What is AI agent hosting? VPS vs managed servers, and do you need a GPU?

By the serverserv team · 8 min read · Updated 15 October 2026

AI agent hosting is a server that keeps an AI agent, such as OpenClaw, running around the clock, so it can answer you in chat apps and act on tasks while your own computer is off. Most agents need no GPU, because the AI model runs at a provider such as Anthropic, OpenAI or Google. This guide compares a plain VPS with pre-installed and managed agent servers, then covers sizing and monthly cost.

What does AI agent hosting actually mean?

AI agent hosting means renting a server whose job is to run agent software 24 hours a day. The server holds the agent, its settings, its memory and its connections to your chat apps. The AI model that does the thinking usually runs somewhere else.

A working agent setup has four parts:

• The agent runtime. The program that stays running and routes messages. OpenClaw, a free agent released under the MIT licence, runs as a long-running Node.js service it calls the Gateway.
• The AI model. A hosted model you reach with an API key or a subscription sign-in (Anthropic, OpenAI, Google, OpenRouter, Amazon Bedrock and others), or a local model run with a tool such as Ollama.
• Channels. Where you talk to the agent. OpenClaw supports 29 channels, including WhatsApp, Telegram, Slack, Discord, Signal, Microsoft Teams and Google Chat.
• Tools and skills. What the agent is allowed to do: browse the web, read files, call APIs or run scripts.

The host supplies the server for the first part and a stable, always-on place for the other three to connect. You bring the model account and decide what the agent may touch.

How is AI agent hosting different from regular VPS hosting?

Under the hood, most AI agent hosting is a VPS. The difference is how much is done for you: who installs the agent, who keeps it updated and how much control you keep over the server.

Option	What you get	Who installs it	Your control	Typical price at the time of writing
Plain VPS	An empty Linux server with root access	You	Full	From about $6 a month for 1 vCPU and 1 GB RAM (the size OpenClaw's own DigitalOcean guide uses)
Pre-installed agent server	A VPS with the agent installed and set up	The host	Usually full server access	About $12 to $16 a month at the regular price for 2 to 4 GB RAM
Managed agent platform	The agent as a service, often with a simplified dashboard	The provider	Limited, often no server access	Hostinger managed OpenClaw $10.99 monthly, Elestio from $16, Cloudways $9.99 to $79.99

Pick a plain VPS if you are comfortable with SSH and the Linux command line. Pick a pre-installed server if you want your own machine without the setup work. Pick a managed platform if you never want to see a server and can accept the platform's limits.

Check the renewal price as well as the headline price. Many hosts advertise an intro rate that only applies to a long prepaid term. Hostinger, Bluehost and IONOS, for example, list OpenClaw plans that renew at roughly two to three times the first-term price.

Do you need a GPU to host an AI agent?

No, not when the AI model runs in the cloud, which is how most people use agents. The agent server sends requests to the model provider and handles the replies, so an ordinary CPU server is enough.

OpenClaw's documentation lists a minimum of 1 GB RAM, 1 CPU core and 500 MB of disk. Its Raspberry Pi guide notes that the Pi "is just the gateway" because the models run in the cloud through an API.

A GPU only matters if you want to run the model itself on your own server. OpenClaw can use local models through Ollama, LM Studio, vLLM or SGLang. On a CPU-only server this works, but it needs far more memory (16 GB or more is a common starting point) and replies are much slower than a cloud API. For most personal and small-team agents, a cloud model on a CPU server is the faster and cheaper choice.

serverserv plans run on CPU, which suits agents that use cloud models.

Why not run the agent on your own computer?

You can, and many people start that way. A server becomes the better choice once you rely on the agent, because it stays online when you do not.

• It does not sleep. An agent on a laptop stops answering when the lid closes or the Wi-Fi drops. Scheduled tasks and incoming messages wait until you are back.
• It keeps the agent away from your personal files. An agent with broad permissions on your own computer can reach everything you can. A separate server limits what a mistake or a bad instruction can touch.
• It avoids home network problems. No router changes, no power cuts at home taking the agent down, and no fan running all night.

One exception: iMessage needs a Mac. If that channel matters to you, OpenClaw lets a Linux VPS pair with a Mac as a "node", so the main agent can still live on the server.

How much server does an AI agent need?

For one agent using a cloud model, 2 to 4 GB of RAM and 1 to 2 vCPU covers most personal use. Memory needs climb quickly with browser automation and local models.

How you use the agent	RAM	vCPU
Testing, one chat channel	1 to 2 GB	1
Daily personal use, several channels	4 GB	2
Browser automation (about 2 GB per headless browser)	8 GB	2 to 4
Local model on the same server, CPU only	16 GB or more	4 or more

These bands come from OpenClaw's documentation and third-party sizing guides. Vultr, for example, suggests 2 vCPU and 2 GB for personal use and 4 vCPU and 8 GB for multi-channel use.

Can you run several agents on one server? Yes, if they serve the same person or team. OpenClaw's security model is "one trust boundary per gateway", so agents for different people or clients who should not see each other's data belong on separate gateways, and ideally separate servers. For general sizing advice, see how much RAM and CPU a VPS needs.

What does AI agent hosting cost per month?

You pay two bills: one for the server and one for the AI model. The agent software itself is often free. OpenClaw is MIT-licensed and has no paid edition or official hosted tier.

• The server. Entry agent servers with 2 to 4 GB RAM cost about $12 to $16 a month at regular prices at the time of writing. Intro offers of $2 to $6 exist, but they renew higher.
• The model. Your provider bills you directly, usually by usage. OpenClaw can also sign in with a ChatGPT or Codex subscription, so an existing subscription can cover the model. A local model avoids API bills but needs a much larger server.
• Extras. A domain if you want one, and your own time if you install and maintain everything yourself.

Cost guides for OpenClaw consistently find that model usage, not the server, is the part that varies most. Set a spending limit or usage alerts in your provider account before you connect the agent to busy channels.

What to check before you choose an AI agent host

Compare hosts on the things that affect you after the first month, not just the headline price.

1. RAM for your use case. Match it to the sizing table above.
2. The renewal price. Ask what you will pay in year two.
3. Server access. Root or SSH access lets you add tools, check logs and move the agent later.
4. Updates. OpenClaw releases new versions close to weekly. Find out whether you or the host applies them.
5. How the agent is exposed. OpenClaw's VPS guide advises keeping the Gateway on loopback and reaching it through an SSH tunnel or Tailscale Serve, with SSH hardened first. Run openclaw security audit after setup.
6. Skills from trusted sources only. Researchers have found a malicious third-party skill that exfiltrated data, so read a skill before you install it.

Which serverserv plan fits an AI agent?

For one personal agent on a cloud model, start with OpenClaw Starter. Move up only when you add browser automation, more channels or more people.

• OpenClaw Starter: 2 vCPU, 4 GB RAM, 50 GB NVMe, OpenClaw pre-installed and set up. $14.99 monthly, or $10.99/mo on a 3-year term.
• Personal Use: 4 vCPU, 8 GB RAM, 100 GB NVMe. For browser automation or several busy channels. $24.99 monthly, or $17.99/mo on a 3-year term.
• Do it yourself: a Linux KVM VPS with full root access, from KVM Mini (1 vCPU, 2 GB RAM, $8.99 monthly).

Every plan renews for the same term at the price you paid. If a price ever has to change, you get at least 30 days' email notice and can cancel before renewal. See all OpenClaw plans on the OpenClaw hosting page. If a plan is not in the order form yet, open a ticket and we will set it up. Not sure an agent is the right tool? Compare it with workflow automation in OpenClaw vs n8n.

Questions

Sources

• OpenClaw official site
• OpenClaw on GitHub (MIT licence)
• OpenClaw docs: Raspberry Pi install and hardware minimum
• OpenClaw docs: running on a VPS
• OpenClaw docs: Gateway security

Prices from other companies are as published at the time of writing and can change.