While some people are buying Mac Minis and handing over their whole digital lives to Open Claw, I’m not that trusting and don’t want to spend about £500 on something I probably won’t use in the long term. I still wanted to try Open Claw, so I decided to use a Raspberry PI 5 that I already had.

The idea was simple. I wanted to see what this agentic AI thing can actually do on a cheap, low-power device, without giving it access to my main computer, files, browser, accounts, and everything else that lives on my daily machine.

I like trying new tools, but I also don’t like pretending that every new AI project deserves full access to my life straight away. If something can run commands, download files, generate things, connect to chats and act on my behalf, then I want to test it somewhere separate first.

Installing OpenClaw

The first thing I did was install OpenClaw on the Raspberry Pi.

The installation process was pretty simple, but I did encounter a bunch of missing node module errors that weren’t difficult to fix (just install the missing module), but were annoying. I used the terminal and ran the install script (from their official site):

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

5 Things that I managed to do with it

• Creating a personal dashboard
• Run a cron job that would send me a reminder every 2 hours.
• Create/Edit Videos with Remotion
• Vibe code a very simple game
• Trade Crypto on Binance (It actually made some money for me)
• Create a YouTube to MP3 downloader where you give it a YouTube link, and it sends you back an MP3 file.

The Big negative

It turns out that running OpenClaw is actually very expensive in terms of API credits, and the longer you run it, the more it costs as it keeps sending all of your chat/session history with each request. In the few weeks that I have been playing around with it, I managed to run through over £26 of API credits using various Gemini models. I initially used OpenAI API, but that was even more expensive, so I switched to Gemini with the idea that, because it has a free tier, I could mainly use that and it would be cheaper. As far as I can tell, it only makes sense to run it if you have the LLM running locally. Unfortunately, on a Raspberry Pi 5, you can’t run any LLMs that are actually fast enough or smart enough for anything useful.

Final Thoughts

Open Claw is an interesting concept, but I don’t think it is something I would keep running all the time, at least not in its current state.

I like the idea of having an AI agent sitting on a small device that can help with random tasks, build small tools, generate files, edit videos, run scheduled jobs, and generally act like a personal assistant that you can message. Some of the things I managed to do with it were genuinely impressive, especially considering it was running from a Raspberry Pi 5.

But the cost makes it difficult to justify. Spending over £26 in API credits just by experimenting with it changed how I look at ot. At first, I thought the Raspberry Pi would make this a cheap setup, but the Pi is only cheap from a hardware and electricity point of view. The real cost is not the device; it is the model behind it.

That is the part that makes me hesitate. If every task keeps sending more and more context back to the API, then the cost slowly becomes unpredictable. For small tests, it is fine, but for something that is supposed to run all day in the background, I don’t want to constantly think about how much each interaction might cost. So for now, I see Open Claw as a fun and impressive experiment rather than something I would fully rely on.

If you already have a Raspberry Pi and some API credits to burn, it is definitely worth trying. It is a good way to understand what agentic AI can do, and it is much safer than giving access to your main computer straight away. But if you are thinking of buying new hardware just for this, I personally wouldn’t rush into it. Running local models is free, but it won’t be as powerful as the latest models from OpenAI, Anthropic or Google. Those are expensive to use, so in the real world, it’s not very practical.

Maybe the whole thing will make more sense once local models become faster, cheaper, and good enough to run on small hardware. Or maybe the project itself will improve how it handles context and API usage. But right now, for my use case, Open Claw is in that awkward place where it is powerful enough to be exciting, but expensive enough to not be usable for everyday use.

Here is a video that I made about it this: