Here be dragons
Running a LLM based AI agent with shell access is dangerous. There is a risk of prompt injection or adversarial input getting sent to the bot via the enabled communications channels or other text it finds online. The bot could run malicious software, or expose API tokens and other private information that it has access to if proper safeguards are not implemented and maintained. Read and follow the advice on the OpenClaw docs security page and be warned that this is cutting edge technology that is rapidly changing. LLMs are notorious for hallucinating and behaving unexpectedly. Even if you follow all of the best practices, there is still some risk with this type of tool.
Use the Raspberry Pi Imager to load a fresh image of Raspberry Pi OS (64-bit) onto a micro SD card. The Raspberry Pi 5 with 8GB RAM is the only model of Pi that has been tested and known to work. Once the SD card is ready, insert it into the Pi 5 and boot up.
Run these commands to update all of the pre-loaded software.
sudo apt update sudo apt upgrade
Installing OpenClaw can be done by downloading the install script from the internet and running it in bash. The "Quick Start" section of https://openclaw.ai/ shows a command that can be used to do this as well as a few alternative installation methods. This is the one-line command to install it:
curl -fsSL https://openclaw.ai/install.sh | bash
curling a shell script and piping directly to the shell is typically considered exercising poor security practices. If you have concerns, you can get the install.sh file then scan it for security issues. Since this guide generally is "off the beaten path", you should follow what security you think best.
The script will set up node.js and all of the other requirements that OpenClaw needs. The install time will depend on your network connection. It took about 2-4 minutes to install over WiFi during testing.Ā
Once the install is complete, the onboarding process will start automatically.
The first step is acknowledging the security risk inherent to using the software.
Read the warning and make sure you understand it. It is also recommended to read the security information atĀ https://docs.openclaw.ai/gateway/security. If you do not understand these risks or do not feel comfortable accepting them, you should skip this project.
If you do not understand these risks or do not feel comfortable accepting them, you should skip this project.
To move forward you must use the arrow keys to select Yes and then press enter.
Leave the default QuickStart onboarding mode selected and press enter.
Select the Model/auth provider that you wish to use. Anthropic with the Opus model is the only one that I tested on the Pi, but there are several other choices if you use a different LLM agent service. This guide documents Anthropic only. If you are using something different you will need to follow along with the wizard andĀ docs here, and here for your provider.
Use arrow keys to select Anthrophic and then press enter.
The next prompt is for the Anthropic auth method.Ā Leave the Anthropic token (paste setup-token) option selected and pressĀ enter. If you already have a Claude Code Anthropic token, you can paste it when prompted.
Using Claude Code CLI To Get an Anthropic Token
If you don't yet have a token, then you'll need to access it using the Claude Code CLI. You can do this on the Pi itself, or any other Linux computer.
To do it on the Pi, open a new terminal window or SSH session so that you can leave the OpenClaw onboarding process paused where it's at. You'll return back to it once you've got the token.
Install Claude Code CLI with this command.
curl -fsSL https://claude.ai/install.sh | bash
Once the installation is complete run this to set up your token.
.local/bin/claude setup-token
It will print an authorization URL starting with https://claude.ai/oauth/authorize? copy the URL and open it in a browser. The browser can be on any PC, it doesn't have to be on the Pi itself. If you are already logged in to Anthropic/Claude on a specific PC or browser, then using that will be easiest.
When you visit the URL in a browser, you'll be prompted to login if you aren't already. Depending on how your account is set up you may be asked about what organization to connect Claude code under. Follow the directions on the page to continue the process.
You will be prompted to authorize Claude Code to connect to your account and warned that it will contribute to your plan usage.
Click the Authorize button.
After authorizing, you'll be given code made of random characters to copy. Paste the code into the Claude Code prompt then press enter. Note that it needs to be pasted into the Claude Code setup-token CLI utility, this is not the token for Open Claw yet.
OpenClaw could use a significant amount of AI tokens, resulting in fees. If you cannot afford significant costs, consider only giving OpenClaw set amounts of tokens which you are comfortable paying for.
After entering the code you'll be presented with an OAuth token that is valid for 1 year.
The token is a string that starts with sk-ant-oat01- and contains many random characters after that.
Keep your OAuth token secret. Store it in a password manager or other secure storage system (not on the P5 itself). If someone else were to get this token they could use Claude on your account and drive up your usage bill!
Copy the token, being careful to ensure you get the full token and no extra spaces or other characters surrounding it. Return to the Open Claw onboarding wizard and paste the Anthropic OAuth token in the prompt then pressĀ enter.
The next step prompts you to name the token if desired. Leave the "default" name and press enter.
The next prompt is for which model to use. As of 1/30/26, the default model is anthropic/claude-opus-4-5. As newer models come out, the default may change.
Leave the default option selected if you are unsure, or select your model of choice if you have experience with them and prefer one over the others.
Use up/down arrow keys to select a model and pressĀ enter once the desired model is selected.
Next is channel selection. The channels are communications services that OpenClaw supports. Telegram, WhatsApp, Discord, and many others are available.Ā
Use arrow down to select the Skip for now option and press enter.Ā
Channels are optional. It's possible communicate with the bot over its local web based management portal or the terminal.
You can always connect channels later with the openclaw channels add command. Each channel requires slightly different steps, but the wizard walks you through them. You can also references the channels openclaw docs, and search for the specific channel in openclaw docs to get more specific info.
The first skill prompt is for Homebrew. Since Homebrew is a Mac utility and we're running OpenClaw on a Raspberry Pi, this is not needed.Ā
Select "No" and then press enter.
Next is the preferred node manager for skill installs.
Leave the default option npm selected and press enter.
Next is a huge list of skills that are available for OpenClaw to integrate with other applications and services. Many things are possible without any of these tools, and you can always install them later if you want to use some.
Select the "Skip for now" option, press space bar to check the box for it, thenĀ press enter.
Select "No" and press enter when prompted for and API key for goplaces.
Continue doing the same for the remaining API key prompts. Select "No" and then press enter.
All of these can be configured later if you decide to add the skills associated with them. But none of them are required just to use Open Claw.
Next is a prompt for Hooks setup.
Use the arrow keys and the space bar to check the boxes for "command-logger" and "session-memory" then press enter.
Once the everything completes, it will output some useful information in 3 different panels.
Control UI - Contains the local IP address URLs to the web based control panel for OpenClaw. You can copy/paste the link with the full token included into a browser on the Pi to open the control panel. The token in the screenshot is obscured because it's a private key, like a password. Keep it secure, anyone with it who can access the Pi on the local network could issue commands to the bot.
Start TUI - Is noted as the best option. It will launch a session to chat with the bot right there in the terminal.
Token - Contains information about the Gateway token that is included in the long URL (obscured by red box in the screenshot). It has reminders of where it is stored and other ways to access it.
Finally! The very last thing it will do is ask how to hatch the bot.
Since this whole setup process has been in the terminal, go ahead and select Hatch in TUI and press enter. This is the simplest option for the initial test to make sure things are working.
Once confirmed to be working, you can then access the web based control panel for a more graphical based UI to configure and interact with the bot.
The bot will wake up and ask you about yourself, as well as ask about a name to call it.Ā
I told it a bit about myself and that I was experimenting with OpenClaw on the Pi 5 to write a learn guide. I let it choose a name for itself and we settled on using the name Pipbot.
Page last edited February 03, 2026
Text editor powered by tinymce.
