OpenClaw Setup for Beginners: Safe Steps and Real Running Costs

OpenClaw is a free, open-source AI agent that runs on a computer you control and answers you through chat apps such as Telegram. The official quick start promises a first chat in about five minutes, and plan on an hour once you add hardening and a chat app. It costs nothing to install. The real costs are the machine it runs on and the AI model it calls, which for light use land between about $1.50 and $72 a month depending on the model and the server you pick. This guide follows the official documentation, read on 21 October 2026 against release v2026.10.5, and it adds the safety stops the quick-start guides skip.

This is general information, not a security audit or professional advice. OpenClaw’s own docs say it is not built to keep apart users who do not trust each other, so read the pillar guide on whether OpenClaw is safe for business before you connect anything that matters.

Key Takeaways
OpenClaw is MIT-licensed with no paid tier, hosted service, or token. You pay for hardware and for the AI model, not for the software.
It needs Node 24.16+ or 26.1+. The official installer adds Node for you when it is missing, and Windows users get a native app (Windows Hub) as the recommended desktop path.
Do not connect a chat app until you have worked through the security checklist. Setup is only the first third of a safe install.
Model usage is the cost that moves. With the illustrative assumptions below, 600 calls a month cost about $1.50 on a small model and about $60 on a large one.
A $12 a month server is enough hardware. A model bill with no limits is the thing that can surprise you.

Is This Setup Right for You?

Check these four points before you install anything. If one is a firm no, stop here and look at a managed workflow tool instead.

You can use a terminal. The install, the status checks, and the pairing steps are commands you type or paste. Windows Hub covers installation, tray status, and chat, but the pairing steps in this guide still use commands.
You have a machine you can dedicate. The docs call a dedicated host “not required, but recommended for reliability and isolation.” A laptop that sleeps and reboots will interrupt the agent.
You will keep it updated. The project ships very fast, so an install you never touch drifts out of date. The checklist covers an update routine.
You are not planning to connect customer, payroll, or banking data. The pillar guide explains why that data should stay away from an experimental agent.

If you would rather run fixed, auditable workflows than an agent that decides its own steps, see Best Self-Hosted Workflow Automation Tools and n8n vs Make vs Zapier.

Step 1: Choose Where It Runs

Where OpenClaw runs decides your monthly cost and how much damage a mistake can do. The docs say the Gateway can run on a laptop, a home server, a Raspberry Pi-class box, or a small VPS (a rented cloud server). You do not need a Mac mini; a Mac is only required for macOS-only tools such as iMessage.

Option	Hosting cost	Good for	Watch out for
Spare PC or Mac at home	$0 plus electricity	Learning with no hosting bill	Sleep, reboots, and a machine that also holds your personal files
Your daily Windows PC (Windows Hub)	$0	The easiest first install	Shares the computer with your own accounts and files
Small VPS, 1 vCPU and 2 GiB RAM	$12 a month on DigitalOcean’s pricing page	Always on, separate from your own files	You must secure SSH access and run commands as a non-root user
Smallest VPS, 1 vCPU and 1 GiB RAM	$6 a month on the same page	Testing only	The docs call 1 GB the “absolute minimum”

The docs recommend “1-2 vCPU, 2 GB+ RAM for headroom” and Ubuntu LTS for a server. Prices above were read from DigitalOcean’s pricing page on 21 October 2026; plans and prices change, and tax is extra.

For a first install, a spare machine is the safest choice because a mistake only touches that machine. Step 1 of the security checklist explains the isolation reasoning in detail. If you rent a VPS, DigitalOcean’s own OpenClaw guide in the docs says to use a clean Ubuntu 24.04 LTS image and to avoid third-party Marketplace 1-click images unless you have reviewed their startup scripts and firewall defaults. That server route involves SSH keys and a root-then-non-root user setup, so treat it as the harder path for a first-timer.

Step 2: Check the Requirements

OpenClaw needs Node 24.16+ or 26.1+, and the docs recommend Node 26. Node 22, 23, and 25 are unsupported. You can check what you have with:

node --version

You do not have to install Node yourself in most cases. The official installer detects your system and installs Node if it is missing. You also need access to an AI model, either an existing Claude Code or Codex CLI login or a provider API key, which onboarding can detect and reuse.

Step 3: Install OpenClaw

Pick the path for your system. Copy commands only from the official docs at docs.openclaw.ai, because these installers run a script from the internet on your machine.

Windows. The docs recommend the native Windows Hub companion app. It is a signed installer for Windows 10 20H2 or newer and Windows 11, it installs without administrator privileges, and its Set up locally option creates a separate WSL (Windows Subsystem for Linux) environment for the Gateway without touching an existing Ubuntu install. Download it from the Windows Hub releases page that the install docs link to. If you prefer a terminal, the PowerShell installer is also supported:

iwr -useb https://openclaw.ai/install.ps1 | iex

macOS, Linux, or WSL2:

curl -fsSL https://openclaw.ai/install.sh | bash

Windows Hub can also register your PC as a “node” so the agent can use the screen, camera, and a system.run command. Leave node mode off for now. The security checklist treats system.run as remote code execution on that device, and you should add it only on purpose.

Step 4: Run Onboarding Carefully

The installer starts a guided onboarding wizard. You can also run it yourself with openclaw onboard. Choose Quick start to reuse AI access OpenClaw detects and open the web dashboard, or Custom setup for every option.

Three choices in this step matter for safety:

Access mode. The guided flow asks whether setup may look around for AI apps, keys, and local runtimes automatically (“full access,” the recommended option) or ask first. Quick start defaults to full access. On a machine that holds anything personal, choosing ask first is the more cautious reading of that choice; this is a judgment call, not a docs recommendation.
Which AI connection to use. Onboarding verifies only the connection you pick, with a real test reply, and it does not switch providers for you if the test fails. My suggestion: use a dedicated API key created for this agent instead of your personal logins, so you can revoke it alone.
Local models. The docs say a local model is “usually no” for casual chats, because small or heavily compressed models raise prompt-injection risk. Start with a hosted model.

When onboarding finishes, Quick start leaves the Gateway running in your terminal window until you press Ctrl+C. Your settings stay saved.

Step 5: Keep It Running and Verify

Press Ctrl+C, then install the background service so OpenClaw survives closing the window:

openclaw gateway install
openclaw gateway status
openclaw dashboard

The status command should show the Gateway listening on port 18789, and openclaw dashboard opens the browser control panel, where you can send a first test message. The service is a LaunchAgent on macOS, a systemd user unit on Linux and WSL2, and a Scheduled Task on native Windows. You can run openclaw --version and openclaw doctor at any time to confirm the install and look for config problems.

Step 6: Stop and Harden Before Connecting Anything

Here is the step quick-start guides skip. At this point you have an agent that works, but it runs with default settings. Per the docs, sandboxing is off by default and the default single-operator setup allows host commands without approval prompts.

Open the OpenClaw security checklist and finish at least these before you add a chat app:

Apply the official hardened baseline config.
Keep the Gateway on loopback (private to the machine).
Run openclaw security audit and fix what it flags.

That checklist walks through all eleven steps, so this guide does not repeat them.

Step 7: Connect Telegram

Telegram is the fastest channel to set up because it needs only a bot token. The docs’ sequence is:

In Telegram, chat with the account named exactly @BotFather, send /newbot, and save the token it gives you.
Add the token to OpenClaw:

openclaw channels add --channel telegram --token <bot-token>
openclaw channels status --probe

Send any message to your new bot. This creates a pairing request. Approve it from the machine running OpenClaw:

openclaw pairing list telegram
openclaw pairing approve telegram <CODE>

Pairing is the default direct-message policy for Telegram: strangers who message the bot get a short code, and their message is not processed until you approve. Codes expire after one hour. Approve only your own account and anyone you have deliberately chosen. Do not add the bot to a group chat until you have read the access-control section of the checklist.

Step 8: Watch the Spending

Every reply and tool call runs on your model provider, and that is the main source of cost. OpenClaw gives you three ways to see usage:

/status in chat shows the session model, context size, and an estimated cost for the last reply.
/usage full adds a usage footer, with estimated cost, to every reply. /usage cost prints a local summary.
The dashboard’s Usage view shows token and estimated-cost totals by provider, model, and day.

These totals come from local session history. The docs state they “describe the available local session history, not a provider invoice,” so also check your provider’s own billing page. Set any spending limit or billing alert your provider offers before you leave the agent running unattended.

What OpenClaw Costs to Run

The software is free. Two bills remain: hosting and model usage.

Hosting is $0 on a machine you already own (plus electricity) or about $12 a month for the 2 GiB VPS in the table above.

Model usage depends on the model and on how much the agent does. OpenClaw sends its instructions, tool list, and workspace files with every request, and the docs cap that injected file text at 60,000 characters, so each call carries a large fixed input. Tool use can also trigger several model calls for a single message.

The table below is an illustration, not a quote. It uses Anthropic’s published API prices, read on 21 October 2026 from its pricing page, and these assumptions of mine: 20,000 input tokens and 1,000 output tokens per model call, with no prompt caching.

Model (price per million tokens, input / output)	Cost per call	600 calls a month (about 20 a day)	3,000 calls (about 100 a day)	12,000 calls (about 400 a day)
Claude Haiku 5.5 ($0.10 / $0.50)	$0.0025	$1.50	$7.50	$30
Claude Sonnet 5.5 ($2 / $10)	$0.05	$30	$150	$600
Claude Opus 5.5 ($4 / $20)	$0.10	$60	$300	$1,200

Add $12 for the VPS and a light user on the small model pays about $13.50 a month, while a busy user on the large model pays well over $1,000. Two things make real bills lower or higher than this table. Prompt caching lowers input cost, since Anthropic bills cache reads at 0.05 times the base input price on Sonnet 5.5 and Opus 5.5, so the table errs high on input. Long tool loops, large attachments, and heartbeat or scheduled jobs push the number of calls up, so it can also err low. Haiku 5.5 also charges more for prompts over 100,000 tokens.

Other features can bill separately, and the docs list them: web search (Brave’s plan is $5 per 1,000 requests with a $5 monthly free credit, per the docs), image and video generation, speech, and remote memory embeddings. Skip these until the basics work.

You can sometimes avoid per-token billing by signing in with a ChatGPT or Codex subscription. The docs say this works, but the quota differs from the ChatGPT website and the limits vary by plan. Check the provider’s current terms before relying on any subscription route.

Mistakes to Avoid

Connecting a chat app before hardening. A working agent is not a safe one. Finish Step 6 first.
Running it on your main computer with your personal accounts signed in. For a shared company agent, the docs say not to sign its runtime into personal Apple or Google accounts or personal browser and password-manager profiles.
Choosing the biggest model by default. Start small and move up only for tasks that need it.
Exposing the dashboard to the internet. Keep the Gateway on loopback and use an SSH tunnel or Tailscale if you need remote access.
Leaving the first API key unlimited. Use a dedicated key and a billing alert.
Skipping updates. Run the update routine from the checklist, then re-run the audit.

If Setup Fails

The docs give one first move for a broken install:

openclaw triage

It runs read-only health checks and writes a sanitized description of the problem. It offers to pass that description to a coding agent it finds on your machine, and nothing leaves your computer until you choose one. You can also run openclaw doctor yourself. The most common install fault is that openclaw is “not found,” which the docs say is almost always a PATH issue (the global npm folder is missing from your shell’s search path).

Frequently Asked Questions

Can I set up OpenClaw without coding?

You do not need to write code, but you do need to type commands. On Windows, the Windows Hub app installs like a normal program, but the Telegram pairing and the security checklist still use commands. If that is a dealbreaker, a managed workflow tool is a better fit.

Is OpenClaw free?

The software is MIT-licensed, and the README says it has no paid tier, hosted service, or token. You still pay for the computer or server and for the AI model you connect.

How much does OpenClaw cost per month?

Hosting runs from $0 on a spare machine to about $12 for a 2 GiB VPS on DigitalOcean’s current pricing. Model usage is the variable part: under the assumptions above, 600 calls a month cost about $1.50 on Claude Haiku 5.5 and about $60 on Claude Opus 5.5.

Do I need a Mac mini?

No. The docs say OpenClaw runs on macOS or Linux, and on Windows through Windows Hub or WSL2. A small VPS, a home server, or a Raspberry Pi-class box also works. You only need a Mac for macOS-only tools such as iMessage.

What is the easiest chat app to connect?

Telegram, according to the docs, because it needs only a bot token. Keep the default pairing policy so unknown senders cannot talk to your agent.

Is the quick-start setup safe for a business?

It gets you a working agent, but it does not harden it. Sandboxing is off by default and the default trusted-operator mode allows host commands without approval prompts. Follow the checklist before connecting any account, and read the pillar guide to decide whether OpenClaw belongs in your business at all.

Continue Learning

Start here:

Safer automation:

Coming next in this series:

OpenClaw vs n8n for Business Automation: Which Fits Which Job

Install in the order the docs give, harden before you connect, and put a ceiling on spending before you walk away.