How to Set Up OpenClaw: From Install to Your First Skill

Have you ever wished your AI assistant could live inside the chat apps you already use, instead of you opening yet another website? That is exactly what OpenClaw does. It is a free, open-source assistant that runs on your own computer or server, replies to you in your everyday messaging apps, and can actually do things for you: browse the web, manage files, and run scheduled tasks while you sleep.

In this guide, I'll walk you through the whole journey: complete installation, running the setup wizard, connecting your first chat app, building your own skills, and fixing the problems that almost everyone hits. I'll also explain why Windows users should go through WSL instead of fighting a native install and will leave you with a cheat sheet of the commands you'll use every day.

Every command is given for both Windows and Mac/Linux. No prior technical knowledge needed. If you can copy, paste, and read an error message, you can do this!

• What OpenClaw is and how is it different for a normal chatbot?
• The full setup, in five steps. 
• • Download and install Openclaw: Download and install the software on your Windows or Mac/Linux system
  • Run the setup wizard: Configure the required settings, and complete the initial onboarding steps.
  • Start the gateway: Enable the gateway service so the platform can communicate with connected applications and external services.
  • Connect your first chat app: Link a chat application to openclaw
  • Approve your first message: Send a test message through the connected app and approve the chat ID to complete the setup.
• Why should Windows users go through WSL? Installing natively on Windows breaks several useful skills. I explain the exact reason, and why one extra command at the start saves you hours of workarounds.
• How to build your own skills from scratch? A skill is simpler than you think, and you'll learn the one tiny mistake that silently breaks them for almost everyone.
• Fixes for the four most common problems, 
• • A skill does not appear: the most common blocker.
  • The agent keeps timing out: usually a sign something in the model setup is slowing it down or not running at all.
  • Runs get cut off mid-task: often caused by a context window that is too low for the job.
  • A scheduled job fails after running: usually an external service, limit, or config issue behind the scenes
• A command cheat sheet for daily use. The handful of commands that cover 90% of what you'll ever do, explained in one table near the end.

OpenClaw is a personal AI assistant that you host yourself. It connects to the messaging channels you already use and replies there, instead of forcing you into a separate app. Think of it as a smart helper that never forgets what you told it.

Four things make it different from a normal chatbot:

• All at one place: Rather than opening another app, you simply message OpenClaw through the chat platforms you already use. 
• It remembers: Past conversations stay available, unlike chatbots that reset after every session.
• It takes action:  It can run commands, manage files, and execute jobs on a schedule. A scheduled job is called a cron job, which is just a fancy name for a task that runs automatically at a set time.
• Your data stays with you: It runs on your machine, so nothing is stored on someone else's platform unless you choose so.

The last point cuts both ways. Because OpenClaw can run real commands on a real computer, a badly secured setup is a real risk. We will fix it near the end of this guide.

Not much, honestly. The official docs ask for Node.js version 22.12.0 or newer, which is the runtime OpenClaw is built on. The install script handles most of this for you, but it helps to check upfront.

Check your Node.js version (Mac/Linux terminal, or WSL terminal on Windows):

node –version

If the number shown is below 22.12.0, go ahead and update first before going further. Researchers tied older versions to known, already-fixed vulnerabilities.

Here’s what you’ll need:

• A computer, a small home server, or a cheap cloud server or simply a spare laptop which you can leave running.
• About 30 to 60 minutes of free time for the first setup.
• An API key from an AI provider of your choice. The setup wizard will ask for it.

Fair question, and I have a concrete reason from my own setup. My use-case involved connecting Google accounts to OpenClaw, which is done through the GOG skill. Setting up GOG requires Homebrew, which runs on macOS and Linux but doesn’t run natively on Windows. Workarounds exist, but they are complicated, fragile, and exactly the kind of thing this guide is trying to avoid.

If you try to run openclaw natively on Windows, you will keep hitting small walls. Run inside WSL and everything behaves exactly like it does on Mac and Linux. One install command, zero surprises.

The goal of this guide is to make OpenClaw easy to set up on any system. On Windows, WSL is what makes it possible. A few extra setup steps upfront save countless workarounds later. 

On Mac or Linux

Open your terminal and paste this one line. It downloads and installs everything:

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

On Windows

OpenClaw runs inside WSL2, the Linux layer that you can run inside Windows 10 and 11. Setting it up takes two minutes.

wsl –install -d <linux-version>

Open the new Ubuntu app from your Start menu, create a username and password. Restart your PC when asked and after restart, check the wsl status

From this point on, whenever I say: run this in your terminal, Windows users should run the commands inside that Ubuntu (WSL) window.

To open your Linux environment from any PowerShell window:

wsl

That's it. One word drops you into Ubuntu, right inside the same window. You can also launch it by typing "Ubuntu" in the Start menu, both routes land in the same place

The commands from this point on are the same for mac and windows users.

Now, run the following command to install openclaw:

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

OpenClaw ships with an onboarding wizard that configures your AI model, security token, chat channels, and background service in one guided flow. Run:

openclaw onboard

The wizard asks three things:

1. Pick your AI model. Choose the provider you have an API key for, and paste the key when prompted.
2. Pick your chat apps. Select where you want the assistant to answer you.
3. Accept the security token. The wizard creates one for you. Store it for later.

The Gateway is a part of OpenClaw that stays running and routes messages between your chat apps and the AI. After you’ve installed the service in Step 2, it should already be running. To check the status, use the following command.

openclaw gateway status

Not running? Give it a hard push:

openclaw gateway --port 18791 --verbose

Or simply run:

openclaw gateway

You can now open a browser on the same machine and visit http://127.0.0.1:18791/ to see the control panel.

One important note: The panel should ONLY be reachable from your own machine. We will come back to the why part (for this) in the security section.

Each chat platform gives you a bot token from its developer settings page.

You can either use the following command or you can do it using the Openclaw config command.

openclaw channels add <your-chat-app>

Replace <your-chat-app> with the channel name the wizard showed you. Paste the bot token when prompted, and you are connected.

Here is a nice safety feature. OpenClaw will not talk to strangers. The first time you message using a connected chat app, it generates a pairing code instead of replying. You approve that code from your terminal, which proves the person messaging the bot is actually you.

openclaw pairing approve <your-chat-app> <code>

After approving, open a new chat in the chat app and write something simple like “What's the weather today?” If openclaw replies, congratulations. The base setup is done.

Now, it’s time for the fun part!

Skills are how you teach OpenClaw new tricks. There is a public marketplace of community skills, but you do not have to trust it. In early 2026, a campaign, nicknamed ClawHavoc, planted 341 malicious skills on that marketplace to steal credentials. 

Writing your own skills is safer, and honestly easier than it sounds. A skill is just a folder with one text file inside.

We can use AI to write, store and use skills for us. Once the app is set up, simply prompt it to write a skill for a use case and watch it will do the rest. Choose whichever approach appeals to you.

This is how it will look like once the skill has been created:

Create the folder structure

Each skill lives in its own folder inside ~/.openclaw/skills, and each folder needs a file called SKILL.md. The squiggle symbol ~ just means your home folder.

Mac/Linux (and Windows inside WSL):

mkdir -p ~/.openclaw/skills/my-first-skill
nano ~/.openclaw/skills/my-first-skill/SKILL.md

The nano command opens a simple text editor right in the terminal. Type or paste your skill content, then press Ctrl+O to save, Enter to confirm, and Ctrl+X to exit. If you prefer a normal editor on Windows, your WSL files are also visible in File Explorer under the “Linux” section in the sidebar, so you can edit SKILL.md with any editor you like.

Write the SKILL.md file

Every SKILL.md starts with a small header block that tells OpenClaw the skill's name and what it does. Keep it short:

---
name: my-first-skill
description: "Summarizes my unread messages every morning"
---

Instructions for the skill go here, written in plain language.

The bug that wastes everyone's afternoon

Pay close attention here, because this exact issue has eaten hours of debugging time for users. If your description line contains certain characters, like square brackets, and it is not wrapped in quotes, OpenClaw silently skips the whole skill. No errors. No warning.

description: This is the [task]      <-- skill gets silently skipped
description: "This is the task"      <-- this works

Enable the skill in your config

Next, tell OpenClaw the skill exists. Open the main config file at ~/.openclaw/openclaw.json and add an entry for each skill. If a skill needs a password or API key for some service, you put it in the env section.

You can also ask OpenClaw to handle it, but do not share secrets or credentials in chat. Most skills do not require API keys. 

Confirm the skills registered

openclaw skills list

Every skill you created should appear in that list. If one is missing, check the quotes on its description line first. This small check fixes it nine times out of ten.

Skill Creation Prompt Example:

Create a skill that checks my email once per day and forwards only relevant messages that mention “Snowflake” to *********@gmail.com.
Requirements:
- Run on a daily cron schedule.
- Search my inbox for emails where the subject or body contains “Snowflake” (case-insensitive).
- Do not forward promotional, marketing, newsletter, spam, or advertising emails.
- Forward the email only if it is a normal human-written or operational message.
- Preserve the original email content and formatting as much as possible.
- Prepend the forwarded message with this exact tag line at the top: “Forwarding as received”
- Avoid forwarding the same email more than once.
- Log which emails were forwarded and which were skipped, with a short reason for each skip.
- Before enabling live forwarding, do a dry-run or test pass and report what would be forwarded.

I simply pasted this prompt into the chat, and OpenClaw created the skill for me. It also set up a cron job to run the skill automatically every day.

When the scheduled run executed, the skill produced the email in my desired format.

Problem 1: A skill does not show up

Covered above, but worth repeating because it is so common. Wrap the description in quotes, save, and check openclaw skills list again. Restarting the gateway alone will not fix it.

Problem 2: The agent keeps timing out

Timeouts usually mean one of three things. 

1. First, if you are running a local AI model, the local model server may simply not be running. Check that before anything else. 
2. Second, the model may be too slow for the job. Complex multi-step tasks need a faster or smaller model, and swapping the model in your config often fixes runs that keep dying. 
3. Third, the context window may be too small.

Problem 3: Runs stop because the context window is too low

The context window is how much text the model can hold in its head at once. If it is set too low, long tasks get cut off mid-run. Raise the context Window value in the models section of openclaw.json. One caveat from real-world testing: some local model servers ignore the value in this file, so you may need to raise the limit inside the model server's own settings too.

Problem 4: A scheduled job ran but returned an error

Scheduled jobs depend on outside services, and outside services have rate limits, which are caps on how many requests you can make per day. In the documented test, a morning job failed simply because a search API hit its daily limit. Check the error log, find which service complained, and either slow your schedule down or raise your plan's limit.

Also, one small warning: the jobs.json file inside the cron folder is managed by OpenClaw itself. Do not edit unless you know what you’re doing.

Useful Commands You'll Keep Coming Back To

You do not need to memorize the whole CLI. In practice, a small set of commands covers almost everything you will do day to day.

Two of these deserve a little more detail, because you will use them constantly.

openclaw config

is your control room. Instead of remembering where the config file lives and what its structure looks like, this command gets you straight to your settings. Model options, gateway options, skill entries, all of it.

openclaw gateway status

gives you a live snapshot of your gateway. Instead of digging through logs or config files, this command shows what’s running, what models are available, which providers are connected, and whether your gateway is healthy. It's the quickest way to verify that OpenClaw can reach the services it depends on.

Security is not the topic of this guide, but I cannot send you off without two minutes on it. Researchers found tens of thousands of OpenClaw setups exposed online in early 2026, mostly running default settings. The basics protect you from being one of them:

• Keep the gateway private. In your config, the gateway should bind to 127.0.0.1 (only your machine can reach it), never 0.0.0.0 (the whole internet can).
• Keep pairing mode on. It is on by default. It is the reason strangers cannot talk to your bot.
• Run the built-in audit once: openclaw security audit --fix handles the common gaps automatically.
• Be picky about community skills. Check who wrote a skill before installing it, or better, write your own. It is just a folder and a text file, remember?

That is it. If you ever expose OpenClaw to the wider internet or handle sensitive data with it, go read our dedicated security and hardening guide. For a personal setup on your own machine, the four points above cover you from basic threats.

Openclaw rewards a bit of patience. The install is genuinely a one-liner, but getting skills and scheduling right takes tweaks and tests. Set it up on a dedicated machine, go through WSL if you are on Windows, quote your skill descriptions, and keep the command cheat sheet handy.

Don't want to spend hours troubleshooting? Let us set up OpenClaw for you. We'll get it running, connect your tools, and build your first automations that you can start using immediately.