Step-by-Step OpenClaw Setup for People Who

TL;DR: OpenClaw is now set up through openclaw onboard --install-daemon, not the older clawdbot or curl | bash flows. As of April 2026, the official repo recommends Node 24 or Node 22.16+, supports macOS/Linux/Windows via WSL2, and treats BlueBubbles as the recommended iMessage path. Heads up before you start: on April 4, 2026, Anthropic blocked Claude Pro/Max subscriptions from authenticating third-party tools like OpenClaw via OAuth. There’s a clean workaround — piping OpenClaw through the local claude CLI that’s already authenticated on your machine — and the setup docs cover it. See the note below before you pick a model provider.

I run an AI assistant on a Mac Mini in my office 24/7.

I treat it like hiring a real assistant. Separate accounts. Separate workspace. Separate memory. Always on.

The important thing is this: a lot of OpenClaw setup advice on the internet is already stale.

If a guide tells you to use clawdbot, install via curl https://openclaw.ai/install.sh | bash, or configure direct iMessage as the default modern path, you’re reading an older snapshot. The current setup is cleaner than that.

This is the version I’d hand a non-technical founder today.

What OpenClaw actually is

OpenClaw is a local-first AI assistant runtime. You run the gateway on your own machine or VPS, connect the channels you already use, and give one or more agents access to tools, files, schedules, and memory.

That means AI stops being a browser tab and starts acting more like infrastructure.

As of April 2026, the official repo has a much broader product surface than early tutorials implied: channels, cron, skills, Canvas, companion apps, talk mode, and a full onboarding flow.

The one thing to know before you start

Use the official onboarding flow.

Do not piece this together from five different old blog posts unless you enjoy debugging renamed commands.

Before you start

You need:

• a machine that can stay on
• Node 24 recommended, or Node 22.16+
• one model provider you already trust and pay for
• one first channel to connect

You do not need to understand every config file up front.

Start with one machine, one agent, one channel, one real task.

Step 1: Install the CLI

The official README currently recommends either npm or pnpm:

npm install -g openclaw@latest

or

pnpm add -g openclaw@latest

If you do not have Node yet, install Node 24 first from nodejs.org or through your package manager.

Step 2: Run onboarding

This is the setup command that matters:

openclaw onboard --install-daemon

That flow is designed to guide you through:

• gateway setup
• workspace setup
• model auth
• channel setup
• daemon installation

It is the recommended CLI path on macOS, Linux, and Windows via WSL2.

A heads-up about Anthropic / Claude Max

Before Step 3, read this.

On April 4, 2026 — the Saturday before Easter — Anthropic sent an email late on Friday announcing that, effective immediately, Claude Pro and Claude Max subscriptions could no longer be used to authenticate third-party tools like OpenClaw. Their stated reason was compute strain and prompt-cache hit rates on agentic harnesses. If you set this up expecting to just plug your $200 Max sub in via OAuth, you will hit this wall.

There is a workaround. Within hours of the announcement, the community (and then the OpenClaw docs themselves) converged on piping OpenClaw through the local claude CLI binary that’s already authenticated on your machine. From Anthropic’s side it looks like native CLI usage instead of a third-party harness, and the OAuth block doesn’t apply. You keep using your Max subscription, you don’t have to move to metered API billing, and your agents don’t feel lobotomized by a weaker default model.

So: if you were planning to use Claude Max, don’t panic. Just know that the setup path now goes through the local CLI adapter, not raw OAuth. Check the current OpenClaw docs for the claude adapter instructions — they’re what I follow.

If you already pay for an API key and don’t care about subscription pricing, this doesn’t affect you.

Step 3: Authenticate your model provider

The current official guidance is simple: prefer a current flagship model from the provider you already use and trust.

During onboarding, connect the provider you actually plan to pay for. Don’t overbuild a multi-model routing setup on day one.

My advice:

• if you already live in Anthropic and pay for Max, route through the local claude CLI adapter (see the note above)
• if you already pay OpenAI, start there
• do not optimize this before you even have one working channel

Step 4: Pick your first channel

This is where people make the setup harder than it needs to be.

Do not wire up five channels in the first session. Pick one.

If your real goal is “I want this assistant on my phone,” BlueBubbles is the current path worth understanding, not older direct-iMessage-only guides.

Step 5: Verify the install

Run:

openclaw doctor

This is the command I wish more people used. It catches risky or incomplete setup states much earlier than waiting for a broken channel to surprise you later.

Then send the assistant one real message on the first channel you connected. Don’t move on until that works.

Step 6: Make it actually useful

A bare assistant is not the point. Context is.

My baseline setup still looks like this:

• separate machine identity
• separate workspace
• clear written instructions
• persistent memory files
• one real recurring job

If you want the assistant to remember decisions, preferences, and business context, pair OpenClaw with a simple file-based memory system or a dedicated brain layer like g-brain.

Step 7: Keep it healthy

The official update path is now:

openclaw update --channel stable
openclaw doctor

If you want more change velocity, there are also beta and dev channels. If you are non-technical, stay on stable until you have a reason not to.

What I would do if I were starting from zero today

1. Install Node 24.
2. Install openclaw@latest.
3. Run openclaw onboard --install-daemon.
4. Connect one provider.
5. Connect one channel.
6. Run openclaw doctor.
7. Give the assistant one real recurring job.

That’s the setup loop.

Not “install everything.” Not “connect every service.” Just get one useful assistant live.

Related guides

• Turn a Mac Mini into Your First AI Employee
• OpenClaw Review: Is It Worth It for Founders in 2026?
• The Ops Layer That Keeps Your OpenClaw Agents Alive
• Why Your OpenClaw Agent Doesn’t Remember You
• What Is g-brain?