Part 4 of the Homelab Telemetry series. The verdict exists (Part 3); now it has to reach me without touching the cloud.

Where this sits in the series:

  1. Part 1 β€” Overview: the whole pipeline, decision by decision.
  2. Part 2 β€” Logs: journald β†’ VictoriaLogs, zero agents on the laptops.
  3. Part 3 β€” A local LLM triages the logs on a four-core Celeron.
  4. Part 4 β€” Notifications: ntfy, and the two rabbit holes that ate an evening. (the end)

A daily verdict sitting in a log file on the NAS is a diary nobody reads. It needs to arrive. Locally. This is where I learned two things the hard way.

ntfy, self-hosted

ntfy is a tiny notification server: POST a message to a topic, subscribers get a push. The public ntfy.sh would mean my data leaving the network, so I self-host it on the NAS (it’s in the TrueNAS catalog). Publishing is just an HTTP call, which is perfect for a stdlib script:

curl -H "Title: log review" -H "Priority: high" \
     -d "sshd: 40 failed logins from one IP" \
     http://192.168.1.12:30184/homelab-alerts

The script maps the model’s SEVERITY to ntfy’s priority: none β†’ 1 (silent βœ“ you can glance at), up to high β†’ 5 (buzz-your-pocket urgent). Quiet days are unobtrusive; real problems are not. That mapping is the whole UX β€” I get pinged in proportion to how much I should care.

ntfy web UI showing a run of daily log-review verdicts at different severities

Rabbit hole #1 β€” browser notifications need HTTPS

I opened the ntfy web UI over plain HTTP on the LAN and it flatly refused to raise desktop notifications: “Notifications are only supported over HTTPS.”

That’s not an ntfy limitation β€” it’s the browser’s Web Notifications API, which only works in a secure context (HTTPS, or localhost). No server setting fixes it. What actually works over plain HTTP:

  • the phone app β€” the native ntfy app talks to the server directly and shows native notifications, no browser API involved. This is the intended receiver anyway;

  • the ntfy CLI on a laptop, which can fire notify-send:

    ntfy subscribe http://192.168.1.12:30184/homelab-alerts \
         'notify-send "$title" "$message"'
    

Only if you specifically want browser pop-ups do you need TLS β€” and then you put ntfy behind a reverse proxy (Nginx Proxy Manager is the friendly choice) with a cert from your own CA, imported into your devices. For a phone-first setup, don’t bother. I didn’t.

Sidebar: TrueNAS’s “HTTPβ†’HTTPS redirect” is for its own UI, not your apps. Your apps each listen on their own port with no TLS, so they speak HTTP. A CA on the box doesn’t magically wrap every app in HTTPS β€” each service needs its own cert or a proxy in front. Keep the internal scriptβ†’app traffic on HTTP; only the browser-facing bits ever need the proxy.

Rabbit hole #2 β€” Proton really doesn’t want scripts sending mail

Email was the optional second channel. I use Proton, so naturally I tried to send from Proton. The chain looked like this:

to send via Proton SMTP you need a token β†’ to get a token you need a verified custom domain β†’ to verify a domain you add a DNS TXT record β†’ …at your DNS provider β†’ for a domain you own.

I did not own a domain. I had invented one in the Proton UI five minutes earlier, and Proton was now asking me to prove I owned it. Proton’s SMTP submission needs a paid plan and a real, verified custom domain; Proton Bridge also needs a paid plan and a local app running. Both are wildly over-spec for “email myself a summary.”

The unlock was realizing I don’t need to send from Proton at all β€” I just need the message to land in the Proton inbox. So the script sends from any account I can already SMTP-auth to (a work relay, or a throwaway Gmail app-password), addressed to the Proton address. Credentials live in a chmod 600 email.json the script reads β€” never hardcoded β€” and email is skipped entirely if that file is absent:

{ "host": "smtp.example.com", "port": 587,
  "user": "sender@example.com", "password": "app-password",
  "from": "sender@example.com", "to": "me@proton.me" }

smtplib handles STARTTLS/SSL/relay. Done. No domain, no token, no Bridge, and no DNS record for a domain I don’t own.

The end state

review.py  β†’  ntfy (phone push, severity-scaled)
           β†’  email (optional, sent TO Proton, not from it)

Two notification channels, both fully local in the sense that matters β€” the logs and the analysis never leave the NAS; only a short verdict reaches me. It runs on two cron jobs while I sleep, and most mornings it tells me, in a calm silent ping, that nothing is on fire.

Which is the entire point. I didn’t build this to stare at dashboards. I built it so I could stop looking, and trust that a small, slow, slightly-taped-together brain in my basement is doing the looking for me.


That’s the series. If you build your own, steal the good decisions and skip the Wazuh detour, the 12B experiment, and the imaginary domain.