OpenClaw is the fastest-growing way to run a personal AI agent, and the first question after "this is amazing" is always the same: where do I keep it running? An agent that dies when your laptop lid closes isn't an assistant, it's a demo.
We operate OpenClaw at production scale — every app hosted on VibeKit gets its own persistent OpenClaw agent, which means we run one of the larger multi-tenant OpenClaw fleets anywhere. This post is the comparison we wish had existed when we started: the five real hosting options, what each actually costs, and the failure modes the tutorials skip.
The short answer
| Option | Monthly cost | Setup effort | Survives your laptop closing? | Maintenance burden |
|---|---|---|---|---|
| Your own machine | $0 | Minutes | No | Low |
| Docker on a VPS | $5–12 | An afternoon | Yes | Yours, forever |
| Raspberry Pi | ~$3 power (after ~$80 hardware) | A weekend | Yes (until the SD card dies) | Yours |
| Shared/app hosting (Hostinger-class) | $5–10 | Fights you | Sort of | High — wrong tool |
| Managed / hosted agent | $0–20 | Minutes | Yes | Provider's |
Every option below is a legitimate choice for somebody. The honest question is how much of your agent's uptime you want to own personally.
Option 1: run it locally (free, and fine for trying it)
npm install -g openclaw and you're running. This is the right first step for everyone, and the wrong permanent home for almost everyone.
What the quickstarts don't emphasize:
- The agent only works while the machine is awake. Scheduled tasks, cron jobs, and anything long-running silently stops with your lid. If you're using OpenClaw as a chat toy, fine. If you're using it as an assistant that does things while you're away, this is disqualifying.
- Model auth lives on that machine. Moving later means re-doing auth profiles, migrating the state directory, and re-learning permissions — budget an evening for the move you'll eventually make.
Option 2: Docker on a VPS ($5–12/mo — the default DIY answer)
A $6 Hetzner or DigitalOcean box running OpenClaw in Docker is the canonical self-hosted setup, and it genuinely works. If you enjoy operating servers, this is a good home.
The parts you're signing up for, from our production notes:
- Memory is the real constraint, not CPU. A Node agent runtime plus a browser tool plus whatever your agent installs will eat a 1GB box. 2GB is the honest floor; 4GB if the agent does real builds.
- Updates are on you. OpenClaw ships fast — multiple releases a week. A version pin protects you until it doesn't (we've watched startup migrations block on stale sidecar configs; the fix was manual). Auto-updating without a health-gate occasionally means waking up to a dead agent.
- State is precious. The agent's memory, session transcripts, and workspace live in its state directory. If your Docker volume strategy is "the default", your agent's entire memory is one
docker system pruneaway from gone. Back up the state dir specifically. - Security is not optional. An agent with shell access on a public box is a real attack surface. Sandboxing (we use bubblewrap in production), firewall rules, and never putting provider keys in world-readable files are the minimum. The tutorials mostly skip this part.
None of this is hard for someone who runs servers anyway. All of it is recurring.
Option 3: Raspberry Pi (the hobbyist's honest choice)
"run openclaw on raspberry pi" is a genuinely popular search, and it works — OpenClaw runs fine on ARM, and a Pi 5 with 8GB handles a personal agent comfortably at ~$3/month of electricity.
Two things kill Pi agents in practice:
- SD card wear. An agent writes constantly — transcripts, memory, logs. Consumer SD cards die under sustained writes, taking the agent's memory with them. Use an SSD over USB, or accept that your agent has a lifespan.
- Residential connectivity. If you want to reach the agent from outside (the whole point of a phone-accessible assistant), you're now maintaining Tailscale or a tunnel, plus dynamic DNS. Each piece is small; the stack of them is why the "check on my agent from my phone" experience on DIY setups stays janky.
Great weekend project. Real infrastructure it is not.
Option 4: shared hosting (Hostinger-class) — mostly a mismatch
Autocomplete suggests a lot of people are trying this ("openclaw hostinger setup"), so the honest answer: most shared hosting can't run OpenClaw properly. You need a persistent Node process, a writable state directory, and usually root or Docker access. Shared PHP-era hosting gives you none of that; their VPS tiers do — at which point you're just in Option 2 with a different logo.
If a "Hostinger tutorial" has you running OpenClaw inside a hosting-panel Node app slot, expect the process reaper to kill your agent between requests. It's not hosted; it's intermittently resuscitated.
Option 5: managed / hosted OpenClaw (what we ended up building)
The gap between "OpenClaw is incredible" and "I want to babysit a Linux box" is exactly why hosted agents exist. In a managed setup, the provider owns the update cadence, memory persistence, sandboxing, and uptime; you own the conversation with your agent.
Disclosure and specifics, since this is our category: VibeKit runs a persistent OpenClaw agent for every app you build with it — the agent builds your app, hosts it on a live domain, and keeps working across sessions, driven from your phone or the web. There's a free tier (rotating pool of free models, no card), and BYOK (Anthropic or OpenAI) is free with no markup. The fleet lessons above — state-dir backups, health-gated auto-updates, bubblewrap sandboxing, group-writable workspace permissions — are things our platform does so users never learn they exist.
Managed isn't automatically better. You trade control for uptime: if you want to hand-tune your agent's runtime, own the box. If you want the agent to just be on, don't.
Can you host OpenClaw for free?
Genuinely free options: your own hardware (a spare machine or Pi you already own — you pay attention instead of rent), always-free cloud tiers (Oracle's ARM free tier fits OpenClaw and survives; expect occasional capacity pain), or a hosted free tier like ours. A $0 VPS that you operate isn't free; it costs the thing agents were supposed to give back.
What we'd actually pick
- Just exploring → your laptop, today, no setup.
- You run servers for a living and enjoy it → 2GB VPS + Docker + a state-dir backup cron. It'll be solid.
- Hobby project, tinkering is the point → Pi with an SSD. Have fun, that's the point.
- You want the agent working while you live your life → managed hosting, ours or anyone's. The agent economy only compounds when the agents stay up.
The mistake isn't picking any of these. It's picking one whose maintenance burden you'll quietly stop paying three weeks in — a half-maintained agent is worse than none, because you stop trusting it exactly when it was starting to be useful.