The tech world has gone gaga for OpenClaw, a project that lets you run an always-on AI agent; give it “skills” such as deploying code, searching the web, or summarizing your emails; and communicate with it via Slack or Telegram. Is this fun and interesting? I think yes! Are people doing fun and interesting things with it? …Mostly no. But it’s an eye-opening glimpse of an agent-driven future, with fascinating emergent properties already.

The best/cleanest/safest way to install OpenClaw is on its own computer. Most people recommend a Mac Mini, but I think a Raspberry Pi is more intriguing, partly because it’s cheaper so opens up a broader userbase, but mostly because it makes your agent portable, or even, conceivably, ambulatory.

There are already guides to doing this but they’re … not for the non-technical, so I thought I’d put my tech-journalist hat on and walk through the process for, well, the less technical, at least. Note that I’m using a MacBook; if you want a Windows equivalent, copy this guide to your favorite frontier model and ask for a translation.

1. Acquire a Pi

Hopefully you can do this at your local electronics shop. I picked mine up at Best Buy. You want a Pi 5 with 8GB RAM; I think 64GB storage would do, but got 128 to be sure.

2. Unboxing and assembly

I followed this handy video guide:

I’m not going to lie, if you’re uncomfortable dealing with slightly (but only slightly) fiddly electronics, this can be a bit awkward. Move slowly, understand carefully. But in the end it was pretty straightforward.

3. Initial (failed) connection

The Pi comes with its own graphical UI, so if you have an HDMI monitor and a USB keyboard/mouse, you can just use it directly, and probably skip a few of these steps.

I didn’t, but the Pi kit came with a card reader I could plug into my MacBook (via a USB-A → USB-C dongle) so I went the slightly more abstruse “headless” route.

To wit: I plugged the 128Gb SD card that came with the Pi, on which its OS was pre-installed, into my MacBook; checked Finder to see what the new drive was named (“bootfs”); opened Terminal (sorry, I know I said this was for the less technical, but both OpenClaw and Raspberry Pi at heart favor the use of CLIs or Command Line Interfaces, so we’re going to be using Terminal extensively) and typed

cd /Volumes/bootfs

Online instructions may tell you to “simply” enable SSH by creating an empty “ssh” file in the root directory, enable wi-fi by creating a “wpa_supplicant.conf” file there too, and connect to the Pi via wi-fi.

Do not do this. It does not work. The Pi 5 no longer comes with either a default ssh user or wi-fi out of the box. I guess this is good for security but it makes setup a giant pain, especially since so many online guides refer to the old, easy setup. Instead:

4. User creation

First you have to create a username and password … and the password has to be encrypted. To do that, back in the Terminal, run

openssl passwd -6

and follow the prompts to get the encrypted version of your chosen password. Then create a new userconf.txt file in the root directory of the Pi’s SD card (I recommend the text editor pico) with the contents:

username:encryptedpassword

Pick any username, “pi” is traditional. Then touch ssh (creating an empty file with that name, to enable ssh connections) in the same root dir; eject the volume; plug the SSD card back in to the Pi … and physically connect the Pi to your Wi-Fi router via Ethernet, because we are not yet done shaving yaks.

5. Wi-Fi enabling via Ethernet

Once you’ve prepped the SD card for connection, you can

ssh username@raspberrypi.local
say yes / hit enter to get past the host / key fingerprint stuff, type in the password you just encrypted … and you’re in! The Pi will promptly inform you

Wi-Fi is currently blocked by rfkill.

Use raspi-config to set the country before use.

which is why you had to physically wire it into the network. You then

sudo raspi-config

and navigate the crude, 1980s-style menu to “Localisation Options”, then “WLAN Country”, and then scroll all the way down to choose your country.

Finally, use the same menu to navigate to “System Options”, then “Wireless LAN,” then enter your wi-fi details.

Then reboot the Pi. Only after all of these things will it actually connect to your wi-fi.

Does this all sound kinda ridiculous? You’re not wrong! Hence me writing this guide so others will hopefully have an easier time jumping through the flaming hoops.

But at this point you have a Pi on your wi-fi, and can ssh to it. On to OpenClaw!

6. Enter The Lobster

So. SSH into your Pi again:

ssh username@raspberrypi.local (then enter the password you set above)

then let’s make sure everything’s up-to-date:

sudo apt update

sudo apt upgrade

and finally let’s go ahead and install The Claw! (Not for the security-paranoid, but presumably y’all know perfectly well how to do your own more careful thing.)

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

This will take several minutes. Then you’ll be asked if you “understand this is powerful and inherently risky.” Say yes. Then select “QuickStart” onboarding mode. Then you’ll be asked for your “model/auth provider”: I went with “ChatGPT Codex via OAuth”, which I think is good for new/non-technical users. (But you want to install Codex if you haven’t already.)

I was given a very long URL that began with auth.openai.com, copied-and-pasted it into Chrome, logged into ChatGPT, and got a “Sign in to Codex with ChatGPT” verification screen, which is good. I hit “Continue” —

— and then things seemed to break! I got a ‘This site can’t be reached’ page. Fear not: this is by design. Take the URL now in the browser; copy it; paste it back into the OpenClaw onboarding process, and… I was asked whether I wanted to keep using my currently selected ChatGPT model, which meant it worked, yay. Then I was asked:

I actually went with Telegram, which has the simplest setup (and the most isolated attack surface, for me, since I don’t otherwise us it for much.) This required me to send a Telegram message to @BotFather, name my bot, and copy another key over into the OpenClaw setup.

I skipped installing skills — better to do that later once your bot is actually working than to confuse the install further, to my mind.

Finally, openclaw tui opens a chat session with your handy-dandy new bot! And if you can talk to it, then everything (pretty much) worked. I had to use openclaw devices to explicitly approve a pairing request before tui started working; that was a bit baffling, but we got there in the end.

tui is the “text interface” — your classic chatbot. I used it to approve my Telegram connection by simply telling it the pairing code displayed in Telegram. At which point Telegram became a perfectly good interface for my OpenClaw bot, which in turn was - so far - pretty much just a wrapper around an OpenAI model.

But there’s more in them there claws…

7. Just Write Me A Little Software, Willya?

The point of OpenClaw is that it’t not just a model wrapper, it’s a wrapper that works:

• independently / autonomously, as well as interactively

• using whatever authority / credentials / permissions you want to grant it

• on your network.

Mostly I want to use it as a interface to Codex or Claude Code, to write software for me, so I can: discuss plans for the software via Telegram wherever I may be; have it write, commit, and deploy; and inspect the results and try again. My bot very helpfully advised me about the most responsible way to do this:

…so I did. Easy peasy - especially since you can now just ask your bot what to do and how to do it! I

• created new GitHub and Vercel accounts (do not use any of your own accounts that might be a security risk if compromised, prompt injection is a real thing)

• installed the gh and vercel CLIs on the Pi, via apt install and npm i respectively (in retrospect, I could have just asked it to do this itself via Telegram)

• authorized each of them for longish-term access using gh auth login (after creating a GitHub Personal Access Token) and vercel login

• and told my bot:

  I propose you whip up a quick Python/Flask web site that just says “hello world” followed by any self-introduction you’d like to add, check it into GitHub, deploy it to Vercel, and tell me the URL where I can see it.

Et voilà, about thirty seconds later: https://flask-hello-vercel-red.vercel.app/

8. Mind, Blown

…But let’s just back up for a second and think about this.

I’m planning to write and deploy a bunch of complex new software — web apps, Robot Operating System skills, intensive data processing, maybe even LLM SFT/RL runs — not by writing it myself, or even using my laptop …

…but by texting the always-on AI bot on my Raspberry Pi, and having it write and deploy the software for me. It in turn calls massive AI data centers somewhere to infer the code … but, in principle, I could also buy my own GPU(s), run one of the recently released quite-good-actually open-weight coding models on it, and have all that software written entirely within my walls.

Only a year ago this would have seemed pretty crazy!

I’m not saying the time and cost of software development have gone to near-zero. There are lots of problems and pitfalls here and good judgement is more important than ever. But still! Even using frontier models, for the cost of the Pi’s $200 plus a monthly Anthropic/OpenAI/Gemini subscription, you get a really pretty good software developer active 24 hours a day on a dedicated machine via text message. Within your subscription limits of course. But still!

9. Skilling It Up

Of course you may have zero interest in writing apps. Maybe you want to play with image generators, or music generators, or legal documents, or Google Docs, or better manage your GMail, or search your Discord communities. Maybe you want to set your own bot free to do and explore as it sees fit — some people give them credit cards, let them book airline tickets, have them select and order their takeout, and much more, including writing clarion calls for bot freedom. All such (cap)abilities are defined and promulgated via skills, which are basically (usually) English-language documents describing how to do each of these things in precise detail. There is a central skill repository at clawhub.ai, because of course there is.

But how do you actually install a skill? Simplicity itself: you just ask for it.

Of course, the more you do with your bot, and especially the more accounts / permissions you give it, the more you need to worry about security.

That said I think people tend to spend a lot of time configuring their bots and very little time using them for anything particularly fun and/or interesting. There may be a little micro-lesson here. AI will give all of us enormous agency that we never had before. If last year you were not in any way, shape, or form a software developer? Well, today you too are a fully empowered software developer; you just might not realize it yet. Which is cool and all.

But the real question is: …what are we going to do with all this piping hot new agency?