The Cheapest Way to Run OpenClaw 24/7, and What Breaks

You can run OpenClaw around the clock for $6.49 a month, all in — server and model included, nothing pirated, nothing that stops working next week.

Two honest caveats before we go further. That $6.49 is a promotional rate on a 24-month prepaid term; the steady-state number is $11.99. And 4GB is the floor rather than a comfortable amount — a truly complete agent wants more, for reasons the rest of this article is about.

This stack works. It also breaks in three specific ways, and knowing which one you’ll hit first is worth more than the $5 you’re saving.

Why 4GB and not less

You’ll see $2–4 plans advertised with 1–2GB. For OpenClaw specifically, don’t.

The agent’s own baseline sits around 1.5–2GB resident. That leaves almost nothing on a 2GB box for the OS, let alone anything the agent actually does. It’ll appear to work, then die the first time it does something interesting.

4GB is the honest floor for an API-backed agent, and it’s the tier this article prices. If your agent is text-only, never touches a browser and runs one channel, you can get away with less — but the saving is about $4 a month and the failure mode is silent.

The free model tier, and its real limit

Groq’s free tier is the part that makes $6.49 possible. No credit card, and it serves genuinely capable open models — Llama 3.1 8B, Llama 3.3 70B, Llama 4 Scout, GPT-OSS 120B and others.

The published limits are 30 requests per minute, 6,000 tokens per minute, and 14,400 requests per day.

Most write-ups quote the request limits. For an agent, the tokens-per-minute figure is the one that binds, and it binds much sooner than people expect, because an agent turn is not a short prompt. It carries a system prompt, tool definitions and conversation history on every single call.

A realistic OpenClaw turn with a handful of tools registered lands in the 4,000–10,000 token range. So the free tier gives you roughly one agent turn a minute — fine for a background assistant that reacts to messages, useless for anything conversational or for a multi-step tool loop that needs ten turns to finish a task.

One mitigation worth knowing: cached tokens cost half and don’t count toward rate limits. If your system prompt is stable, caching it materially raises your effective ceiling. That’s the single highest-value change you can make to this stack, and it’s free.

If the free tier throttles you, moving to paid usage on a small open model costs a couple of dollars a month rather than tens — what OpenClaw costs to run breaks down where token spend actually goes, and the answer is usually the heartbeat interval rather than your actual usage.

What breaks at $6.49, and what fixes it

1. Browser automation is the wall

This is the one that gets everyone, and it’s not gradual.

A headless Chromium instance wants 1–2GB entirely to itself. Agent baseline plus browser plus OS overshoots 4GB the moment a page loads. The kernel’s out-of-memory killer then terminates something, and — this is the cruel part — OpenClaw’s own log shows nothing, because the process didn’t decide to stop. It was killed.

If your agent vanishes mid-task with a clean log, that’s an OOM kill, not a bug.

The fix is 8GB, and it costs $3.00/month at renewal rates — KVM 1 at $11.99 to KVM 2 at $14.99. If browser automation is part of your plan at all, buy this tier at the outset rather than migrating later. Three dollars is not worth the afternoon.

2. One vCPU is thin for tool loops

The cheapest tier gives you a single shared core. For an agent that mostly waits on API responses, that’s genuinely fine — the work is I/O-bound and the CPU is idle.

It stops being fine during a long tool loop: the agent shells out, parses a large response, drives a script, and does it repeatedly. On one shared core those steps queue behind each other, and shared vCPUs are contended — you get some fraction of a core, varying with what your neighbours are doing.

The fix is the same upgrade as above. KVM 2 gives 2 vCPU along with the 8GB, so the $3.00 buys both. That’s why it’s the upgrade worth making: it fixes the two most likely failures at once.

A caveat I want to be straight about: we have not measured steal time on any of these providers. That shared vCPUs are contended is a property of the product, not a finding of ours, and how much it affects you depends on your neighbours and the host’s overcommit ratio.

3. Backups — and the brief for this article was wrong here

I expected to write that the cheap tier gives you no snapshots, so one bad config costs you the instance. That’s the conventional wisdom and it’s not accurate for Hostinger.

Its VPS plans include free automated weekly backups and one free snapshot — you can only hold one snapshot at a time, but you have one. Take it before you change anything significant and the “bad config destroys everything” scenario largely goes away.

What’s not included is daily backups, at $6.00/month. That’s a real cost and the largest delta in the table, which makes it the one to think hardest about: a week of lost agent state matters if the agent holds meaningful memory, and doesn’t if it’s stateless.

The free move is the one to make first: take a snapshot before you touch the config. It costs nothing, and it’s the single habit that most reduces the risk of the cheap tier.

What it costs to fix everything

Buy the 8GB tier and add daily backups and you’re at $20.99/month at renewal rates, against $11.99 for the bare stack.

That’s the honest top of this article’s range, and it’s still less than most managed alternatives. But note the shape: $3.00 fixes the two failures that will actually happen to you, and $6.00 fixes the one that might. Spend the three dollars; decide on the six based on whether losing a week of state would matter.

The providers

Hostinger is the pick and the basis for every figure here — the cheapest 4GB tier at $6.49 promotional and $11.99 at renewal, an 8GB tier $3 above it, weekly backups and a snapshot included, and one-click templates that skip the install. On price per gigabyte it leads our whole VPS index among plans you can actually buy.

Cloudzy is worth knowing for its billing rather than its price — month-to-month with no term, so nothing is stranded if you abandon the project. Its 2GB Advanced is $8.97 and 4GB Professional $17.37 monthly, so you pay a premium for that flexibility.

UltaHost sits mid-field, and unusually its discounted 2-year rate is also its renewal rate rather than a step-up — $15.87 for a 4GB, 3-vCPU plan. More cores than Hostinger’s entry tier, more money.

AccuWeb Hosting is a long-standing budget option worth pricing at the specific tier you need; we haven’t verified its current figures at source for this article.

Hetzner would normally be the value answer, and currently isn’t: its cheap CX tier has been showing as unavailable, and the CPX line you’d fall back to is €19.99 for 4GB after a June 2026 repricing. Check stock before assuming.

For the install itself, self-hosting OpenClaw on a VPS does it with hardening first — which matters more on a cheap box than an expensive one, because an exposed agent costs the same to compromise either way. And if you’d rather not run it at all, managed OpenClaw vs self-hosted prices the alternative.

How we checked this

Hostinger’s KVM 1 and KVM 2 pricing ($6.49 and $8.79 promotional, $11.99 and $14.99 at renewal, on 24-month terms), Cloudzy’s monthly-billed rates and UltaHost’s 2-year rates are the providers’ published figures, read in August 2026 for this series.

Groq’s free-tier limits — 30 requests per minute, 6,000 tokens per minute, 14,400 requests per day, no credit card required, with cached tokens costing half and not counting toward rate limits — are from published documentation of that tier rather than from Groq’s own console, which we could not access without an account.

Hostinger’s backup terms — weekly automated backups and one snapshot included, daily backups at $6.00/month — are from its own support documentation and plan pages.

A correction to this article’s brief. It assumed the cheapest tier comes with no snapshots, making a bad config catastrophic. Hostinger includes weekly backups and one snapshot on all VPS plans, so that premise doesn’t hold there. The genuine gap is daily backups, and we’ve priced that instead.

The turns-per-minute table is arithmetic, not measurement. It divides Groq’s published 6,000 tokens-per-minute limit by plausible agent turn sizes. We have not instrumented an OpenClaw instance to measure real turn sizes, and yours will depend on how many tools you register and how long your history grows. The direction is reliable — an agent turn is large, so TPM binds before RPM — but treat the specific numbers as illustrative.

On steal time: we have not benchmarked shared vCPU contention on any provider here, and we won’t quote figures we didn’t measure. The claim is only that shared cores are contended by design.

What we did not do: we haven’t run this stack for a month, hit the free tier’s limits in practice, or tested an OOM recovery. The failure modes are drawn from the documented resource requirements rather than from our own incident log.

The hosting links above are affiliate links. Groq’s free tier, snapshots before config changes, and prompt caching — the three things that most improve this stack — cost nothing and earn us nothing.

FAQ

What’s the cheapest way to run OpenClaw?

A 4GB VPS at $6.49/month promotional and a free model API tier, for $6.49 all-in. The steady-state figure once the promotional term ends is $11.99.

Can I run OpenClaw on a $5 VPS?

Close to it. The constraint isn’t the price but the memory — 4GB is the honest floor, and the cheapest 4GB plans start around $6.49 promotional. Cheaper 1–2GB plans will appear to work and then get killed by the kernel.

Is a free LLM API tier enough for OpenClaw?

For a background agent, often yes. Groq’s free tier allows 6,000 tokens per minute, which works out at roughly one agent turn a minute once you count system prompt, tool definitions and history — fine for reactive work, too slow for conversation.

Why does my cheap OpenClaw server keep dying?

Almost certainly the out-of-memory killer, and almost certainly the moment browser automation launched. A headless Chromium wants 1–2GB by itself, which 4GB can’t spare. The upgrade to 8GB costs about $3/month.

Do cheap VPS plans include backups?

Hostinger’s include weekly automated backups and one snapshot at no extra cost; daily backups are $6/month more. Don’t assume this across providers — check before you rely on it.

How much does it cost to fix everything that breaks at the cheapest tier?

$20.99/month at renewal rates, against $11.99 bare. But $3.00 of that fixes both failures you’re likely to hit; the remaining $6.00 buys daily rather than weekly backups.

Should I pay for a bigger plan up front?

If browser automation is anywhere in your plan, yes — buy 8GB at the outset. Migrating an agent later costs more of your time than the $3/month difference ever will.