Cheap VPS for AI Agents: What You Actually Need in 2026

The Real Cost Question Is About Specs, Not Brands

The phrase "AI agent" suggests a hungry workload, but most self-hosted agents are not running the model. They are calling an API. The model lives on OpenAI, Anthropic, OpenRouter, or your local Ollama box, and your VPS is doing something much smaller: holding the conversation state, running a Telegram or Slack gateway, and orchestrating tool calls.

That changes the buying decision entirely. You do not need a GPU. You do not need 32 GB of RAM. You need a small Linux box that stays online and does not get killed when memory pressure spikes. Most popular agent runtimes, including Hermes Agent, will run comfortably on 2 GB of RAM and a couple of vCPUs.

This post breaks down what you actually need, then compares the four cheapest options for the job in 2026.

What an API-Driven Agent Actually Uses

The minimum baseline for a personal AI agent that calls an external LLM API and serves one or two messaging channels:

CPU: 1 to 2 vCPU is enough for the runtime, the messaging gateway, and a small process supervisor.
RAM: 1 GB works for lightweight runtimes; 2 GB gives you headroom for occasional headless browser tasks or a second agent process.
Disk: 20 to 40 GB SSD or NVMe is plenty. Memory files, skill files, and logs do not grow fast.
Bandwidth: 1 to 2 TB/month is far more than you will use; the agent's network traffic is mostly small JSON payloads to the model provider.

Hermes Agent's documented minimum, for reference, is 2 GB of RAM and 10 GB of disk. If you plan to run two agents, attach a headless browser, or experiment with self-evolution loops that batch run skills overnight, bump RAM to 4 GB.

If you plan to run the model itself locally with Ollama, Mistral, or Llama, this entire conversation changes - you need a GPU box or a CPU server with 16 to 32 GB of RAM, and you should be reading a different guide. Everything below assumes the model is remote.

Specs trade-off illustration showing a small VPS box connected to remote model APIs

The Four Cheapest Real Options in 2026

Pricing changes, but the relative positioning of these four providers has been stable for years. All quotes below are May 2026.

Hetzner Cloud (Best Performance per Euro)

Hetzner's CX22 is the default recommendation in self-hosting communities and it earns it. The plan ships 2 vCPU, 4 GB RAM, 40 GB NVMe, and 20 TB of traffic, and as of the April 2026 price adjustment it sits at €3.79/month for EU regions. US-based US-East and US-West regions are slightly more, around $4.59 to $4.99/month equivalent.

Why it stands out:

Real benchmark performance under sustained load. The vCPUs do not get oversold the way they do at some lower-cost competitors.
Snapshots, backups, and a clean API are included. Spinning up, destroying, and rebuilding the agent is a one-command operation.
Hourly billing. If you destroy the box on day 7 of testing, you pay one-fourth of the monthly rate.

Trade-off: account verification can take hours to a day, and Hetzner sometimes asks for ID for new accounts. If you need a server in five minutes, this is not it.

Contabo (Most RAM per Dollar)

Contabo's entry-level VPS 10 lists at €3.60/month on a 12-month contract (around $4.95 USD), and bundles 4 vCPU and 8 GB RAM. On paper, that is two to four times the spec of Hetzner's CX22.

In practice:

The CPU cores are oversubscribed more aggressively than at Hetzner. Sustained CPU benchmarks usually favor Hetzner's "smaller" 2 vCPU plan.
Provisioning can take minutes to hours after payment.
The 12-month commitment locks the price; monthly contracts cost more.

For an API-driven agent that mostly idles between messages, Contabo's specs are overkill and the extra RAM is real. If your workload is bursty and CPU-bound (running scheduled skill batches, doing local data processing), the Contabo CPU oversubscription will hurt you more than the RAM helps.

Hostinger VPS (Easiest Setup)

Hostinger's VPS offering ships templates for OpenClaw and other open-source AI runtimes. The starter plan begins around $6.99/month on a 24-month commitment (renewing at $14.99/month after the initial term).

What you trade:

A locked-in multi-year contract for the discounted rate.
Less control over the runtime stack - templates assume a specific layout.
A fast first-deploy experience. If the template covers your runtime, you can be talking to the bot in 10 minutes.

Hostinger fits the user who would rather pay a small premium than read a deployment guide. For Hermes Agent specifically, the template ecosystem is thinner than for OpenClaw, so a manual install is usually faster than waiting for a template that may not exist.

Oracle Cloud Always Free Tier (Best Free Option)

Oracle's Always Free Tier is the only one of the four that costs nothing. The Ampere A1 ARM shape gives you up to 4 OCPUs and 24 GB RAM, configurable as one VM or split across up to four. You also get 200 GB block storage and 10 TB of outbound traffic per month.

The catch is real:

Always Free capacity is not always available in your region. New accounts often see "out of capacity" errors for weeks before a slot opens up.
Account verification requires a credit card. Oracle does not charge it, but the friction is real.
ARM-only on the free shape. Most agent runtimes ship ARM builds, but verify before you commit.
Oracle has been known to reclaim idle Always Free instances. Set up a small heartbeat so yours does not get flagged as inactive.

If you can tolerate the setup pain, this is the cheapest real option for a 24/7 personal agent. Many self-hosters quietly run their entire homelab on Oracle's free tier.

Decision Framework

| Your Situation | Pick | |---|---| | Want the best performance-per-dollar ratio | Hetzner CX22 | | Want the most RAM and CPU cores on paper | Contabo VPS 10 | | Want a one-click install and don't mind a longer commitment | Hostinger VPS | | Want it free and don't mind setup pain | Oracle Always Free Tier | | Want zero infrastructure work at all | Hermify managed hosting |

If you have never deployed a Linux service before, Hostinger or managed hosting is genuinely faster. If you have, Hetzner is the default and Oracle Free is the free fallback.

Photorealistic dark server room with a single rack illuminated by a green status LED

What You Actually Spend Once the Agent Is Running

The VPS is rarely the largest line item. The model API is. A rough monthly budget for a personal agent that handles a moderate volume of Telegram messages, scheduled tasks, and occasional skill execution:

VPS: $0 (Oracle) to $7 (Hostinger entry)
Model API (Claude Sonnet, GPT-4o, or similar): $3 to $25 depending on volume
Domain name (optional, only if you need a public webhook): ~$1
Telegram, Discord, Slack bot tokens: free

Total: roughly $5 to $35/month for a fully self-hosted personal agent. If your provider offers BYOK and you connect to OpenRouter or a cheaper provider, the API line drops further. The detailed breakdown for Hermes specifically is in our cost guide.

When the Cheap VPS Math Stops Working

There is a point where the cheapest VPS is not the cheapest path. That point is when you start spending hours debugging the deployment instead of using the agent.

Things that flip the math:

Telegram gateway dropping connections at 3 AM and you don't notice until lunch.
Memory volume not mounted correctly, causing the agent to forget everything on every restart.
A package update breaking the runtime and requiring a manual rollback.
Spending a Sunday wiring up systemd, log rotation, and basic monitoring that a managed platform would have given you in 30 seconds.

If those things are interesting to you, run your own VPS. If they sound like a tax on time you would rather spend elsewhere, the comparison shifts. Our hosting vs self-hosting guide walks through the full math, including what your time costs.

Get started with Hermify and have an agent running on managed infrastructure in under 30 minutes, with persistent memory, Telegram already wired up, and BYOK to your model provider of choice. The VPS bill, the systemd service, and the Sunday afternoon of debugging are skipped.

The Recommendation

For most self-hosters in 2026:

Start with Hetzner CX22 at €3.79/month if you have time to set up Hetzner's account verification. Real performance, fair price, no commitment.
Use Oracle Always Free Tier if you are budget-bound and patient. Free is hard to beat.
Use Hostinger if you want a guided install and don't mind the multi-year contract.
Skip Contabo for AI agent workloads unless you genuinely need the extra RAM. Hetzner usually wins on real performance.

If you do not want to think about infrastructure at all, that is what managed hosting exists for. The right answer is the one that matches how you want to spend your time.