Deploy OpenClaw to the Cloud: A Private AI Agent on Your Own VPS

Introduction

Every few months a new wave of cloud-hosted AI assistants shows up, and most of them want your data, your money, and an always-on internet connection in exchange for a chat box. OpenClaw flips that bargain. It’s a private AI agent you can deploy on a VPS you own, wire up to your own Telegram or WhatsApp, install “skills” on demand, and schedule like a cron job — without renting a SaaS seat or leaking conversation history to a third party.

This guide walks through deploying OpenClaw to the cloud end to end using Hostinger’s one-click Docker template as a concrete example. The same patterns apply if you’re bringing your own VPS on DigitalOcean, Hetzner, or AWS — the moving pieces are a small Linux box, Docker, a gateway token, and optional messaging-channel credentials. By the end you’ll have a running private assistant, a clean mental model of the architecture, and the security and cost knobs you should know about before you scale it past a hobby workload.

Table of contents

What OpenClaw actually is

OpenClaw is a self-hostable private AI agent. Think of it as a personal assistant that lives on a server you control, talks to whichever LLM you point it at, and exposes itself through a web UI plus optional channels like Telegram and WhatsApp. It isn’t a model — it’s the runtime, the chat surface, the tool layer, and the scheduler stitched together.

The mental model worth keeping: OpenClaw is the body, the LLM API you connect is the brain, and skills are the hands. The cloud VPS is where the body lives.

Why deploy it to the cloud instead of locally

Running OpenClaw on your laptop works for a weekend of tinkering. The moment you want it to actually behave like an assistant — checking your email, pinging you on Telegram at 9am, running an overnight research job — you need it on a machine that doesn’t sleep, doesn’t move networks, and isn’t fighting your browser for RAM.

The trade-off is clear: cloud hosting costs a few hundred rupees or a handful of dollars a month, but it buys you reliability, real channels, and the ability to forget the box exists most days.

Prerequisites and sizing your VPS

OpenClaw itself is fairly modest, but skills, vector stores, and the LLM SDKs that run alongside it add up. The realistic floor for a usable single-user deployment is roughly:

You’ll also need:

• A VPS account at a provider that exposes Docker template provisioning or root SSH
• API credentials for whichever LLM you want the agent to use
• Optional — a Telegram bot token (via @BotFather) and/or WhatsApp Business credentials
• A domain name if you want a clean URL instead of http://<ip>:port

Step-by-step cloud deployment on Hostinger

Hostinger ships a one-click OpenClaw template under its KVM VPS plans, which is the fastest way to go from zero to a running instance. The same flow generalises to any provider — the only real difference is whether you click a template or run a Docker command yourself.

If you’re bringing your own VPS, the equivalent of step 6–8 is a single Docker run command from the OpenClaw docs, plus a reverse proxy (Caddy or Nginx) in front for HTTPS. Most providers’ one-click templates are a thin wrapper over that.

First-run configuration — gateway token, login, channels

The first screen after deployment asks for your OpenClaw gateway token. This is the secret printed during setup — paste it in and click Login. You should land on a chat interface that looks a lot like a generic AI chat app, with extra panels for agents, skills, sessions, and channels.

Quick sanity check — ask the assistant “what can you do?” The default reply lists the broad capability areas: answering questions, research, summarisation, brainstorming, planning, working with files, and managing reminders. If you see that response, the LLM connection is healthy and you can move on to wiring up real capabilities.

Installing skills and wiring up real tools

A bare OpenClaw can chat. A configured one can read your calendar, file GitHub issues, search arXiv, push tasks to Todoist, and write to Notion. Those capabilities ship as skills — modular tool integrations you install from a registry called Claw Hub.

Conversational install looks like this — ask, confirm, paste credentials, done:

You: please install the arxiv skill

Agent: Installing arxiv... done. No API key required.
       Try: "search arxiv for ‘attention is all you need’ and summarise"

You: search arxiv for "attention is all you need" and summarise it

Agent: [calls arxiv.search tool]
       Vaswani et al., 2017. Transformer architecture replacing recurrence
       with self-attention…

Scheduling tasks with the cron-job feature

The piece that makes a private agent feel like an assistant rather than a chatbot is scheduling. OpenClaw lets you create cron jobs in plain English — the agent translates your request into a stored schedule and runs the underlying prompt at the configured interval.

You: every morning at 9am IST, search arxiv for recent agentic-AI
     research and summarise the top papers

Agent: [calls cron.create tool]
       Created daily job "agentic-ai-digest" running 09:00 IST.
       I’ll post results into this channel.

You: run it now to verify

Agent: [executes immediately]
       Digest: NextMem (latent factual memory for LLM agents),
       MirrorThinker (heavy-duty research with self-verification),
       Recursive Language Models under uncertainty…

A few practical notes on the scheduler:

• Job timeouts default to roughly 60 seconds — bump them for research jobs that need to call multiple tools
• Output is posted into the channel the job was created from — pick a dedicated channel for noisy recurring jobs
• You can list, edit, and delete jobs in chat or from the UI’s scheduler panel
• Long-running jobs use isolated workers — expect a small delay between trigger and result

Connecting Telegram and WhatsApp safely

Channels turn OpenClaw from a browser app into something that actually shows up where you live — your phone’s messaging apps. The Telegram path is the cleanest, so it’s worth walking through.

WhatsApp setup follows the same pattern — create a WhatsApp Business app, grab the credentials, paste them into Channels. The cost and approval friction is higher than Telegram, so most people start with Telegram and add WhatsApp once the workflow is proven.

Where else you can host it — provider trade-offs

Hostinger’s one-click flow is the lowest-friction path, but it isn’t the only one. Different providers shine in different scenarios.

Realistic cost breakdown

The sticker price of the VPS is rarely the whole story. Here’s a rough monthly budget for a personal OpenClaw deployment that actually does useful work.

Compared to paying for a multi-seat SaaS AI assistant, a self-hosted OpenClaw is usually cheaper above two or three users and competitive even for a single power user — especially once you factor in the lack of vendor lock-in.

Security and privacy posture

The promise of self-hosting is privacy — but only if you configure it that way. A misconfigured OpenClaw on a public IP is an open door to whoever finds the URL first.

The reassuring bit — because OpenClaw lives entirely on your VPS, your conversations, files, and integrations don’t round-trip through a vendor’s SaaS plane. The only third party that sees content is whichever LLM provider you point it at, and you can swap that in the settings any time.

Troubleshooting common deployment issues

Best practices for running it in production

Common mistakes to avoid

• Undersizing the VPS. The 4 GB tier looks fine on paper and falls over once you add Docker, two skills, and a vector index. Start at 8 GB.
• Skipping the rotate-token step. Default tokens often end up in screenshots and support threads. Rotate before you do anything else.
• Wiring every channel at once. Get one channel solid — usually Telegram — before you add WhatsApp and Slack on top.
• Treating the agent as autonomous. A private AI assistant is still an assistant. Review what it’s about to do for anything that touches money, identity, or production systems.
• Ignoring observability. If you don’t watch logs and job histories, you won’t notice a silently failing cron until it’s a week late.

Conclusion

Deploying OpenClaw to the cloud is no longer the weekend project it used to be. A one-click template plus a sensible VPS plan gets you from credit card to working private assistant in under thirty minutes, and the architecture — container on a VPS, your LLM keys, your channels, your skills — is simple enough to reason about even when things go sideways.

The interesting work starts after the deployment. Picking the right skills, scheduling jobs that actually save you time, locking down channels, and treating the box like infrastructure rather than a toy is what separates a private AI agent that gathers dust from one that quietly becomes part of your daily workflow. Start small, keep the surface area tight, and let the assistant earn each new permission you grant it.

Explore More on DevShelf

• →
  OpenClaw — Tool Profile

  Full overview of OpenClaw's agent runtime, skill marketplace, channel integrations, and pricing.

• →
  Run OpenClaw Locally with Ollama

  The local alternative to cloud deployment — zero API costs, full privacy, using Ollama for on-device inference.