AI Agent VPS Hosting in 2026: Running OpenClaw on Your Own Server
The new AI-agent VPS category, sizing, security and cost, compared.
Your AI assistant needs a place to live
Sometime in the last year, the question changed. It used to be "which chatbot should I subscribe to?" Now it's "where do I run my agent so it actually works while I sleep?" That's a hosting question, and the VPS industry has noticed.
OpenClaw, the open-source self-hosted AI assistant formerly known as Clawdbot and Moltbot, links the messaging apps you already use, WhatsApp, Telegram, Slack, Discord, Signal, even iMessage, to whatever model you bring: Anthropic Claude, OpenAI, Gemini, or anything reachable through OpenRouter. It answers messages, runs scheduled tasks, browses, automates. None of that works on a laptop that closes at midnight.
So the agent needs an always-on box. In September 2026 that need became a product category: providers started selling "AI agent VPS" plans by name. This guide walks through what's actually on the market, what hardware an agent really needs, how to secure it, and what it costs compared with rolling your own.
What OpenClaw does to a server
The OpenClaw Gateway is a Node.js process that owns two things: state and workspace. Your agent's memory, config, conversation history, installed skills, all of it lives in files on whatever machine runs the Gateway. The official docs are blunt about the implications.
- The VPS is the source of truth. Lose the box without backups and your agent's memory is gone. The docs tell you to back up state and workspace regularly, not as an afterthought.
- You connect from anywhere. Laptop, phone, Control UI, usually over Tailscale or an SSH tunnel. The Gateway itself stays bound to loopback by default.
- API keys live on the server. Anthropic, OpenAI, Gemini, OpenRouter keys sit in the Gateway config. That makes key hygiene and spend caps a server-administration task, not just an app setting.
None of this is exotic. It's a long-running Node process with a filesystem and an internet connection. But the always-on part changes the hosting math, and the keys-and-memory part changes the security math.
The new agent-ready VPS plans
ServerHost OpenClaw VPS
ServerHost announced its OpenClaw VPS line on September 3, 2026, and it's the most concrete sign the category has arrived. Two platforms: Xeon Linux for steady always-on agents, and Ryzen/i9 "CPU Optimized" for heavy browsing and automation. Every plan ships NVMe storage, a 1 Gbps unmetered port, full root KVM access, and support for Anthropic, Gemini, OpenAI and OpenRouter keys. Ready-to-Go builds are typically online in under 15 minutes.
- Starter: 2 GB RAM, 1 Xeon core, 30 GB NVMe, $8/month
- 4 GB: 2 cores, 60 GB NVMe, $16/month
- Power: 8 GB RAM, 4 cores, 120 GB NVMe, $30/month
- Heavy Duty: 16 GB RAM, 4 cores, 240 GB NVMe, $56/month
The September promotion on LowEndTalk pushed the annual numbers down hard: the 4 GB OpenClaw plan at $49.99/year with code OCSEPT4GB, 8 GB at $79.99/year, 16 GB at $129.99/year. That's roughly $4.17/month for 4 GB, which undercuts most general-purpose clouds by a wide margin. Read the terms before committing, promo pricing and recurring pricing are different things.
ClawVPS: the managed alternative
ClawVPS, from the ServerAvatar team, attacks the same problem from the opposite end. No sysadmin work at all: the agent deploys in about five minutes, updates and patches are handled for you, backups run daily, and a one-click factory reset restores a broken instance to a fresh state with the same config. Chrome comes pre-configured so the agent can browse from its first prompt, and "Elevated Mode" grants it full root inside its own sandbox. There's a 24-hour free trial with $1 of API credits, plus $10 in credits on signup, and it supports 42 model providers with no token markup, you bring your own keys.
The pitch is honest: DIY on Hetzner or DigitalOcean is cheaper on paper, but costs you 4-10 hours of setup and a couple hours a month of maintenance. Whether that trade is worth it depends entirely on how much your time costs.
How much machine does an agent actually need
Here's where marketing and reality diverge. The Gateway itself is not heavy. What eats resources is everything around it: headless browsers for web tasks, embeddings and memory indexing, multiple concurrent skills, and the occasional model-context binge.
- 2 GB: fine for one personal agent doing messaging, reminders, light API calls. You'll feel the squeeze the first time it opens a browser.
- 4 GB: the sweet spot for most people. Agent plus headless Chrome plus breathing room. This is where ServerHost's $49.99/year promo lands, and it lands there for a reason.
- 8 GB and up: multi-agent setups, heavy scraping, local tooling alongside the Gateway, or a shared team agent.
CPU matters less than you'd think until it suddenly matters a lot. A 2 GHz Xeon core handles message round-trips fine; browser automation and page rendering want high clock speeds, which is exactly why ServerHost sells the Ryzen/i9 tier for $11/month and up. The official docs also note that CLI commands can feel slow on low-power VMs and ARM hosts, and suggest enabling Node's module compile cache to compensate.
Storage is the quiet one. Agent memory grows. Skills accumulate. Browser caches balloon. 30 GB fills up eventually; 60 GB is comfortable; NVMe is non-negotiable for snappy context loads.
The DIY route: generic clouds work fine
You do not need a specialized plan. The official docs list Azure, DigitalOcean, Fly.io, GCP, Hetzner, Hostinger, Northflank, Oracle Cloud, Railway, and even a Raspberry Pi as deployment targets. AWS EC2 and Lightsail work well too. Any KVM VPS with 2 GB+ RAM and root access will do.
- Hetzner: shared vCPU instances start around EUR 4-6/month with generous NVMe. The default European budget pick.
- Oracle Cloud Always Free: up to 4 ARM OCPUs and 24 GB RAM for nothing. Overkill specs, with the usual caveats about capacity availability and account reclaim.
- DigitalOcean / Vultr: from roughly $4-6/month, hourly billing, mature APIs. Good if you already run infrastructure there.
- Budget annual providers: RackNerd-class deals around $20/year get you 2-3 GB KVM. Fine for a hobby agent, but you're accepting single-node consumer-grade reliability for something you've tasked with running your digital life.
Rule of thumb: the more the agent touches your real accounts, your email, your files, your payments, the less you should optimize for the cheapest possible host.
Security: the part most guides skip
The official documentation spends most of its VPS page on hardening, and that's the right instinct. An always-on agent holding your API keys and your chat channels is a juicy target.
- Keep the Gateway on loopback. Access it through an SSH tunnel or Tailscale Serve. If you bind it to lan or tailnet instead, the Gateway requires a shared secret, gateway.auth.token or gateway.auth.password, unless auth is delegated to a trusted proxy.
- Harden SSH before anything else. Install Tailscale first, join the box to your tailnet, verify a second SSH session over the tailnet IP, then restrict public SSH. No Tailscale? Apply equivalent hardening to your SSH path before exposing services.
- Dedicate the runtime. For a shared team agent, keep it on a dedicated VPS with its own OS user. The docs are explicit: never sign that runtime into personal Apple, Google, or password-manager accounts. If users might be adversarial to each other, split them by gateway, host, or OS user.
- Cap your API spend. Set provider-side spending limits on every key stored on the server. A compromised or hallucinating agent with an uncapped key is a bill, not a breach.
- Back up state and workspace. This is the agent's brain. Rsync to object storage, restic snapshots, whatever you'll actually run on a schedule. Daily is the floor.
One more from the docs worth repeating: you can pair nodes on your local devices with a cloud Gateway. The gateway stays in the cloud as the source of truth while your Mac or phone provides screen, camera, and local run capabilities. That hybrid keeps the always-on part cheap and the trusted-hardware part local.
What it costs: three honest tiers
Running the numbers for a personal agent kept online for a year:
- Free-ish: Oracle Always Free ARM, $0, with availability and reclaim risk. Raspberry Pi at home, hardware you already own, plus electricity.
- Budget self-host: ServerHost OpenClaw promo at $49.99/year (4 GB), or a Hetzner-class cloud at roughly EUR 5/month. Call it $50-70/year plus API tokens, which you'd pay regardless of host.
- Managed: ClawVPS or similar. Higher monthly cost, but setup in minutes, daily backups, patching handled, humans on support. You're paying to never SSH into anything.
Notice what's missing from every tier: token costs. The VPS bill is the small line item. A month of moderate Claude or GPT usage will dwarf a year of $8/month hosting. Choose your host for reliability and security, not for the last two dollars of savings, because downtime and a leaked key both cost more than the spread between any two providers on this list.
Before you deploy: a ten-minute checklist
- Pick sizing honestly: 2 GB if it's messaging-only, 4 GB once a browser is involved, 8 GB for multi-agent or team use.
- Create the VPS with a fresh OS user; disable root password login on SSH.
- Install Tailscale, verify tailnet SSH, then lock public SSH down.
- Deploy the Gateway bound to loopback; expose nothing publicly.
- Add API keys with provider-side spend caps already set.
- Schedule backups of state and workspace, and test a restore once.
- Run the agent for a week with low-stakes tasks before you hand it anything that matters.
The category is new. The discipline isn't. An AI agent on a VPS is still a daemon holding credentials, and everything we learned in two decades of running daemons holding credentials applies. Get the box right, get the keys capped, get the backups running, then let the agent earn your trust one scheduled task at a time.