OpenClaw Setup in 2026: Requirements, Steps, Real Costs

Installing OpenClaw genuinely takes minutes: the official installer plus the onboarding wizard get you to a first chat fast, and the docs say as much↗. The setup that matters takes longer, because installed is not the same as useful: you still need a server that stays on, a channel your assistant lives in, a model API key, and a security pass on the Gateway.

This guide walks the honest version of all of it, with real numbers. And if you want the destination without the journey, Tessera Founder runs a managed Hermes instance (the open-source agent by Nous Research, OpenClaw-style) connected to your Telegram, with the AI included at a flat rate, for €55 a month.

What you need before you start

Four things: a machine, a supported Node version, a model API key, and a free hour. The official requirements↗ ask for a current Node runtime, Node 22 or newer, with Node 24 recommended as the default; older lines like Node 18 and 20 fail with a version error.

For an always-on assistant you want a small server rather than your laptop. Community sizing guides converge on 2 vCPU and 4 GB of RAM as the comfortable floor, 8 GB if you plan browser automation, which lands in the $4-24 per month VPS range depending on provider.

The model API key is the quiet decision that sets your monthly bill. Most tutorials default you to a metered frontier key from Anthropic or OpenAI; keep that choice open until you have read the cost section below.

The install, step by step

The official flow is short, and it works as advertised. In order: run the platform installer (bash on macOS and Linux, PowerShell on Windows), run openclaw onboard --install-daemon to configure your model provider and API key, confirm the Gateway is listening on port 18789 with openclaw gateway status, open the control UI with openclaw dashboard, and send your first message from the browser.

The onboarding wizard covers the pieces people usually forget: it registers the daemon so the Gateway survives reboots, stores your provider credentials, and leaves you with a working control UI. If you skip an optional step, you can rerun onboarding later without reinstalling.

Docker is optional and officially supported with pre-built images. The docs position it for isolated or throwaway environments and warn that containers under 2 GB of RAM risk out-of-memory kills during setup (OpenClaw docs↗).

So why does a popular community guide describe itself as born from 15 days of struggle↗? Because everything after the install is where the real work lives: channel pairing, security, memory, and skill tuning. Budget your time for that, not for the installer.

Pairing a channel: start with Telegram

Telegram is the fastest channel to connect, and the official docs say exactly that: it needs just a bot token. You create the bot with BotFather, paste the token, and the Gateway reaches Telegram by polling outward, so you do not open inbound ports or configure a reverse proxy. The Bot API↗ itself is free, with no per-message charge at any volume.

WhatsApp is the channel most founders ask about, and it needs an honest caveat in 2026. Meta banned general-purpose AI assistants from the WhatsApp Business API on January 15, 2026, and charges per service message from October 1, 2026; unofficial QR bridges end in a banned number. We covered what still works on WhatsApp separately: defined-function business bots through the official Meta API, general assistants on Telegram.

OpenClaw also speaks Discord, Signal, Slack, and Microsoft Teams, among others. Each channel pairs independently, so a misconfigured one does not take down the ones already working, and the assistant is never locked to a single app.

Lock down the Gateway before anything else

Is OpenClaw safe to install? Yes, with one non-negotiable rule that security writeups keep repeating: never expose the Gateway to the open internet↗. It listens on port 18789 for you, not for the world; keep it bound locally or behind a private network.

Remember what this box holds: your model API key, your bot tokens, and an agent with permission to act on your accounts. Treat it like a production credential store. Patch the OS, use SSH keys instead of passwords, and be deliberate about which skills you enable, because an agent that reads untrusted content can be steered by prompt injection.

None of this is exotic sysadmin work, but it is recurring work. It is also the part the five-minute framing leaves out.

Keeping it running

An OpenClaw box is a small piece of infrastructure, and it behaves like one. The npm package ships new releases frequently, so plan for regular updates, an occasional breaking change in a beta line, and a daemon restart after each upgrade.

You also inherit the boring parts: OS patches, disk space for logs and memory files, and backups of your configuration so a dead VPS does not take the assistant’s memory with it. If the Gateway dies while you are away, the assistant is simply offline until you notice, so monitoring is one more small job you take on. None of it is hard; all of it is yours now.

A fair budget is an hour or two per month once things are stable, plus the occasional evening when an update changes behavior. A recurring calendar reminder to update and reboot is enough process for most solo setups. If that sounds like a hobby, self-hosting will suit you; if it sounds like a tax, read the last section.

The line every tutorial skips: your token bill

The server is the cheap part of self-hosting; the tokens are the bill that surprises people. Agents like OpenClaw and Hermes carry roughly 15,000-20,000 tokens of tool definitions in every request, so even short exchanges are token-heavy.

Run the numbers on a real assistant that handles 40 requests a day: at about 17,500 overhead tokens each, that is roughly 700,000 tokens a day, or about 21 million a month, before the model does any actual work. On a metered frontier API that shows up as variance: community estimates for self-host running costs span $6-74 per month with tokens dominating, and there are public reports of a single month costing $500. We break the full math down in how much OpenClaw really costs.

The structural fix is a flat-rate, OpenAI-compatible endpoint: OpenClaw-style agents talk to any OpenAI-compatible base URL, so the model key does not have to be metered. That is the half of the problem most hosted-OpenClaw services leave on your plate; we compare that market in OpenClaw alternatives.

When you should not self-host

Self-hosting is the right call if you enjoy running servers, want full control, and treat the maintenance as a hobby you like. The install is genuinely accessible, the official docs are good, and the community is large and active. The math also changes if you already operate servers for other reasons.

It is the wrong call if you are a solo founder whose scarcest input is time. Between the VPS, the security pass, the updates, and a metered token bill, you are operating a small piece of infrastructure to save a subscription fee.

That trade is why Tessera Founder exists: we operate your private Hermes instance on European servers, connect it to your Telegram, and include the AI at a flat rate with no token counter. The only limit is pace, 20 queries per minute, and it costs €55 a month with a first-month money-back guarantee.

FAQ

How long does OpenClaw setup take?

The install and onboarding wizard take minutes, as the official docs claim. A production setup with an always-on server, channel pairing, and security hardening is more honestly measured in hours; community guides describe days of iteration to make it reliable.

Is it safe to install OpenClaw?

Yes, if you keep the Gateway off the open internet, protect the server like a credential store, and are deliberate about skills that read untrusted content. Most reported incidents trace back to exposed Gateways or leaked keys.

Can I run OpenClaw without a VPS?

Yes, it runs on a Mac or Linux machine at home, and that is fine for trying it out. An assistant that answers while your laptop is closed needs an always-on machine: a small VPS or a managed service.

Do I need Docker to run OpenClaw?

No. The normal flow installs OpenClaw as a Node package with a daemon. Docker is optional and officially supported with pre-built images for isolated or throwaway environments; small containers risk out-of-memory kills.

Which model should I connect it to?

Any provider with an OpenAI-compatible endpoint works. A metered frontier key is the default in most tutorials; an open-weight model like Qwen3.6-35B-A3B on a flat-rate endpoint keeps the bill fixed, which matters more for agents than for chat.

What does it cost to run after setup?

The server is $4-24 per month. The variable is the model API: metered keys produce $6-74 typical months with spikes far beyond, while a flat plan like Tessera Founder fixes AI and hosting together at €55.