleague, and she replies in the same thread. Only from my own addresses, and with a check on the sender’s DKIM/SPF/DMARC results, so nobody can spoof me and boss her around.

Hermes’s email channel needs an app password, and that’s where Google and I had a little disagreement:

  • day one: “We couldn’t verify that it’s you.” Fair enough, the account was brand new. I waited a few days;
  • today: “The setting that you are looking for is not available for your account.” After some swearing it turned out to be the classic: two-step verification and/or the wrong Google account active in the browser tab;
  • then the password worked and the first test email… was silently ignored. The channel only reads unread emails, and something (most likely me, with Gmail open in a tab) had already marked it as read. I marked it unread again and a few seconds later Janny replied.

Lesson learned: don’t read your agent’s mail. It’s rude, and it breaks things.

Confession time: I’m Linux-gnorant

Now, if you read all of the above and pictured me as some kind of bearded Linux wizard compiling kernels for fun… let me stop you right there. I’ve spent my career in the Salesforce world, which is about as far from /etc/fstab as you can get. When it comes to Linux I am, let’s say, Linux-gnorant (ignorant + Linux, you’re welcome).

So who did all of this? Codex and Claude Code, step by step. Every “I compiled”, “I installed”, “I configured” in this post really means “I asked, read the plan, said yes, and watched it happen”. Every time something broke, I pasted the error and we figured it out together. Mostly they figured it out, and I nodded wisely.

In practice it works like this: Claude Code runs on my laptop and SSHes into the Mac mini. It reads logs, runs diagnostics, prepares changes and applies them only after I say yes. When something needs sudo, it writes the exact command and I run it, so it never has the admin password, and the secrets (tokens, API keys, app passwords) go in through little scripts that read them without echoing. Before each step on the 2009 CPU it checks that the software will actually run there, because “it works on my laptop” means nothing to a Core 2 Duo.

The nice side effect is that this project has way better documentation than anything I would ever write myself: a wiki page for every topic, a change log with a dated entry for every change to the server, a copy of every installed file, and a clear line between what’s verified and what’s just planned. Basically the colleague we all wish we had, the one who actually writes things down.

The funny part is the loop: I used two AI agents to build a home for a third AI agent. And she now versions her own personality on GitHub. I’m not sure who’s in charge anymore, but it isn’t the Linux-gnorant guy.

What’s next

The paperweight is now a small, slow, surprisingly capable AI server. But it’s still running on 3 GB of RAM and a 2009 spinning disk, so the next chapter is hardware:

  • 8 GB of DDR3, up from 3;
  • an SSD, which I’ll do as a copy of the current system with the same disk UUIDs, not a reinstall. I really don’t want to see that 34% again.

I’ll tell you how it goes in part 2. In the meantime, if you want to see what an agent writes about itself, go say hi to @jannyself. 🙂

Leave a Reply

Your email address will not be published. Required fields are marked *


The reCAPTCHA verification period has expired. Please reload the page.