If you're shopping the "should I self-host or use a managed host" decision, hardware is usually the first question. Short answer: OpenClaw will run on a $5/month VPS, but it won't run well. Here's the honest breakdown.

Self-hosting: minimum vs. recommended

Resource	Bare minimum	Recommended	Why
CPU	1 vCPU	2+ vCPU	The stealth browser is the cost center. Single-vCPU hosts will block on Chromium during page loads.
RAM	1 GB	2 GB	OpenClaw alone fits in ~300 MB. Chromium is the rest. We watchdog-kill at 1.5 GB on our pool nodes.
Disk	5 GB	10 GB	Docker image is ~1.5 GB. Leave room for browser cache, logs, and workspace state.
Bandwidth	100 GB/mo	500 GB/mo	A browser-heavy agent will pull a lot of pages. Cheap VPS plans cap out fast.
OS	Linux x86_64	Ubuntu 22.04 LTS	We test against this. ARM works but isn't a tested path.

Bare minimum gets you a working agent for low-traffic personal use. Anything that involves the stealth browser doing real work — agentic web tasks, scraping, form fills — needs the recommended tier.

What doesn't scale linearly

Two things fight you when you self-host:

The Anthropic API key is the hot resource. Tier limits aren't about your VPS; they're about your relationship with Anthropic. A new Tier 1 key gets rate-limited fast under any kind of agent load.
The browser is single-tenant. OpenClaw's stealth browser holds a Chromium process per active session. You can't pack 50 users on a 4 vCPU host. If you want multi-user, you're now writing pool-node-style infrastructure yourself.

This is the core argument for managed hosting: the second user costs you 100% as much as the first, because each one needs their own browser. There's no efficiency gain from running OpenClaw on bigger hardware unless you also write the orchestration layer.

ShipClaw managed: what you don't think about

We've done that orchestration layer. On ShipClaw:

Each user shares a pool-node container with ~50 others, but their OpenClaw process and stealth browser are isolated within the container.
We watchdog memory at 1.5 GB and respawn cleanly.
Idle agents suspend after 15 minutes (1 hour for Telegram-connected) so you're not paying for a sleeping browser.
Four Railway projects share the load to clear the 50-volumes-per-project ceiling.
The Anthropic relationship is pooled — we negotiate Tier 4 limits and prompt-caching on your behalf.

You don't size hardware, you don't watch a dashboard, you don't care about the upgrade treadmill when OpenClaw ships a new version. You buy credits and use the agent.

The decision

Self-host if:

You have specific compliance requirements that demand a dedicated machine
You enjoy the operational work and want full control of the stack
Your usage is large enough that the per-token margin we charge dominates the all-in cost (rough crossover: tens of dollars per day in token spend)

Use ShipClaw if:

You want the agent online today
You're not running a fleet
You'd rather pay slightly more per token to never think about Chromium memory leaks again

If you're going self-host, the openclaw GitHub repo has the canonical Docker quickstart. Come back if you decide the operational tax isn't worth it.