OpenClaw

OpenClaw: Install and first job

This is the second of eight OpenClaw pages, after Pick it and plan. It installs OpenClaw on a Windows 11 PC or a Mac, turns on the safety switches, connects a free Gemini key, and gives the agent one small job.

These steps were checked on 14 November 2026, when the latest stable OpenClaw was 2026.11.3.

Install on Windows 11

The OpenClaw docs call the Windows Hub app the easiest desktop path, and it sets the agent up inside its own hidden Linux system (WSL). This guide could not confirm how to turn on the approval switch below for that setup, so this page uses the PowerShell path, where the safety commands are documented and run in the same window.

1. Open PowerShell

Click Start, type PowerShell, and open Windows PowerShell. Do not choose "Run as administrator."

What you should see a window with a line ending in > and a blinking cursor.

2. Paste the install command

Copy this line, paste it into PowerShell, and press Enter:

& ([scriptblock]::Create((iwr -useb https://openclaw.ai/install.ps1))) -NoOnboard

This downloads OpenClaw's official install script from openclaw.ai and runs it. -NoOnboard tells it to skip the setup questions for now, so you can turn on the safety switches first. Only paste install commands that come from openclaw.ai.

OpenClaw needs Node.js, a program that runs it, and if Node.js is missing, the script tries to install it for you.

What you should see lines of text scrolling past, then the > line comes back.

3. Check that it worked

Paste this and press Enter:

openclaw --version

What you should see a version number, such as 2026.11.3 or newer.

If PowerShell says running scripts is disabled, do not change that setting. Use this line instead:

& "$env:APPDATA\npm\openclaw.cmd" --version

If you needed that line, use & "$env:APPDATA\npm\openclaw.cmd" in place of the word openclaw in every command on this page.

Now skip to Turn on the safety switches first.

Install on a Mac

1. Open Terminal

Press Command and Space together, type Terminal, and press Return.

What you should see a window with a blinking cursor.

2. Paste the install command

Copy this line, paste it into Terminal, and press Return:

curl -fsSL https://openclaw.ai/install.sh | bash -s -- --no-onboard

This downloads OpenClaw's official install script from openclaw.ai and runs it. --no-onboard tells it to skip the setup questions for now, so you can turn on the safety switches first.

If Node.js is missing, the script installs it through Homebrew, a tool that installs Mac programs, and installs Homebrew first if you do not have it. If it stops with an error about Homebrew or Node, see the official install page.

What you should see lines of text scrolling past, then the cursor comes back.

3. Check that it worked

Paste this and press Return:

openclaw --version

What you should see a version number, such as 2026.11.3 or newer.

OpenClaw also has a Mac menu-bar app, but skip it for now. It needs macOS 26.2 or later, and once its permissions are granted, its computer control (clicking and typing on your screen) starts on.

Turn on the safety switches first

Out of the box, OpenClaw has its sandbox off and its shell approvals off. The shell is the part that runs commands on your computer, so with both off, the agent can run commands on your real computer without asking.

Do these five things before you connect a model.

1. Keep its own folder

OpenClaw's workspace is the folder the agent starts in, and by default it is ~/.openclaw/workspace. The ~ means your user folder, such as C:\Users\YourName on Windows. Keep that default.

Never point the workspace at your user folder, Documents, or Desktop. The workspace is a starting place, not a wall, so with the sandbox off, a command can still reach other folders if it names them in full, and so can the file tools until you finish step 3.

What you should see nothing to change. You are leaving the workspace where it is.

2. Turn on approvals

Paste this and press Enter (Return on a Mac):

openclaw config set tools.exec.mode ask

ask means the agent runs only commands on its allowed list, and asks you before anything else. You will check this again once the agent is running.

Leave the sandbox switch off for now, because turning it on needs Docker or a similar tool. The safety chapter covers it for later, and so does Advanced features.

Never set tools.exec.mode back to full. Also skip /elevated in the chat: it only matters once the sandbox is on, and then it moves commands back onto your real computer.

What you should see the command finishes with no error.

3. Keep the file tools inside the workspace

The approval switch covers commands only. OpenClaw's own file tools, which read, write, and edit files, never ask first, and out of the box they can write to any folder your account can open. Paste these two lines to keep them inside the workspace:

openclaw config set tools.fs.workspaceOnly true
openclaw config set tools.exec.applyPatch.workspaceOnly true

What you should see both commands finish with no error.

4. No email, bank, or chat bot in the first hour

What you should see no chat apps and no skills turned on.

5. Cap your spending

The free Gemini key has no card attached, so it cannot bill you.

If you later use a paid key from OpenAI or Anthropic, set a monthly spend limit on their website before you paste the key, because OpenClaw has no monthly dollar cap of its own. The screen names are in the safety chapter.

What you should see no paid key yet, or a monthly cap saved on the provider's website.

Connect a free model

1. Get a free key

Go to Google AI Studio and create a free key, then copy it. Do not share it, and never paste it into the chat.

What you should see a long key copied and ready to paste.

2. Run setup with the Gemini key

Paste this and press Enter:

openclaw onboard --auth-choice gemini-api-key

When it asks for your key, paste it there. If it asks Keep OpenClaw running in the background?, choose Yes, and skip any chat apps or extras it offers.

This guide could not confirm every screen the setup shows.

What you should see setup finishes. It may open a page in your web browser.

3. Start the agent in the background

If you did not answer Yes above, install and start OpenClaw's Gateway yourself by pasting these two lines, one at a time:

openclaw gateway install
openclaw gateway start

Then check it:

openclaw gateway status

What you should see the Gateway listening on port 18789.

4. Switch to a free Gemini model

Gemini key setup picks google/gemini-4-argon as the main model, but Google does not offer that model on the free tier. Switch to google/gemini-4-hydrogen, which is free:

openclaw models set google/gemini-4-hydrogen
openclaw gateway restart

If it shows an error, run openclaw models list --provider google and copy the name exactly.

If you run the Gemini key setup again later, it switches back to the paid model, so run these two lines again after it.

To see which Google models your key can reach:

openclaw models list --provider google

What you should see google/gemini-4-hydrogen in that list.

5. Check the approval switch again

openclaw approvals get
openclaw exec-policy show

What you should see an effective policy of ask, or allowlist with on-miss, which is the same thing. If the effective line says full with off, run openclaw config set tools.exec.mode ask again, then openclaw gateway restart.

6. Open the chat page

openclaw dashboard

This opens OpenClaw's chat page in your web browser. The page runs on this computer, so type Hello.

What you should see a chat page in your browser, and a reply from the agent.

Or use a local model

First install Ollama and download a model with Run a local model with Ollama. The steps below replace steps 2 and 4 above, while steps 3, 5, and 6 stay the same.

Run openclaw onboard, choose Ollama, then Local only, rather than Cloud + Local, which also routes Ollama's cloud models. When it asks for the Ollama base URL, use http://127.0.0.1:11434, which is Ollama on your own computer, and it needs no real key. Then check the list and make your model the main one, putting its name after ollama/:

openclaw models list --provider ollama
openclaw models set ollama/gemma4

What you should see your model in the list as ollama/ followed by its name, then a reply when you type Hello in the chat page.

Your first safe job

In the chat page, type:

In your workspace folder, write a new file named first-note.txt with three short tips for keeping a houseplant alive. Only work inside your workspace. Do not run commands, do not delete or move anything, and do not open any website.

If it asks to run a command, read the request, and deny anything that mentions delete, move, your user folder, Documents, Desktop, email, or a browser login. A file the agent writes with its own tools does not ask first, which is why step 3 of the safety switches keeps those tools in the workspace.

The request appears in the chat page as Command approval needed, with Allow once, Always allow, and Deny. Use Allow once or Deny, never Always allow, in the first week. If no approval screen can be reached, or you do not answer in time, OpenClaw refuses the command.

Then look at the folder yourself. On Windows, paste this into PowerShell:

explorer "$env:USERPROFILE\.openclaw\workspace"

On a Mac, paste this into Terminal:

open ~/.openclaw/workspace

You may also see files the agent keeps as its memory, such as MEMORY.md or USER.md. Leave those alone for now, because the next page shows how to read and change them.

What you should see first-note.txt in the workspace, and nothing changed anywhere else.

Sources

All checked 14 November 2026.

Next

Continue to Files, the browser, and memory.

Checked 14 November 2026 · prices, limits, and versions change