VPS vs Mac Mini vs Old Laptop: The Sponsor-Free Math on Hosting a 24/7 AI Agent (July 2026)

Bottom Line (July 2026)

• The harness is not the hard part. OpenClaw or a Hermes agent idles at a few hundred MB of RAM and near-zero CPU between jobs. Hosting it is a cheap problem that tutorials make sound expensive.
• Where the model runs is the real decision. API vs local dominates your bill by an order of magnitude. Pick that first, then pick the box.
• VPS (~$5–10/mo): about $72–120 for 12 months, no hardware, no home power draw. It is a rented computer — your secrets sit on someone else’s disk and the model has to be an API.
• Mac mini (~$599): about $613 for 12 months including power, and it can also serve small local models. Resale recovers a large chunk if you quit.
• Old laptop (~$115 used): about $125 for 12 months. Cheapest working setup. One commenter runs a 24/7 Hermes agent on a £90 smashed-screen M1 MacBook Air.
• Raspberry Pi class (~$80): fine for harness-only. It will not run a useful local model, so budget for API tokens.

Why Every Tutorial Ends at the Same Signup Page

Search for how to run an agent 24/7 and you will watch six videos that all install it on the same VPS provider. The comments have worked this out:

“Feels surprisingly like a commercial for hostinger.”

“everybody is sponsored by hostinger and only shows how to install using their service.”

And the one that actually explains the situation, sitting at 32 likes:

“Hostinger isn’t necessary and neither is openrouter… A VPS is just another computer somewhere else.”

That last line is the whole article. There is nothing special about a rented Linux box. The agent harness does not know or care whether the kernel it is running on lives in a datacenter or under your desk.

Separate the Two Decisions

The single most expensive mistake here is treating “where do I host the agent” and “where does the model run” as one question. They are independent:

You can run the harness on a $80 Pi and call a frontier API. You can run it on a Mac mini and serve a local model from the same machine. The combinations are free to mix, and the model side is where the money is — see the real token cost of running a Hermes agent and hybrid routing between local and frontier models for that half of the problem.

Everything below is about the harness half.

Option 1: Cheap VPS ($5–10/mo)

What you get. An always-on machine with a static IP, a real uptime SLA, and no hardware in your house. Nothing to unplug when you move apartments. If your home internet dies at 3am, the agent keeps working.

What it actually costs. At $6/mo you are at $72 for 12 months, and $72 again the year after, forever. There is no equity — you are renting.

The catches nobody mentions in the sponsored version:

• The model has to be an API. A 2–4 GB RAM VPS cannot run anything useful locally. So the “cheap” option quietly commits you to a metered token bill on top of the $72.
• Your secrets live off-site. The agent needs credentials to be useful — email, calendar, repos, possibly your bank. On a VPS those tokens sit on a disk you do not control, on a host that can snapshot the VM. That is a real risk, not a theoretical one.
• Latency to home devices. If the agent’s job involves anything on your LAN — a NAS, Home Assistant, a printer, a local model server — you now need a tunnel back into your house anyway, which erases most of the simplicity advantage.

Pick this if you travel a lot, you do not want a machine at home, or your agent’s work is entirely cloud-to-cloud. Setup specifics are in deploying OpenClaw to a VPS.

Option 2: Mac Mini or a Dedicated Small Box (~$599)

What you get. A one-time purchase that is silent, sips power, and can double as a local model server. The data stays in your house. When you stop caring about agents, you sell it or use it as a desktop.

What it actually costs. $599 up front. Power is the part people wildly overestimate: an M-series mini averages roughly 10 W running an idle-ish agent workload, which is 87.6 kWh a year, which at $0.16/kWh is about $14/year. Fourteen dollars. It is not a GPU rig — a 3090 pulling 350 W under load is a completely different conversation, covered in the local LLM electricity break-even math.

The community math on this is blunt:

“buying a mac mini is the cheaper option in the long run.”

That is true but load-bearing on “long run.” Year one, the mini costs more than the VPS. It gets cheaper from month 30 onward at $6/mo, sooner at $10/mo, and much sooner once you count that the mini can serve a small local model and cut your API bill to near zero for routine tasks.

The catches: your home internet and power are now the uptime story, and you own the failures. A UPS is $60 and worth it.

Pick this if privacy matters, you want the option of local models, or you plan to keep doing this for more than a year. Setup: running OpenClaw on a Mac mini and keeping the gateway alive with launchd.

Option 3: Old Laptop or Used Machine (~$0–150)

The cheapest setup that genuinely works is the machine already in your closet. The receipt from the comments, at 66 likes:

“£90 smashed-screen M1 MacBook Air off eBay… runs 24/7 Hermes agent.”

A broken screen is a discount, not a defect, on a headless server. You SSH in. An M1 Air idles around 7 W, so figure $10/year in power and roughly $125 all-in for the first 12 months — cheaper than the VPS in year one and every year after.

The caveats are real but small:

• Sleep settings. A laptop’s default is to suspend on lid-close, which kills your agent silently at 2am. On macOS, sudo pmset -a disablesleep 1 and set sleep 0. On Linux, mask the suspend targets. Do this before you trust it with anything.
• Thermals. Closed lid on a soft surface is how you cook a machine. Give it a hard surface and airflow. Elevated on a wire rack is ugly and fine.
• Battery. A used battery is an unknown-quality UPS. It might give you 4 hours through a power cut; it might swell. Check it, and do not put a swollen battery in a closet.
• Reliability. Consumer laptops are not built for continuous uptime. Assume it dies eventually and keep your agent’s config in git so a rebuild is 20 minutes.

Raspberry Pi class (~$80) is the same story with less risk and less capability: perfectly adequate for a harness that calls an API, hopeless for local inference. See running an agent on a Pi for where that ceiling actually sits.

12-Month TCO, With the Assumptions Written Down

Assumptions, so you can substitute your own: electricity at $0.16/kWh (roughly the US residential average — check your bill, it varies 2–3x by state), continuous operation for 8,760 hours, VPS at $6/mo, hardware prices at July 2026 street prices, resale estimated at 12 months. Power draw figures are average agent-idle workloads, not peak.

None of these numbers include the model. If you route to a frontier API, your token bill will likely exceed every row in this table within the first month of real use. That is the point: the hosting argument everyone is having is an argument about the smallest line item.

Two more things the table hides. The VPS row repeats every year at full price while the hardware rows drop to just power — by month 30 the mini is ahead of a $6/mo VPS and much further ahead of a $10/mo one. And the hardware rows can eliminate a chunk of the token bill by serving a small local model, which the VPS row structurally cannot.

Decision Guide

• Privacy matters, or the agent touches your accounts and files → home box. A Mac mini if you want it to also serve models; anything you already own if you do not.
• You travel, or you do not want hardware at home → VPS. Accept that the model is an API and put real effort into secret handling.
• Budget is the constraint → old laptop. A $115 used machine does the same job as a $599 one for harness-only work.
• You want to try this for a week before committing → the machine on your desk, sleep disabled. Zero dollars, and you will learn what you actually need before buying anything.
• You want local models too → this stops being a hosting question and becomes a RAM question. Start at what local LLM fits my machine and the honest OpenClaw/Hermes hardware requirements.

Getting To It From Outside the House

If you host at home, you need to reach the agent from your phone. There are two paths and only one of them is a good idea.

Do this: put the machine on a private overlay network — Tailscale, or a Cloudflare Tunnel. Your phone and the agent join the same private network, nothing is published to the internet, and there is no port to scan. It is a 10-minute setup and it is free at personal scale.

Do not do this: forward port 3000 on your router to the agent gateway. An agent gateway is a much more attractive target than a normal web app, because it holds credentials and can execute tools. If it is reachable, it will be found. If you must expose something publicly, put authentication in front of it and read the OpenClaw security checklist first — and consider running the agent in a Docker sandbox regardless of where it lives.

The tradeoffs between reaching in and hosting out are covered in more detail in remote vs in-home OpenClaw setups.

Related Guides

• The real token cost of a Hermes agent — the line item that actually dominates your bill
• Deploy OpenClaw to a VPS — if you pick the rented box
• OpenClaw on a Mac mini and keeping the gateway alive with launchd — if you pick the home box
• Honest OpenClaw/Hermes hardware requirements — what you need if the model runs locally too
• Local LLM electricity break-even — the power math when a GPU is involved
• OpenClaw security checklist — before you expose anything
• Self-hosted vs cloud OpenClaw — the wider version of this decision