I was paying around $84/year for UptimeRobot's Pro plan plus another $20/month for a Datadog free-tier-with-overage thing on a side project. None of it made sense. The side project does maybe 200 requests an hour. I'm not running Netflix.
So over a long weekend in 2026 I spun up a $3/mo VPS, installed 7 self-hosted monitoring tools one after the other on the same box, and kept whichever one didn't make me reach for rm -rf at the end.
Same hardware for all of them: 1 vCPU, 2 GB RAM, 25 GB disk, Ubuntu 24.04. Same workload: 4 production websites pinged from outside + the host itself.
The 7 tools (and how I ranked them)
| Tool | Memory at idle | Setup time | First-time UX | Notes |
|---|---|---|---|---|
| Beszel | 12 MB (hub) + 8 MB (agent) | 4 min | "Wait, this is it?" | Single binary, agent + hub. Zero config. |
| Netdata | 320 MB | 6 min | Overwhelming | Auto-detects everything. So many charts. |
| Grafana + Prometheus | 410 MB combined | 35 min | Painful | You install one to be useful, the other to make the first one useful. |
| Pulse (Proxmox VE) | 95 MB | 8 min | Niche | Lovely if you run Proxmox. Empty otherwise. |
| Glances + InfluxDB | 240 MB | 20 min | Average | Glances is fine. InfluxDB is the tax. |
| Prometheus only | 180 MB | 15 min | Terminal-grade | You're querying PromQL or you're nothing. |
| Dozzle | 18 MB | 3 min | Logs only | Brilliant at one thing. Not a monitoring tool. |
The memory column matters more than it looks. On a 2 GB VPS, Netdata's 320 MB is 16% of your RAM gone before you've shipped any user code. Beszel's 12 MB is a rounding error.
Why I uninstalled the obvious choices first
Netdata. Auto-detects PostgreSQL, nginx, Docker, journalctl, every interface, every cgroup. Great. Also: 320 MB at idle and a dashboard so dense I scrolled for 90 seconds and gave up. It's the Salesforce of self-hosted monitoring - extremely capable, exhausting to use casually.
Grafana + Prometheus. I love Grafana. I do not love that the answer to "how is my server doing" is "let me first configure scrape targets in a YAML file, then write a PromQL query, then build a dashboard, then realize the dashboard library on GrafanaLabs has 600 versions of CPU usage." 35 minutes to first chart. On a side project.
Pulse. I don't run Proxmox. It told me so within 90 seconds. Fair.
Glances. Glances itself is great as a CLI. The moment you want web UI + history + alerts, you need InfluxDB or some other time-series store, and now you've got a moving-parts problem on a 2 GB box.
Dozzle. Spectacular for tailing Docker logs. Doesn't pretend to be monitoring.
Prometheus alone. PromQL is wonderful if you live in it. I don't. I want to glance at my phone in the queue at the coffee shop and know whether prod is up.
The one I kept: Beszel
Beszel is two binaries:
- The hub runs on one server. Web UI, history, alerts, push notifications.
- The agent runs on every server you want to watch. Reports back to the hub over an encrypted channel.
Setup, end to end:
# On the hub server
mkdir beszel && cd beszel
curl -L https://github.com/henrygd/beszel/releases/latest/download/beszel_Linux_x86_64.tar.gz \
| tar -xz
./beszel serve
# On every agent server
curl -L https://github.com/henrygd/beszel/releases/latest/download/beszel-agent_Linux_x86_64.tar.gz \
| tar -xz
KEY="$(cat ~/.ssh/id_ed25519.pub)" ./beszel-agent
# Add the agent in the hub web UI, paste the public key, done
That's 4 minutes if you've made a single typo, 3 minutes if you haven't.
The web UI shows CPU, memory, disk, network, temperature, Docker container stats, and a clean event log. Alerts go to email, Discord, ntfy, Pushover, Telegram - the usual stack. History as far back as you tell it to keep.
What sealed it for me: the same setup works for 1 server or 50. I started with the one VPS, then added the agent to my home homelab box (an old laptop running Debian under the TV), then to a Raspberry Pi running Home Assistant. Same dashboard. Same alerting. No reconfiguration.
What I'd actually pay attention to
Three things matter on a small server: memory headroom, disk filling up overnight, and "did the process I care about die." Beszel covers all three out of the box. Netdata covers them too, but you'll need to learn the dashboard. Grafana covers them once you've spent a weekend wiring it together.
If you run multiple boxes - even just "VPS + home lab + Pi" - the single-hub model is the unlock.
Costs (real numbers)
| Setup | Monthly cost | What you give up |
|---|---|---|
| UptimeRobot Pro | $7 | External-only checks, no resource metrics |
| Datadog Pro | $15+ per host | Per-metric overage. The trap. |
| Grafana Cloud Free | $0 | 10K metrics ceiling, then it scales fast |
| Beszel on a $3 VPS | $3 | Hosting it yourself, owning your data, no per-host fee |
| Beszel on a managed pod | $3 | Setting it up yourself |
The last row is shameless - I work on InstaPods and we ship a pre-baked Beszel image for people who don't want to run the install commands above. If you'd rather not touch a VPS, that path exists. If you'd rather own the box, the install commands above are the whole guide.
What I'd skip
- Don't self-host monitoring on the same box you're monitoring if it's a one-server side project. If that box goes down, your monitor goes down with it. Beszel's hub is small enough to run on a different cheap VPS, a Raspberry Pi at home, or a free-tier Oracle box.
- Don't pay for SaaS monitoring on a side project. The pricing model assumes you're running 10 production hosts and have a budget. You're running one production host and a budget of "ramen tonight or not." The math never works.
- Don't pick the tool with the prettiest dashboard until you've tried the tool with the smallest memory footprint. The prettier dashboard you can grow into. The 320 MB you can't grow into on a 2 GB box.
Try-this list (15 minutes total)
- Spin up the cheapest VPS you can find. Beszel's hub needs maybe 50 MB and a few MB of disk.
- Install Beszel hub. Open the web UI.
- Install the agent on whatever server you care about. Paste the SSH public key into the hub. Done.
- Set one alert: "CPU > 90% for 5 minutes." That's the single alert that catches the most "something is wrong" cases.
The whole thing costs less than one month of UptimeRobot Pro and tells you more.
What's your monitoring stack right now? And how many alerts do you actually act on vs. silently ignore? Curious how this lands with people running multi-box homelabs - the agent/hub model seems made for it but I've only run it on three boxes so far.
Top comments (0)