The cheapest VPS for AI agents — what you actually need

The measurements

Real servers, our own one-click installs, memory read thirty seconds after a live model request:

The “shape” column matters more than the numbers. Claude Code and Codex are invoked per session and leave nothing running in between, so an idle box sits at its operating-system baseline. OpenClaw and Hermes hold a process open — which is precisely what lets them work while you are away, and what costs the 120–278 MB.

Why so little?

Because the hard part happens somewhere else. The model runs on a provider’s hardware; your server assembles a request, waits on the network, and applies the result. It is an I/O-bound errand-runner, not a compute job — which is why RAM matters far more than vCPUs here, and why doubling your CPU count buys you almost nothing.

The exception is running models locally. That moves inference onto your box and changes the requirement by orders of magnitude — realistically a GPU. If that is your plan, none of the numbers on this page apply to you.

The uncomfortable part: this is not the bill that matters

We sell servers, so take this in the spirit it is offered: for a single agent doing real work, token spend usually dwarfs the server. The gap between a $2 plan and a $5 plan is rounding next to your model bill.

So optimising the server is worth doing when you are running many agents, or when the box is doing something beyond hosting an agent. If you are hunting the absolute cheapest plan to save a few dollars a month, you are optimising the smaller number — tiering your models will save you more, and it is a config change.

What to look for in a cheap VPS

• RAM before vCPUs. 1 GB is the measured floor for a CLI agent; 2 GB if you want a web workspace. Extra cores are largely wasted on an I/O-bound workload.
• Disk is usually the real constraint. Agents are small, but Docker images, a package cache and logs are not. 10 GB gets tight faster than you expect.
• Beware “unlimited” anything. Very cheap plans are often heavily oversold; the number on the pricing page is not the throughput you get at 9pm.
• Check the bandwidth allowance. Agent traffic is small, but a runaway loop against an API is not.
• You need a real IP and root. Shared hosting will not run any of this.

Our plans, for what it is worth

These are the plans that include one-click agent installs:

Being straight about it: our measurements say a CLI agent runs on less than our cheapest agent-enabled plan provides. We size that entry plan for headroom — an agent plus a database, or a build, or the thing you add three weeks later — not because the agent needs 4 GB. If all you want is one agent on the smallest possible box, plenty of providers sell a 1 GB machine that will run it, and now you know it will.

What you get from us instead is the install being done for you, HTTPS and a subdomain configured, credentials surfaced in a dashboard, and the ports bound so the thing is not open to the internet. Whether that is worth a couple of dollars is genuinely your call.

Method

Each agent was installed on a fresh DigitalOcean server through our own provisioning, verified end to end, and measured thirty seconds after a live model request, then destroyed. The figures are memory used by the agent’s own processes, not whole-system usage. Two honest limits: Claude Code and Codex were installed and verified but not driven through a task, so their figures describe an idle box rather than one mid-session; and everything was measured on one provider’s hardware, so a heavily oversold $2 box elsewhere may behave worse at the same specification.

Related

• The comparison matrix — sort agents by smallest server.
• Which model to run — the bill that actually matters.
• Managed vs self-hosted — what running it yourself really costs.