How to Run an AI Agent on a VPS for £5 a Month

Quick Answer

You can run a basic OpenClaw agent on a cheap VPS (virtual private server) for around five pounds per month. The setup involves renting a small Linux server, installing Ollama with a lightweight model, deploying OpenClaw via Docker, and connecting it to a messaging channel like Telegram or WhatsApp. The agent runs 24/7 without tying up your personal computer, giving you an always-on AI assistant at the price of a coffee.

Why a VPS?

A VPS solves the biggest problem with running a local AI agent: uptime. When your agent runs on your laptop, it stops working when you close the lid, restart for updates, or take it somewhere without internet. A VPS runs around the clock in a data centre with reliable power and connectivity. Your agent is always available, no matter what you are doing with your own computer.

VPS hosting has become absurdly cheap. Providers like Hetzner, Contabo, and DigitalOcean offer capable Linux servers starting at three to five pounds per month. These are not powerful enough for large AI models, but they can absolutely run small models that handle everyday agent tasks.

The trade-off is clear: you get 24/7 uptime and reliability, but you are limited to smaller, less capable models compared to what you could run on a beefy desktop or a managed platform like Tulip.

Choosing a VPS Provider

For running an AI agent on a budget, you want a VPS with at least 8GB RAM (16GB is better), two or more CPU cores, 80GB or more of storage, and a location near you for low latency.

Hetzner offers some of the best value in Europe. Their CPX21 plan gives you 4 vCPUs, 8GB RAM, and 80GB storage for about four euros per month. Their CPX31 with 8 vCPUs and 16GB RAM runs about eight euros and is significantly more comfortable for AI inference.

Contabo is another budget option with generous specs. Their VPS S plan offers 6 vCPUs, 8GB RAM, and 200GB storage for under five pounds. The servers are not the fastest, but they are reliable and the specs are hard to beat at this price.

DigitalOcean and Vultr are slightly more expensive but offer better performance and more polished management interfaces. Their 8GB RAM droplets run about 30 to 40 pounds per month, which is more than our five pound target but still very affordable.

For the true budget setup, Hetzner or Contabo are the way to go.

Step 1: Set Up Your Server

Once you have your VPS, connect to it via SSH. Most providers give you a root login immediately after creating the server. The first thing to do is basic security: create a non-root user, set up SSH key authentication, disable password login, and enable the firewall. This takes about 10 minutes and protects your server from automated attacks.

Install Docker, which you will use to run OpenClaw. Docker keeps everything contained and makes setup much simpler. Most VPS providers offer Ubuntu 22.04 or newer, which has straightforward Docker installation.

Step 2: Install Ollama and Pick a Model

Install Ollama on your VPS. The installation is a single command on Linux. Once installed, pull a small model that fits within your RAM constraints.

With 8GB RAM, your best options are quantised 3B to 7B models. Phi-3 Mini is excellent for its size — remarkably capable for a 3.8B model and very fast on limited hardware. Gemma 2 at 2B is even smaller and runs comfortably on 8GB, though it is less capable. Qwen 2.5 at 7B (quantised to 4-bit) fits in 8GB RAM and offers better quality at the cost of slower inference.

With 16GB RAM, you can run 7B-8B models comfortably and even squeeze in some 14B quantised models. Llama 3.2 at 8B or Qwen 3 at 8B are both solid choices that handle tool calling reasonably well.

Remember: on a VPS without a GPU, all inference runs on the CPU. This means responses will take 5 to 15 seconds for short answers and longer for complex tasks. It is not instant, but it is perfectly usable for an asynchronous assistant that you message and check back on.

Step 3: Deploy OpenClaw

With Docker installed and Ollama running, deploy OpenClaw using the official Docker image. Configure it to use Ollama as its model provider by pointing to the local Ollama endpoint.

Set up your SOUL.md file with the personality and instructions you want. Keep it concise on a budget VPS — shorter system prompts mean fewer tokens to process and faster responses. Focus on the most important instructions and skip anything nice-to-have.

Configure OpenClaw to start automatically when the server boots. Docker's restart policy handles this — set the container to restart unless stopped, and your agent will come back online automatically after server reboots or crashes.

Step 4: Connect a Messaging Channel

Telegram is the easiest channel to set up on a VPS because it requires no additional infrastructure. Create a bot through Telegram's BotFather, get the API token, and add it to your OpenClaw configuration. Within minutes you will be messaging your agent through Telegram.

WhatsApp is also possible but requires a bit more setup, typically involving the WhatsApp Business API or a bridge service. Discord is straightforward — create a bot application, get the token, and configure. Slack works similarly through a Slack app.

For the simplest possible setup, start with Telegram and add other channels later.

Step 5: Install Lightweight Skills

On a budget VPS, be selective about which skills you install. Each skill adds context to every interaction, consuming tokens and slowing down the model. Start with two or three essential skills and add more only if you genuinely need them.

Good lightweight skills for a VPS agent include a web search skill (gives your agent access to current information), a timer or reminder skill (lets your agent remind you about things), and a note-taking skill (lets your agent store and recall information for you).

Avoid heavyweight skills that require significant processing or large responses on a budget server. Save those for a more powerful setup.

Step 6: Set Up Auto-Restart and Monitoring

You want your agent to stay online without babysitting. Docker's restart policy handles crash recovery. Set up a simple health check script that pings your agent periodically and alerts you (via email or Telegram) if it goes down.

Also configure log rotation so your server's disk does not fill up with agent logs over time. Docker can handle this with a logging configuration that limits log file size and count.

What You Can Expect

A five pound VPS agent is not going to match the performance of a powerful desktop or cloud deployment. Response times will be measured in seconds rather than milliseconds. Complex reasoning and multi-step tool calling will be slower and occasionally less reliable. Very long conversations may strain the limited context window of smaller models.

But for an always-on personal assistant that can answer questions, look things up, set reminders, take notes, and handle basic tasks through Telegram? It is remarkably useful. Many people find that having an agent always available — even a slower one — is more valuable than a fast agent that is only available when their laptop is open.

When to Upgrade

If you find yourself wanting faster responses, more capable models, or more complex skills, you have several upgrade paths. A more powerful VPS with 16-32GB RAM gives you room for better models at 15 to 30 pounds per month. A dedicated GPU server from providers like Hetzner or RunPod gives you fast inference but costs significantly more. Or you can move to a managed platform like Tulip, which handles all the infrastructure and gives you access to large, powerful models without managing servers.

The VPS approach is a brilliant way to get started cheaply and learn what you want from an AI agent. Once you know your needs, you can make an informed decision about where to run your agent long-term.

Frequently Asked Questions

Is five pounds per month really enough?

For basic agent tasks with a small model, yes. You will not be running Llama 4 Maverick, but you can run a 3B-7B model that handles everyday assistant tasks. The main limitation is speed, not capability.

Can I run multiple agents on one VPS?

Technically yes, but each agent and model instance uses RAM. On an 8GB VPS, one agent with one model is the practical limit. With 16GB, you might fit two lightweight agents.

What about security? Is my data safe on a VPS?

A properly configured VPS with SSH key authentication, a firewall, and up-to-date software is reasonably secure. Your data is stored on the VPS provider's infrastructure, so choose a reputable provider. For maximum privacy, consider encrypting the disk or using a provider with strong privacy policies.

How does this compare to running on Tulip?

Tulip costs more but gives you access to much larger, more capable models without managing any infrastructure. The VPS approach trades capability and convenience for cost savings. If your budget allows, Tulip is the easier path. If you want the cheapest possible always-on agent, a VPS is hard to beat.