Running OpenClaw 24/7 feels light when you npm i -g openclaw on a laptop, but keeping an agent online day-and-night is a different animal. Power bills, fan noise, and upstream bandwidth suddenly matter. This post summarises what I learned after punching numbers on five common options the community keeps bringing up: Mac Mini M4, Raspberry Pi 5, DigitalOcean droplet, Hetzner VPS, and the dusty laptop on a shelf. I’m assuming a single agent connected to Slack + Gmail + browser control, moderate shell tasks, and a log-retention window of 30 days.

Baseline requirements: how much does OpenClaw actually need?

The gateway (web UI) is CPU-light; the daemon does the polling, schedule handling, and context writes. What bites is memory—tool integrations and browser sessions stack up. Here’s the minimum I’d budget if you want no midnight OOM hiccups:

  • CPU: 4 logical cores. Node 22’s JIT likes it, and Chromium-based browser control spawns a couple extra processes.
  • RAM: 4 GB absolute floor, 8 GB comfortable. Memory leaks aren’t rampant but they happen when you load 20+ Composio tools.
  • Disk: 15 GB free. Logs + the node_modules folder are the hogs.
  • Network: 5 Mbps sustained upstream is enough. If you enable Whisper for voice, double it.

All five machines clear those numbers, but with wildly different headroom, bills, and hassle factors.

Mac Mini M4 (~$499 on refurb): community favourite, mostly overkill

Half the GitHub issues that mention self-hosting are folks running a Mac Mini under a desk. Apple silicon’s single-thread performance is absurd, and it sips power. My numbers are based on the entry M4 chip, 8 GB RAM, 256 GB SSD.

Pros

  • Silent at < 10 dB. You literally forget it’s on.
  • Draws 6–12 W idle, 20 W peak during Chromium runs. That’s ≈ 88 kWh/year → $13 in electricity (US average 15¢/kWh).
  • macOS is a first-class Node target. No musl/gcc dance.
  • Crazy headroom: you can run three agents, each with its own memory store, before you notice load.

Cons

  • Up-front $499 is real money if you just want one chatbot.
  • macOS updates reboot whenever they like unless you babysit softwareupdate --schedule off.
  • No ECC; if you’re allergic to silent corruption, look elsewhere.

1-year cost

$499 hardware + $13 power = $512.

Raspberry Pi 5 8GB (~$80 board + $30 accessories)

I spent three evenings compiling Chromium so you don’t have to. It works, but the Pi 5 is the slowest option on this list and the most temperature sensitive.

Pros

  • Absurdly cheap entry cost.
  • 3–4 W idle, 12 W under full browser load. Annual power: ≈ 55 kWh → $8.
  • Fanless if you under-clock, otherwise a 30 mm fan hum at 35 dB.
  • You can stash it behind a router with PoE and forget.

Cons

  • CPU bound: Puppeteer page loads are 4× slower than on the Mac Mini.
  • MicroSD lives a hard life with write-heavy logs. Use an NVMe HAT or expect corruption.
  • Some npm native modules ship pre-built binaries for glibc 2.31+. Raspberry Pi OS might lag.

1-year cost

$110 hardware (board + 256 GB NVMe + HAT + case) + $8 power = $118.

DigitalOcean droplet (s-2vcpu-4gb, $24/mo)

The “just pay and forget” route. I picked the regular Intel droplet, not Premium AMD, because price/perf is close for this workload.

Pros

  • No hardware to maintain, no power bill surprises.
  • Backup snapshots ($2.40/mo) are one click.
  • DO’s network is fat; 1 Gbps egress bursts make WhatsApp media uploads instant.

Cons

  • $24/mo = $288/year before Bandwidth. First free TB is plenty unless you stream audio.
  • You’re sharing an SSD. I/O latency spikes during NYC midnight backup window.
  • 24/7 monitoring: outages happen; to get SLA you need to pay for uptime monitors.

1-year cost

$288 droplet + $29 snapshots + $0 bandwidth (if <1 TB) = $317.

Hetzner CPX11 VPS (€4.35/mo ≈ $4.75)

European readers swear by Hetzner. The CPX11 (2 vCPU, 2 GB RAM) is tight on memory out of the box, so the common trick is to add 2 GB more via --memory 2G for €1.70 extra.

Pros

  • Ridiculously cheap: €73/year even with backups.
  • AMD EPYC Milan cores benchmark 30% faster per vCPU than DO’s Intel ones.
  • 40 TB traffic quota – basically unlimited for a bot.

Cons

  • Latency from the US West Coast is 160 ms. Fine for chat but bad for real-time voice.
  • German data residency may conflict with some US-only compliance checklists.
  • Support is ticket-only and blunt.

1-year cost

€73 VPS + €20 (backups) ≈ $101.

The repurposed old laptop (cost $0, or is it?)

I took a 2015 Dell XPS 13 (i5-5200U, 8 GB RAM, 256 GB SSD) and installed Ubuntu 24.04 LTS.

Pros

  • Free if it’s in the closet.
  • Comes with built-in UPS (battery) for short power flickers.
  • Integrated screen + keyboard = emergency console when SSH is down.

Cons

  • 25 W idle, 40 W under load. That’s 263 kWh/year → $40 in power.
  • Fans spin at 45 dB when Chromium wakes up. My spouse noticed.
  • Consumer SSD at 9 years old: S.M.A.R.T shows 14% life left. Backup or cry.

1-year cost

$0 hardware + $40 power + $20 replacement SSD (when, not if) = $60.

Power, noise, and bandwidth cheat sheet

  • Mac Mini M4: 10 W avg, 0 dB, 1 Gbps NIC
  • Pi 5: 5 W avg, 35 dB (with fan), 1 Gbps NIC
  • DigitalOcean: Your wall plug draws 0 W; data center fans are somebody else’s problem.
  • Hetzner: Same as DO.
  • Old Laptop: 30 W avg, 45 dB, 802.11ac Wi-Fi unless you keep the lid open for Ethernet.

12-month total cost of ownership

  • Mac Mini M4: $512
  • Raspberry Pi 5: $118
  • DigitalOcean droplet: $317
  • Hetzner VPS: $101
  • Old laptop: $60 (assuming you already own it)

Mac Mini’s price looks steep, but resale value after a year is $400+; factor that and you’re near parity with a droplet. The Pi 5 is cheapest but labor-heavy. Hetzner is the sleeper hit if latency isn’t critical.

Decision matrix: which rig for which use-case?

  • I need zero maintenance: pick DigitalOcean, flip the backups switch, budget $26/mo.
  • I want a silent box in the home office: Mac Mini wins.
  • I’m building a fleet of 10+ hobby agents: Hetzner. Stack them, automate with Cloud-Init.
  • I have $100 total and enjoy tinkering: Raspberry Pi 5. Accept slower browser automation.
  • I literally have no budget: dust off the laptop, install TLP, and accept the hum.

Practical next step

Once you pick a host, run the same smoke test:

# Node 22 LTS curl -fsSL https://deb.nodesource.com/setup_22.x | sudo -E bash - sudo apt-get install -y nodejs # or brew install node@22 on macOS # Install and start OpenClaw groupadd --system openclaw && useradd -s /usr/sbin/nologin -g openclaw openclaw sudo -u openclaw npm i -g openclaw@latest openclaw daemon --memory 4G --port 3000

Leave it for 48 hours, check htop for steady RAM, then wire your chat channels. If the box idles under 70% memory and the CPU rarely spikes past 50%, you picked well. If not, scale up or swap rigs before your agent starts ghosting users.