Hermes Agent VPS Cost: Does the '$5 VPS' Claim Hold Up?

The “$5 VPS” claim isn’t a forum rumour — it’s in Nous Research’s own README, which says Hermes Agent can run on “a $5 VPS, a GPU cluster, or serverless infrastructure.”

That’s a wide range, and the low end is doing a lot of work in people’s heads. So: does it hold?

Mostly, yes — and for a more interesting reason than you’d expect. The thing everyone blames for breaking small servers, the Camofox anti-detection browser added in v0.7.0, is one of the lightest browsers you could pick. What actually kills a $5 box is a configuration default that permits five times more memory than the machine has.

What you’re actually running

Hermes Agent is an MIT-licensed agent framework from Nous Research, built around a closed learning loop: complete a task, auto-generate a skill document from what worked, reuse it next time. Nous describes it as “the agent that grows with you.” Alongside that it does persistent memory, conversation search and multi-platform messaging.

The install pulls uv, Python 3.11, Node.js, ripgrep and ffmpeg. Docker Compose files ship in the repo. None of that is heavy.

What matters for sizing is that Nous publishes no hardware requirements at all. There is no minimum RAM figure, no vCPU count, no disk recommendation in the documentation. Every “Hermes Agent VPS requirements” table you’ll find online — including the ones quoting confident numbers — is somebody’s inference. This one included. The difference is that we’ll show the arithmetic.

If you’re still deciding whether you want an always-on agent at all, what an AI agent actually is covers why this turns into a hosting question so quickly.

The three tiers, and what fails at each

2GB is genuinely viable for a text-only agent. The model runs on someone else’s hardware; your box orchestrates. That’s the whole reason this category is cheap, and it’s why you don’t need a GPU either.

4GB is the honest floor for a complete agent — one that can also click things on the web.

8GB is where you stop thinking about it. Whether that’s worth the money depends entirely on the section after next.

The Camofox correction

Here’s the claim I set out to test, because it’s repeated across nearly every guide on this keyword: the Camofox browser in v0.7.0 adds roughly 2GB, so the real floor is 4GB.

That misreads the source. The 2GB figure comes from the Camofox browser server’s system requirements, which state “2GB+ RAM (browser + contexts require significant memory).” That’s a total system floor for running the browser service — not an increment you add to whatever else you’re running.

And Camoufox, the engine underneath it, is unusually light. Its own documentation describes it as “headless-first, <200MB footprint,” and puts it at “~200MB vs Chrome’s 800MB+.” It’s Firefox-based rather than Chromium-based, and that’s precisely why.

So the browser is roughly a quarter the weight of the headless Chromium that most agent stacks reach for. If anything, adding browser automation to Hermes is cheaper in memory terms than adding it to a Playwright-based setup.

The naming confuses this further, and it’s worth untangling once:

• Camoufox is the anti-detect browser engine, originally by daijro, now maintained by Clover Labs.
• Camofox Browser Server is a REST API that wraps that engine for AI agents, with OpenClaw-compatible endpoints and MCP support.
• Hermes v0.7.0 integrates the second one — its release notes list “Camofox browser — Camoufox-based anti-detection browser with VNC debugging.”

One u, two different things, and the memory figures attached to them are not interchangeable.

The setting that actually kills small boxes

This is the part no sizing guide mentions, and it’s the answer to why people’s cheap servers die.

The Camofox browser server ships with CAMOFOX_MAX_SESSIONS set to 50 and CAMOFOX_MAX_TABS set to 10.

Fifty concurrent sessions at roughly 200MB each is about 10GB of permitted allocation — on a machine you bought with 2GB in it.

The default isn’t wrong — it’s sized for a dedicated browser host, which is a perfectly reasonable thing for that project to assume. It’s wrong for you, and nothing warns you, because a ceiling only matters on the day you approach it.

The failure mode is the nasty kind. You don’t get an error from Hermes. The kernel’s out-of-memory killer picks a process and terminates it, and the agent’s own log shows nothing at all — because the agent didn’t decide to stop. If your agent vanishes mid-task with a clean log, that’s an OOM kill, not a bug in the software.

Set both variables down before you need them. A personal agent doing sequential browsing wants a handful of sessions, not fifty. This single change does more for stability on a small server than doubling the RAM.

One genuinely thoughtful default in that server, worth crediting: it implements a two-stage idle lifecycle, cleaning up and then exiting the daemon after inactivity — 30 minutes per stage by default. So an idle browser isn’t holding memory indefinitely. Idle isn’t your problem. Bursts are.

What each tier costs

Prices checked in August 2026, on each provider’s own pricing page.

Every one of those numbers has a condition attached, and in a cost article the conditions are the article.

Hostinger’s are 2-year promotional rates, paid upfront. KVM 1 renews at $11.99/mo and KVM 2 at $14.99/mo, and the site states plainly that “all plans are paid upfront” with the monthly rate being the total divided by the term. So $6.49/mo is really $155.76 today for 24 months. It’s a genuinely good price; it is not a $6.49 monthly commitment. Hostinger also has no 2GB tier — its entry plan is 4GB, which for this workload is the one you wanted anyway.

Cloudzy’s headline prices are the annual rate. The pricing page defaults to a “Pay Yearly −50%” toggle, and the figures people quote from it — $7.48 for Advanced, $14.48 for Professional — are that annual rate. Switch to monthly billing, which is Cloudzy’s actual selling point, and it’s −40% instead: $8.97 for 2GB and $17.37 for 4GB. Those are the numbers in the table above, because month-to-month with no commitment is the reason to choose Cloudzy in the first place.

Hetzner’s include the IPv4 address. The plan itself is €11.49, €19.49 and €35.49; a primary IPv4 is €0.50/month on top, and the price on the pricing page is the combined figure. Prices are ex-VAT at its German locations, and they’re monthly caps — billing is hourly, so a server you delete mid-month costs less.

DigitalOcean’s are simply the price, month to month, no term and no promotion. Worth noting given everything above.

Why the usual Hetzner advice is now stale

If you’ve read anything about cheap agent hosting written before this summer, it told you to buy Hetzner. That advice needs updating, and this is the most time-sensitive thing on this page.

Hetzner raised cloud prices on 15 June 2026, and not gently. CPX plans went up 2.4× to 2.75× in Germany and Finland and up to 2.98× in the US; CCX rose 2.1× to 2.73×. CPX32 went from €13.99 to €35.49. Existing subscriptions weren’t grandfathered — the new pricing applies to new orders and to rescales, so resizing an old server re-prices it.

The CX and CAX lines took a much softer 1.3×–1.4× increase, and they remain by far the best value here: a CX23 is 2 vCPU and 4GB for €5.99, which would undercut everything else in this article.

It is also, along with every other plan in that tier, marked not available.

That’s not a glitch. Hetzner’s own explanation is that the cost-optimized tier runs on “proven hardware generations that continue to be used,” and that this “resource-saving use means that the number of available servers is limited.” Cheap because it’s leftover capacity, and sometimes there isn’t any.

So the practical position today: the Hetzner plans you can actually buy for this workload are the repriced CPX ones, and a 4GB CPX22 at €19.99 is roughly three times a 4GB Hostinger KVM 1 at $6.49. That inverts the recommendation this category has run on for years. If the CX line comes back in stock, it inverts straight back — check availability before you buy rather than trusting either of us.

The providers

Hostinger is the straightforward answer right now, and unusually it’s both the cheapest and the easiest — its 4GB KVM 1 is the tier this workload wants, and one-click AI templates mean you’re not assembling a Python 3.11 environment by hand. Budget the $11.99 renewal.

Hetzner still gives you the best hardware per euro once you’re above these tiers, and its network and NVMe storage are genuinely excellent. For a 4GB agent box in August 2026, it is no longer the cheap pick, and pretending otherwise would be repeating advice rather than checking it.

Cloudzy earns its place on billing terms rather than price. At $17.37 for 4GB month-to-month it costs more than Hostinger’s promo — but nothing is paid upfront, there’s no term, and there’s a 14-day money-back guarantee. That’s the right shape for “I want to find out whether I’ll still be running this in December,” which is the honest state of most people deploying an agent framework for the first time. Compare $17.37 you can stop against $155.76 you’ve already spent, and it isn’t obviously the more expensive option.

DigitalOcean is the reference point everyone knows and is the most expensive option here at every tier — $24 for 4GB against Hostinger’s $6.49. You’re buying documentation, predictable API behaviour and a decade of Stack Overflow answers. That’s worth something, just not usually four times. Its tracked link isn’t working on our end, so we’ve named it without linking.

For the full provider comparison rather than a cost-focused one, see best VPS for Hermes Agent. If price is the only axis you care about, cheapest VPS for AI agents goes lower still.

The verdict

The smallest configuration that survives a week without intervention is 4GB — and buy it even if you’re starting text-only, because the gap between 2GB and 4GB is a few dollars and the gap between “it works” and “it was killed at 3am” is the whole point of paying for hosting.

At $6.49 promotional and $11.99 at renewal, the $5 VPS claim is close enough to true that arguing about it is a waste of time. Nous’s README is honest. What it can’t tell you is that the software’s own defaults will happily try to allocate 10GB on the box you bought.

The one caveat I’d put on the cheapest route: $6.49/mo at Hostinger means $155.76 today. If you’re not yet sure you’ll still be running this agent at Christmas, Cloudzy’s $17.37 month-to-month is the cheaper mistake, and a first agent deployment is very often a mistake.

Lower CAMOFOX_MAX_SESSIONS, buy 4GB, and the cheap answer is the right answer.

One cost this article doesn’t cover: the model API. For most people it’s the larger half of the bill, and the mechanism is identical across agent frameworks — we broke it down in what a self-hosted agent costs to run. And before you expose any of this, securing a VPS that runs an agent covers a genuinely privileged piece of software.

How we checked this

Last verified against Hermes Agent v0.7.0, August 2026. This is fast-moving software and the browser integration is new, so re-check the figures against your version.

The description of Hermes, its MIT licence, the Python 3.11 and Node.js requirement, the installer’s payload and the “$5 VPS, a GPU cluster, or serverless infrastructure” line are from the Nous Research repository, read in August 2026. The v0.7.0 feature list — pluggable memory providers, approval UI, inline diff previews, session persistence, the Camofox browser with VNC debugging, and credential pool rotation — is from release coverage of that version.

The “2GB+ RAM (browser + contexts require significant memory)” requirement, the CAMOFOX_MAX_SESSIONS default of 50, the CAMOFOX_MAX_TABS default of 10 and the two-stage 30-minute idle lifecycle are from the Camofox browser server’s own documentation. The “<200MB footprint” and “~200MB vs Chrome’s 800MB+” figures are Camoufox’s own published claims.

Provider pricing was read off each provider’s own pricing page in August 2026, and the screenshots above are those pages as we saw them. Two details we checked rather than assumed, because both are easy to get wrong and both are wrong in most write-ups of this keyword:

• Cloudzy’s pricing page defaults to its annual toggle. The widely-quoted $7.48 and $14.48 are the “Pay Yearly −50%” rates. We switched the toggle to monthly to get $8.97 and $17.37, and used those, since month-to-month is the reason to pick Cloudzy.
• Hetzner’s displayed price bundles the IPv4 address. Its plan data lists CPX12, CPX22 and CPX32 at €11.49, €19.49 and €35.49, with a primary IPv4 at €0.50/month; the page figure is the sum. We’ve quoted the sum because that’s the bill.

Hostinger’s renewal figures and the “all plans are paid upfront” condition are its own disclosure on that page. The Hetzner increase — effective 15 June 2026, CPX up 2.4×–2.75× in Germany and Finland and up to 2.98× in the US, CPX32 from €13.99 to €35.49, no grandfathering of existing subscriptions — is from published analysis of that change, and the current CPX figures we quote match Hetzner’s own pricing data. The unavailability of the CX and CAX tiers, and the explanation about limited hardware, are what Hetzner’s cost-optimized page itself displayed when we checked. That one may well have changed by the time you read this.

Two corrections to the brief for this article. It asserted that Camofox “adds roughly 2GB,” making 4GB the floor. The 2GB figure is the browser server’s total system requirement rather than an increment, and Camoufox is roughly four times lighter than Chromium — so the conclusion happens to be defensible while the reasoning behind it isn’t. It also assumed a 1 vCPU / 2GB documented floor for a bare install; Nous publishes no hardware requirements at all, and we’ve said so rather than inventing one.

What we did not do: we have not deployed Hermes Agent and measured its resident memory across workloads. The tier table is reasoning from the published component figures — the browser server’s stated floor, Camoufox’s stated footprint, and the session caps — not from htop on a running instance. Anyone quoting you a precise Hermes baseline is doing the same thing with less transparency about it. We also haven’t tested how the self-improving loop’s accumulating skill documents affect disk use over months, which is a real cost dimension none of this coverage addresses.

The hosting links above are affiliate links. Hermes Agent is free and MIT-licensed.

FAQ

Can Hermes Agent really run on a $5 VPS?

For a text-only agent calling a hosted model, yes — that claim is Nous Research’s own and it holds. Buy 4GB rather than 2GB anyway: at current pricing it’s a few dollars more and it’s the difference between running and being killed by the kernel.

How much RAM does Hermes Agent need?

Nous publishes no requirement. Reasoning from components: 2GB works for text-only, 4GB is the honest floor once the Camofox browser is enabled, and 8GB covers browser work at volume. Treat any more precise figure — including ours — as inference.

Does the Camofox browser need 2GB of extra RAM?

No. The 2GB figure is the browser server’s total system floor, not an increment. Camoufox itself is about 200MB per instance against Chrome’s 800MB+, because it’s Firefox-based and headless-first.

Why does my Hermes Agent keep dying with no error?

Almost certainly the out-of-memory killer. The likeliest cause on a small server is the browser session cap: CAMOFOX_MAX_SESSIONS defaults to 50, which at ~200MB each permits far more memory than the machine has. Lower it.

Is Hetzner still the cheapest option for Hermes Agent?

Not as of August 2026. Hetzner raised CPX pricing up to 2.75× on 15 June 2026 and its cheaper CX tier is currently showing as unavailable, which puts a 4GB CPX22 at €19.99 against Hostinger’s $6.49. Check availability yourself — this one changes.

Do I need a GPU to run Hermes Agent?

Not unless you run the model locally. The agent orchestrates on your server while the model runs on the provider’s hardware, which is exactly why a small VPS is enough.

How much does the model API cost on top of hosting?

Usually more than the server. Hosting is the predictable half of the bill; token spend depends on your model choice and how often the agent wakes itself up, and that’s where the surprises live.