The question that decides server size is not "which agent?" but "where does the model think?". An agent such as OpenClaw or Hermes Agent that sends its questions to a model over an API runs comfortably on KVM 1 at $6.49 a month; a language model you run on the server itself with Ollama needs far more memory, and the sensible floor starts at KVM 4.
This is the difference most excited tutorials skip: the same words, "AI", describe two workloads that have nothing in common when it comes to resources.
Both workloads in one table
| Agent calling an API | Local model (Ollama) | |
|---|---|---|
| Where the thinking happens | On the model provider's servers | On your server |
| Memory needed | Little | A lot — the model's size and then some |
| CPU | Barely used | Maxed out on every reply |
| The bill | Small server + API usage bill | Big server, no per-word bill |
| Answer quality | The providers' large models | Usually smaller, weaker models |
| Speed | Fast | Relatively slow on a CPU with no GPU |
| Starting plan | KVM 1 | KVM 4 |
Prices on this page are read from our table the moment you open it, not typed into the text. Check the comparison page for the latest figures before buying.
Ready templates in hPanel
When you set up a KVM server, Hostinger offers one-click application templates, and these now include OpenClaw, Hermes Agent and Ollama. A template saves you the installation step, so you start from a working interface rather than an empty command line.
But a template solves installation, not sizing. Pick the Ollama template on a small server and it will install fine — then on the first question you discover the model does not fit in memory, or that a reply takes minutes. Choose the plan first, then the template.
Template details and names can also change between panel releases; what matters stays the same: what kind of workload will you run?
Case one: an agent working through an API
OpenClaw, Hermes Agent and their kind are programs that run on your server around the clock: they take your messages from a chat app or an interface, carry out tasks, and keep context. But when they "think", they send the request to a model at an outside provider and wait for the answer.
Why does a small server do? Because the server's job here is waiting and coordinating, not heavy computation. KVM 1 with 4 GB of RAM carries one agent comfortably.
When should you move to KVM 2?
- You run more than one agent, or one agent per client
- You let the agent browse the web with a real browser on the server — the browser alone takes noticeable memory
- You add n8n, a database or a website beside it
And the cost you do not see in the server price: the API bill. An agent running constantly and doing long tasks can burn more tokens than you expect. Set a spending limit in the provider account before you leave it running on its own.
Case two: a local model with Ollama
Here the maths flips. Ollama runs the model itself on your server, so the whole model is loaded into memory and the CPU computes every word of the reply.
The fact to know before buying: Hostinger's KVM servers have no graphics card (GPU). The model will run on the CPU alone. That works, but it is much slower than the demos you see on GPU machines.
The rough memory rule: the model file's size plus headroom for the system and the context. Models are usually distributed compressed (quantized), which shrinks them, but a bigger model is still smarter and heavier.
| Model (roughly) | What it needs | Realistic plan |
|---|---|---|
| Very small models (up to a few billion parameters) | Moderate memory | KVM 2 for testing |
| Common mid-size models (around 7–8 billion) | Comfortable memory with headroom | KVM 4 |
| Larger models (over 10 billion) | Lots of memory, and patience | KVM 8 |
These are guideposts, not guarantees; real usage varies with the model, its compression and the context length. Test the exact model you want before you build a service on it.
When is a local model worth it?
- Your data is sensitive and you do not want it to leave your server
- Your tasks are batch jobs that need no instant reply: summarising documents overnight, sorting messages, extracting data
- Your volume is so high that the monthly API bill exceeds the price gap between servers
And when is it not?
- You want fast chat at the quality of large models — an API is faster, smarter and cheaper for that
- You expect GPU performance from a CPU — you will be disappointed
- Your usage is light — a big server running all month for a handful of questions is waste
Prices side by side
| Plan | CPU | RAM | Per month | Renewal | |
|---|---|---|---|---|---|
| KVM 1 | 1 core | 4 GB | $6.49 | $11.99 | Offer |
| KVM 2 | 2 cores | 8 GB | $8.79 | $14.99 | Offer |
| KVM 4 | 4 cores | 16 GB | $12.99 | $28.99 | Offer |
| KVM 8 | 8 cores | 32 GB | $25.99 | $49.99 | Offer |
Prices are for a 24-month term, paid upfront: KVM 4, for example, is $311.76 today. The renewal column is the price after the first term. The KVM 4 table, every term:
| Term | Per month | You pay today |
|---|---|---|
| 2 years | $12.99 | $311.76 |
| 1 year | $14.99 | $179.88 |
Security: an agent on a server is an employee with keys
An agent that runs commands and connects to your accounts holds real permissions. Before you leave it running:
- A separate server for the agent, not the one carrying your clients' sites. One mistake by the agent should not take down a site someone pays you for.
- Least privilege: do not give it keys it does not need, or delete rights where read access will do.
- API keys stored safely, not in plain text that gets copied along with files.
- A spending cap at the model provider.
- Backups of its settings and memory; Hostinger's weekly backup is a start, not a full plan — see website backups.
Who should not get a VPS for this at all
- Anyone who just wants a personal chat assistant: the models' official apps do that with no server and no upkeep.
- Anyone who never wants a command line: the template installs, but updates and fixes need some comfort with servers. Start with what is a VPS to see what you take on.
- Anyone expecting GPU speed from a local model: KVM servers are not built for that.
Quick answers
Can the agent and n8n share one server? Yes, on KVM 2 and up, as long as the model runs over an API. We size n8n in the best VPS for n8n.
Is shared hosting enough for an AI agent? No. An agent is a background service that is always on, and shared hosting does not allow that. See shared hosting vs VPS.
Can I start small and upgrade? Yes, and upgrading between KVM plans does not require moving the server. It is also the cheapest path: do not pay for local-model memory before you know you need it.
Is the cheapest server on the market enough? If all you need is lots of memory at the lowest price, compare with the best cheap VPS, including Contabo VPS S at —. Mind the difference in support and ready templates.
The short answer
An agent working through an API (OpenClaw, Hermes Agent): start on KVM 1 at $6.49 a month, and move to KVM 2 when you add agents or services. See the offer
A local model with Ollama: do not go below KVM 4 with 16 GB if you want a mid-size model to run acceptably. See it with the discount
Before buying: compare every plan on the comparison page.
Read next: What is web hosting? · The best hosting company · Cheapest web hosting · Hosting contract terms · Compare plans
Disclosure: some links on this page are affiliate links. They cost you nothing extra, and nobody buys a ranking here with them — how we work.