My Server Now Guilt-Trips Me Over Text. Here's the Build.

Every home lab has a dirty secret: you spend a weekend making everything redundant, monitored, and bulletproof — and then you stop looking at it. The dashboard goes unopened. The Grafana panels gather digital dust. Out of sight, out of uptime.

I hit that wall hard earlier this year. My little rack in the closet was fine, probably. Probably is not a monitoring strategy. So I did what any reasonable person with a spare Raspberry Pi and a Twilio trial account would do: I built an alert system that texts me. And then, because a plain "Disk usage 91%" wasn't going to get me off the couch, I made the messages dramatic. Petty, even. My server texts like an ex who has thoughts.

The Stack Is Boring On Purpose

Here's the whole thing:

• Uptime Kuma on the Pi, watching my services over the LAN
• Twilio for SMS, because I wanted texts to feel urgent in a way push notifications never do
• cron jobs running a handful of shell scripts for the stuff Uptime Kuma can't see — disk space, temperature, backup freshness
• A single Python script, maybe 80 lines, that formats everything into a message and fires it off

That's it. No Kubernetes, no message queue, no "observability platform." If this setup takes you more than an afternoon, you've over-engineered it, which is ironic given the subject matter.

The Pi itself is a 4B that was previously serving time as a retro gaming machine. It sips power, it's been rock solid, and if it dies, the irony of the monitor needing monitoring is a problem for future Aleks.

The Copy Is Not Boring At All

Here's where I deviated from every tutorial I read. The default alert tone of the entire monitoring industry is clinical: "WARNING: High CPU usage detected." Clinical doesn't motivate me at 10pm on a Tuesday. So I wrote a copywriting layer.

When disk space runs low on the media server:

"hey. so. /dev/sda1 is at 92% again. i'm not going to say it's the Plex library. but it's the Plex library."

When a service goes down:

"Jellyfin is down. third time this month. i've stopped taking it personally. have you?"

When the closet temperature sensor crosses 85°F:

"it's 86 degrees in here. i am a computer. i do not have sweat glands. please open a door."

Something about the tone shift works. A red banner in a dashboard is ambient noise. A text message written like a disappointed roommate demands a response. My median time-to-fix dropped from "whenever I next wander downstairs" to "within the hour."

I keep the templates in a YAML file so I can rotate them — partly for variety, partly because even a joke gets stale on the fortieth repetition.

What a Month of Alerts Taught Me

I expected the interesting part to be the build. It wasn't. The interesting part was the data.

Over four weeks, the system sent me 61 texts. Breaking it down:

• 31 were disk space warnings on one particular VPS I refuse to clean up
• 11 were temperature alerts, almost all during a July heat wave, almost all resolvable by leaving the closet door open
• 9 were a flaky smart plug rebooting my switch — the single most useful discovery of the month
• 10 were legitimate service outages, and 7 of those were caused by that same switch

So roughly 80% of my alerts traced back to two root causes. That's the real lesson hiding in this project: once you make failure visible, you discover your infra doesn't have fifty problems. It has two problems that have learned to cosplay as fifty.

I fixed the smart plug, deleted 40GB of docker images I'd been hoarding "just in case," and the alert volume dropped by two-thirds. The dramatic texts now arrive maybe twice a week, which is a sustainable relationship.

Rules I'd Give Past Me

If you're going to build something like this, a few guardrails from hard experience:

1. Rate-limit everything. My first version texted me every five minutes during an outage. Twelve identical texts is not urgency, it's spam, and you'll mute the thread. One alert, then a follow-up only when state changes.
2. Alert on symptoms you can act on. "CPU at 80%" made me anxious and changed nothing. "Backup hasn't run in 26 hours" sent me straight to the terminal.
3. Keep the drama in the copy, not the thresholds. The triggers should be boring and conservative. The personality lives in the message, never in a 70%-CPU warning.
4. Twilio costs real money per SMS. It's pennies, but a runaway loop at 3am is how you learn what "pennies" means at scale. Cap it.

Was It Worth It?

An afternoon of work, a Pi I already owned, and about $1.40 in Twilio credits for the month. In exchange: I fixed a hardware fault I didn't know existed, reclaimed disk I didn't know was full, and my closet server now has more emotional intelligence than most SaaS products I pay for.

The full scripts and the YAML of alert copy are in the lab repo — fair warning, the comments are also written in-character. Apparently I couldn't help myself.

Next experiment on the bench: teaching the system to send a single, smug "everything is fine, not that you asked" message on Sunday evenings. Morale matters, even for hardware.