An AI agent that only works while your laptop is open is a toy. The version people actually want answers a Telegram message at 3 a.m., finishes a long task after you close the lid, and keeps working while you sleep. That is exactly what OpenClaw is built for — and the only way to get it is to host OpenClaw on a VPS that never sleeps.
OpenClaw is a free, open-source personal AI agent. You give it your own model key (Claude, OpenAI, Gemini, or OpenRouter), connect it to Telegram, WhatsApp, or Discord, and it runs as a long-lived gateway on your server — answering messages, running tools, and doing real work around the clock. In November 2026 it has become the default way people run a personal agent, and a whole market of managed OpenClaw hosts has sprung up around it: Hostinger, Cloudways, xCloud, Agent37, and a handful of purpose-built platforms.
So should you pay one of them, or run it yourself on a $5–$10 VPS? This guide answers that honestly — with real November 2026 prices, the full install in about 30 minutes, and the security steps most tutorials skip. By the end, your agent will be online 24/7, and you will know exactly what it costs you every month, including after the promo price ends.
- What Is OpenClaw — and Why It Needs a Server That Never Sleeps
- How Much VPS Do You Need for OpenClaw?
- Managed OpenClaw Hosting vs Your Own VPS
- The Real Monthly Cost: OpenClaw Hosting Over 36 Months
- Step 1: Pick and Prepare Your VPS
- Step 2: Install Docker and OpenClaw
- Step 3: Connect Your Model and Your First Channel
- Step 4: Give OpenClaw Tools with MCP Servers
- Step 5: Keep OpenClaw Running 24/7
- Secure Your OpenClaw VPS Before You Connect Anything
- Troubleshooting: The 5 Problems Everyone Hits
- FAQ
- The Bottom Line
What Is OpenClaw — and Why It Needs a Server That Never Sleeps
OpenClaw is a personal AI assistant you run on your own hardware. It is not a chatbot website and not an app on your phone. It is one long-running process — a gateway — that holds your chat connections open, talks to your language model, and executes tools on your behalf. You message it on Telegram the way you would message a friend, and it replies, runs commands, checks things, and reports back.
That “long-running” part is the whole point, and it is also the catch. Close your laptop and a locally-run agent dies with it. An agent that is supposed to watch a channel, finish a task overnight, or reply while you are away needs a machine that is always on, always connected, and always reachable. That machine is a VPS.
A VPS fits OpenClaw better than almost anything else you can buy for the job:
- It is always on. A $5–$10 a month VPS has better uptime than your laptop, your home internet, and your patience combined.
- It is isolated. Your agent holds API keys and chat credentials. Running it on a separate server keeps those away from your personal computer and your home network.
- It is cheap for what it does. OpenClaw needs no GPU when you use an API model — the heavy thinking happens at Anthropic or OpenAI. Your server just runs the gateway, which is a light Node.js process.
- You keep root access. Managed agent hosts run your agent in their container, on their terms. On your own VPS you choose the model, the tools, the channels, and the data stays yours.
There is one honest trade-off: you do the setup and the occasional update yourself. That takes about 30 minutes the first time, and about five minutes a month after that. This guide walks through all of it.
How Much VPS Do You Need for OpenClaw?
OpenClaw is light, but “light” depends on what you ask it to do. A chat-only agent idling on Telegram sips resources. The same agent driving a browser, running code, or orchestrating long coding tasks needs real headroom. Size for the workload, not the software:
| What your agent does | vCPU | RAM | Disk | Typical VPS |
|---|---|---|---|---|
| Chat-only (Telegram/WhatsApp replies, reminders, simple tools) | 1 | 2 GB | 25 GB | Cheapest tier of any major host |
| Chat + browser automation + file work | 2 | 4 GB | 50 GB | Hostinger KVM 1, Hetzner CX23, DigitalOcean 2 GB–4 GB |
| Long coding tasks, Claude Code / Codex sessions, n8n alongside | 2–4 | 8 GB | 80–100 GB | Hostinger KVM 2, Hetzner CX33/CAX21 |
| Multiple agents, heavy automation, local models | 4+ | 16 GB | 200 GB | Hostinger KVM 4 or a GPU plan — see our guide to running AI models on a VPS |
Two sizing notes that save people from the two most common mistakes. First, do not buy a 1 GB VPS to build OpenClaw from source — the build can be killed for running out of memory. Use the pre-built Docker image in this guide and 2 GB is fine for a chat agent. Second, RAM matters more than CPU here. The gateway is mostly waiting on the model API, not computing. A 2 vCPU / 4 GB box beats a 4 vCPU / 2 GB box for this job every time.
If you are also planning to run coding agents or automation next to OpenClaw, read our guide to running AI coding agents on a VPS and our VPS sizing guide for AI automation — the 8 GB tier is the sweet spot once real work is involved.
Managed OpenClaw Hosting vs Your Own VPS
November 2026 is the first month you can buy OpenClaw hosting the way you buy web hosting — and the offers arrived all at once. Cloudways made its Managed AI Agents generally available on August 17, 2026. Hostinger now ships a one-click OpenClaw template on its VPS plans plus a fully managed OpenClaw product. xCloud, Agent37, ClawHost, and OpenClaw Launch all sell managed agent hosting. Here is how they actually compare with doing it yourself:
| Option | Type | Starting price | Renewal price | Setup time | Best for |
|---|---|---|---|---|---|
| Hostinger Managed OpenClaw | Managed | $4.99/mo (24-mo term) | $11.99/mo (+140%) | Minutes | Cheapest managed start |
| OpenClaw Launch Lite | Managed container | $3.00/mo (first month) | $6.00/mo (+100%) | Minutes | Trying OpenClaw with zero setup |
| OpenClaw Launch Pro | Managed container | $20.00/mo flat | $20.00/mo | Minutes | Managed, no price jump |
| Cloudways Managed AI Agents | Managed cloud | $9.99–$79.99/mo | Same range | ~2 minutes | Teams already on Cloudways |
| xCloud | Managed panel | $24.99/mo flat (6 GB) | $24.99/mo | ~5 minutes | 24/7 agents with no server work |
| Agent37 Basic | Managed | From $3.99/mo | From $3.99/mo | Minutes | Lowest managed entry price |
| ClawHost | Managed platform | $25.00/mo | $25.00/mo | Minutes | OpenClaw-only managed hosting |
| Your own VPS (Hostinger KVM 1) | Self-managed VPS | $6.49/mo (24-mo term) | $11.99/mo (+85%) | ~30 minutes | Full control, lowest long-term cost |
| Your own VPS (Hetzner CX23) | Self-managed VPS | €5.49/mo excl. VAT | €5.49/mo | ~30 minutes | Lowest raw cost in the EU, no renewal jump |
Prices checked early November 2026; hosting promos change often, so confirm on the provider’s page before you buy.
Read that table the right way. Managed hosting buys you one thing: you never touch a terminal. Updates, restarts, backups, and SSL are handled for you. What you give up is root access and flexibility — a managed container is not your server, you cannot install whatever you want next to the agent, and your agent’s data lives on someone else’s platform.
Your own VPS is the opposite deal. You spend 30 minutes once, and in return you get a real server: OpenClaw plus anything else you want on it — n8n, a second agent, a personal dashboard — for roughly the same money or less. If you are comfortable pasting commands into a terminal, the VPS wins on price and on control. If the words “SSH into your server” make you close the tab, pay for managed hosting and skip the rest of this guide. Both answers are correct; they are just for different people.
One warning that applies to every managed option with a promo price: check the renewal, not the headline. Hostinger’s managed OpenClaw jumps 140% at renewal, from $4.99 to $11.99 a month. OpenClaw Launch Lite doubles after a single month. The next section does the full math, because the cheapest first month is rarely the cheapest three years.
The Real Monthly Cost: OpenClaw Hosting Over 36 Months
AI agent hosting has a renewal trap, and November 2026 is the month it got measured properly. A price check of 17 agent hosting plans on November 8, 2026 found that six of them advertise an introductory rate that jumps 71% to 140% on renewal. The honest comparison is not the first month — it is what you pay over three years:
| Plan | Intro price | Renewal price | 12-month total | 36-month total |
|---|---|---|---|---|
| Hostinger Managed OpenClaw | $4.99/mo | $11.99/mo | $119.76* | $263.64 |
| OpenClaw Launch Lite | $3.00 (month 1) | $6.00/mo | $69.00 | $213.00 |
| OpenClaw Launch Pro | $20.00/mo | $20.00/mo | $240.00 | $720.00 |
| Hostinger KVM 1 (own VPS) | $6.49/mo | $11.99/mo | $155.76* | $299.64 |
| Hostinger KVM 2 (own VPS) | $8.79/mo | $14.99/mo | $210.96* | $390.84 |
| Hetzner CX23 (own VPS) | €5.49/mo | €5.49/mo | €65.88 | €197.64 |
*Intro prices on 24-month terms are paid upfront; the 12-month column shows the prepaid amount. Prices checked November 8, 2026.
Three things jump out. First, OpenClaw Launch Lite is the cheapest way to try managed hosting — $213 over three years — but you get a 1 vCPU / 2 GB container with no root access, and its promo lasts exactly one month. Second, Hostinger’s managed OpenClaw is cheaper over three years than its own KVM 1 VPS ($263.64 vs $299.64), because the managed promo runs 24 months while the VPS promo renews sooner — but the VPS gives you a full server you can use for other things, and the managed plan gives you only the agent. Third, Hetzner is the cheapest honest price in the table: €5.49 a month, no promo, no renewal jump, no surprise in month 25. That is the RAM-shortage reality of late 2026 — see our explainer on why VPS prices are rising — and it is why flat pricing is worth more than a low intro rate.
And one cost that belongs in every honest calculation: the model API is separate. OpenClaw itself is free and open source. What you pay every month, on top of hosting, is your language-model usage — Anthropic, OpenAI, or OpenRouter, billed per token. A chatty agent on Claude Opus can cost more than the server it runs on. Start with a cheaper model (Claude Sonnet or Haiku, or GPT-5 mini) and a spend limit on your API key, then upgrade only if the answers disappoint. That one setting is the difference between a $5 month and a $50 surprise.
If budget is the deciding factor, our cheapest VPS guide and our free VPS hosting guide cover the $0–$5 options, including Oracle’s Always Free tier — though a free tier that can be reclaimed is a poor home for an agent you depend on.

Step 1: Pick and Prepare Your VPS
Buy a Linux VPS — Ubuntu 24.04 is the safest choice, because every command in this guide is written for it — with at least 2 vCPU and 4 GB of RAM. That size runs OpenClaw with browser automation and leaves headroom; drop to 2 GB only for a chat-only agent. Pick a location near you or near your users, and make sure the provider gives you a clean IPv4 address (some cheap tiers are NAT-only, which complicates nothing for OpenClaw but is worth knowing).
Before you install anything, run through the first-boot basics. If you have never set up a server before, our new VPS setup checklist covers them in order: update the system, create a non-root user with sudo, disable password login for root, and set up the firewall. The short version is three commands:
sudo apt update && sudo apt upgrade -y
sudo apt install -y curl git ufw fail2ban
sudo ufw allow OpenSSH && sudo ufw enable
That last line matters more than it looks. OpenClaw’s dashboard listens on port 18789, and the single most common security mistake in OpenClaw hosting is exposing that port to the whole internet. We are not going to. Keep the firewall closed to everything except SSH for now — you will reach the dashboard through an SSH tunnel in Step 5, the safe way.
Step 2: Install Docker and OpenClaw
There are two ways to install OpenClaw: the native install, which needs Node.js 24.16 or newer, or Docker, which carries its own runtime so the host needs nothing but Docker itself. On a VPS, Docker is the better choice — the install is the same on every host, upgrades are one command, and a broken update cannot take your server down with it.
Install Docker Engine with the official convenience script, then add your user to the docker group so you do not need sudo for every command:
curl -fsSL https://get.docker.com | sh
sudo usermod -aG docker $USER
newgrp docker
docker --version
Now install OpenClaw with the official setup script. It pulls the official image (ghcr.io/openclaw/openclaw, mirrored on Docker Hub as openclaw/openclaw), runs the onboarding wizard, writes your gateway token to a .env file, and starts the gateway with Docker Compose:
git clone https://github.com/openclaw/openclaw.git
cd openclaw
export OPENCLAW_IMAGE="ghcr.io/openclaw/openclaw:latest"
./scripts/docker/setup.sh
The setup script will ask you the onboarding questions: which model provider to use, your API key, and which chat channel to connect first. Answer them and the gateway starts. You can check it is alive with the built-in health endpoint:
curl -fsS http://127.0.0.1:18789/healthz
If you would rather skip the script and run the container by hand, the manual version is one command. Note the three details that most copy-pasted Docker commands get wrong: state lives at /home/node/.openclaw inside the container, the port is published on loopback only, and a gateway token is set so the dashboard is never open to strangers:
docker run -d --name openclaw \
--restart unless-stopped \
-p 127.0.0.1:18789:18789 \
-v openclaw-data:/home/node/.openclaw \
-e OPENCLAW_GATEWAY_TOKEN='a-long-random-token-you-invent' \
ghcr.io/openclaw/openclaw:latest
Pin an exact version tag instead of latest for anything you depend on — latest moves under you on the next pull. Tags follow the release calendar (for example 2026.9.3), and the extended-stable channel exists for people who would rather update quarterly than weekly.

Step 3: Connect Your Model and Your First Channel
OpenClaw is bring-your-own-model. During onboarding you pick a provider and paste an API key. Anthropic’s Claude is the default most people start with; OpenAI, Google Gemini, and OpenRouter all work, and you can switch providers later without reinstalling. Two settings deserve your attention on day one:
- Pick the right model tier. The most capable model is also the most expensive per token. A personal agent answering Telegram messages does not need the flagship model — a mid-tier model answers just as well for a fraction of the price. Start mid-tier, upgrade only if you notice worse answers.
- Set a spend limit. Every provider lets you cap monthly API spend. Set one before you connect a public channel. An agent with a bug — or a stranger who finds it — can burn through tokens fast, and a $10 cap turns a disaster into a shrug.
Then connect your first channel. Telegram is the easiest: message BotFather on Telegram, create a bot, copy the token, and give it to OpenClaw during onboarding or later from the dashboard. The dashboard itself lives at http://127.0.0.1:18789/ on your server — open it through an SSH tunnel from your own computer:
ssh -L 18789:127.0.0.1:18789 your-user@your-vps-ip
Then open http://127.0.0.1:18789/ in your browser and paste the gateway token from your .env file. Message your bot on Telegram. If it answers, the hard part is done — your agent is live.
WhatsApp and Discord work the same way, through their own bot credentials. Add one channel at a time and test each before adding the next; when two things break at once you will not know which one did it.
Step 4: Give OpenClaw Tools with MCP Servers
A model that can only talk is a chatbot. A model that can do things is an agent — and the tools come from MCP servers. MCP (Model Context Protocol) is the open standard that lets an agent call external tools: read files, search the web, query a database, trigger an n8n workflow, check a calendar. OpenClaw connects to MCP servers from its config, and each one adds a set of tools your agent can use in conversation.
Start small. Two MCP servers cover most personal-agent jobs: a web search/fetch server so the agent can look things up instead of guessing, and a file or notes server for its own workspace. Add more only when you have a job for them — every tool you add is another thing the agent can misunderstand and another credential to protect.
And be picky about which MCP servers you trust. In November 2026 a German hoster disclosed a breach that started through an MCP server — the attacker was inside for a week before being blocked. That is not a reason to avoid MCP; it is a reason to connect servers from sources you trust, keep them updated, and never give an agent tools it does not need. Our What Is an MCP Server guide explains the protocol, the risks, and the setup in full — including that breach as a lesson.
If you want OpenClaw to run automations on a schedule — a morning briefing, a daily report, a watch on a price — pair it with n8n on the same VPS. Our guide to building an AI automation server with n8n covers that setup, and n8n and OpenClaw share the same sizing tier, so one 4 GB box comfortably runs both.
Step 5: Keep OpenClaw Running 24/7
An agent that stops when you log out is not a 24/7 agent. Three things keep it alive:
1. Restart policy. The --restart unless-stopped flag in the Docker command (or restart: unless-stopped in Compose) brings the container back after a crash or a server reboot. Verify it survives a reboot — reboot the VPS once, wait a minute, and check the health endpoint. Better to learn it works now than to discover it does not at 3 a.m.
2. Log watching. When something breaks, the logs tell you first:
docker logs -f openclaw
docker ps --filter name=openclaw
If the container shows as restarting in a loop, read the last twenty lines of the log before changing anything. Nine times out of ten the answer is in there — usually a bad API key, a full disk, or a config typo.
3. Updates on a schedule, not on a whim. OpenClaw ships often. Update monthly, not daily, and back up first — your whole configuration is the /home/node/.openclaw directory, so copy it before you pull:
docker cp openclaw:/home/node/.openclaw ~/openclaw-backup-$(date +%F)
docker compose pull
docker compose up -d
docker exec openclaw openclaw doctor --fix
The doctor --fix step migrates your state to the new version. If you pinned an exact image tag, change the tag in your Compose file first, then pull. Never update right before you need the agent for something important.
Secure Your OpenClaw VPS Before You Connect Anything
This is the section most OpenClaw tutorials skip, and it is the one that matters most. Your agent holds model API keys, chat bot tokens, and possibly credentials for your email, calendar, and files. It is a juicier target than a static website, so treat it like one.
- Never expose port 18789 to the internet. Bind it to
127.0.0.1and reach the dashboard through an SSH tunnel, as in Step 3. Some one-click templates — including Hostinger’s — put the dashboard on a public IP over plain HTTP; if you used one, close it off with the firewall and an SSH tunnel before you connect anything real. - Set a long gateway token. The dashboard must never be reachable without one. Rotate it if you ever suspect it leaked.
- Use a non-root user and SSH keys only. Disable password login, disable root login over SSH, and install fail2ban. Our VPS security guide walks through all of it step by step.
- Give the agent least privilege. A Telegram bot that only needs to chat does not need your email password. Connect tools one at a time, and use separate API keys with spend limits for the agent — never your main account keys.
- Keep the server updated. Enable automatic security updates so the next OpenSSL or SSH vulnerability is patched without you remembering. Patch Tuesday in November 2026 alone fixed four RDP flaws and a critical CredSSP bug — see our November 2026 RDP vulnerabilities roundup for how fast unpatched servers get targeted.
- Back up the config off the server. Copy
/home/node/.openclawto your own machine weekly. If the VPS dies, the agent is one restore away instead of a full rebuild.
Troubleshooting: The 5 Problems Everyone Hits
1. “I can’t reach the dashboard.” The gateway binds to loopback by default, so your-vps-ip:18789 will not answer — that is by design. Use the SSH tunnel from Step 3, or check the container is actually running with docker ps.
2. “The container keeps restarting.” Read the logs first. The usual causes are an invalid API key, a bad config file after a hand-edit, or not enough RAM. Remember: building from source needs 6 GB and can be killed on small servers; the pre-built image in this guide avoids that entirely.
3. “My Telegram bot doesn’t reply.” Check three things in order: the bot token is correct, the bot is started (message /start to it first), and BotFather’s privacy settings allow it to see messages. If the bot works in a private chat but not in a group, privacy mode is the culprit — disable it for the bot or add it as an admin.
4. “My API bill exploded.” Switch to a cheaper model tier, shorten the agent’s memory window, and set a provider spend cap. Long coding sessions on the flagship model are the usual cause — move those to a mid-tier model and keep the flagship for questions that need it.
5. “An update broke my config.” Restore the backup you made in Step 5, pin the previous image tag, and restart. Then update again on a quiet day, reading the release notes first. The doctor --fix command fixes most migration issues automatically — run it before editing config files by hand.
FAQ
Can I host OpenClaw on a free VPS?
Technically yes — Oracle’s Always Free Ampere instance (2 OCPUs, 12 GB RAM) can run it. Practically, be careful: idle free-tier instances can be reclaimed, and an agent you rely on should not live on a server that can vanish. Use free tiers to learn OpenClaw, then move to a $5–$10 VPS for anything you depend on.
Do I need a GPU to host OpenClaw on a VPS?
No. When you use an API model (Claude, OpenAI, Gemini), the heavy computing happens on the provider’s servers. Your VPS only runs the lightweight gateway. You need a GPU only if you run a local model on the server itself — see our guide to running AI models on a VPS for that setup.
OpenClaw vs managed OpenClaw hosting — which is cheaper?
Over 36 months, OpenClaw Launch Lite is cheapest ($213) and Hostinger’s managed OpenClaw ($263.64) beats its own KVM 1 VPS ($299.64) on price alone — but the VPS is a full server you can use for other things, while managed plans give you only the agent. Hetzner’s flat €5.49/month VPS (€197.64 over 36 months) is the cheapest honest price with no renewal jump.
How much does running OpenClaw cost per month in total?
Hosting is $5–$12 a month for a typical setup, plus your model API usage — often $5–$20 a month for a personal agent on a mid-tier model, more on a flagship model. Set a spend cap on your API key so the total never surprises you.
Is it safe to connect OpenClaw to WhatsApp or Telegram?
Yes, with the usual precautions: use a dedicated bot token, never expose the dashboard publicly, and give the agent only the tools it needs. A personal agent on a private chat is low-risk; one with access to your email, files, and payments deserves the full security checklist above.
Can OpenClaw run alongside n8n on the same VPS?
Yes — a 2 vCPU / 4 GB VPS comfortably runs both, and they complement each other: OpenClaw for conversation and decisions, n8n for scheduled workflows. See our n8n automation server guide for the setup.
The Bottom Line
Hosting OpenClaw on a VPS is a 30-minute job that buys you an AI agent that never sleeps, on a server you actually own. A 2 vCPU / 4 GB VPS is the right size for most people, Docker is the right install method, and the dashboard stays behind an SSH tunnel — not on the open internet.
If you never want to touch a terminal, managed OpenClaw hosting is a fair deal: OpenClaw Launch Lite is the cheapest way in, and Hostinger’s managed plan undercuts its own VPS over three years. But read the renewal price before the headline price, because the 140% jump in month 25 is where the cheap plans stop being cheap.
Either way, the real cost is not the hosting — it is the model API. Start on a mid-tier model with a spend cap, connect one channel, add one MCP server at a time, and let the agent earn its keep before you give it more power. Your agent will be answering Telegram messages at 3 a.m. by tonight. For the official reference, see OpenClaw’s documentation.


