First things first, let’s address some hype that seems to have become gospel thanks to influencers. It’s amazing that even the smartest of my friends just bought into the hype without fact-checking.
Let’s debunk some OpenClaw mis-truths.
You need a mac mini to run it. You don’t. I have mine running on a VM with 4GB running tens of docker containers. It works great. The container has about 40GB of RAM assigned. The best set up is either a docker container or a standalone VM.
OpenClaw is unsafe. It’s mostly not. For the average person, it may be. If you work in tech, probably not. OpenClaw is a pretty simple JS app that just routes things places and has a heartbeat and some memory. There’s nothing inherently crazy about that. As with all software, you need to know what it does and be careful with it. Copying and pasting command line instructions that you don’t understand can be more risky. For my install I did not enable ClawHub to install new skills. That said, they are all now scanned by VirusTotal, so they should be safer. For me though, it’s simple enough to just ask OpenClaw to make new skills for itself – it’s really good at it!
It’s expensive to run. Again, it’s not. There are so many stories of people racking up huge bills with Opus 4.6. You need not worry. You can run OpenClaw with a selection of free models from OpenRouter. I started this way. I also found Gemini 3 Flash to be surprisingly capable and cheap. I’ve been running non-stop for a week and haven’t even spent £15 ($20) yet. Even then, Google gives $300 of cloud credit, so it’s actually been free. I highly recommend you start with a free model or the Gemini 3 Flash model. Once you’re vibing with your agent, then you can experiment with adding more agents that have bigger and better models to help you complete those more demanding tasks.
Installation
Ready to get started? Great!
There are three approaches I recommend for running OpenClaw, although I’ve only tried one of them. The other two are listed because they seem sensible and give you some options.
1. Use Docker
Docker worked best for me. The app runs mostly isolated from the host operating system. I can give OpenClaw access to other services on my network, if I want to. I spun up a browser container that it uses to do things in Chrome. Installing docker is fairly straightforward, so this is my recommended way. There are some gotchas with this approach e.g. updating might be difficult, I haven’t crossed that bridge yet. Also if you let the agent install things, they might disappear if you rebuild the container, because they’re kind of ephemeral unless you mount the storage.
Installation was a little bit of a pain because I decided the build the docker container myself and it took ages and kept hitting some odd permission errors. It worked out in the end though. There are pre-built images, so you can avoid that pain, but I wanted to be really sure the code I was running was the real thing, especially if I’m going to tell my agent my deepest darkest secrets.
TIP: If you do install via docker, you only need one container. You don’t need the openclaw-cli one. If you attach to the container you won’t be able to run openclaw instructions e.g. openclaw doctor –fix. Instead you’ll need to run node dist/index.js doctor –fix. This is a weird quirk of how the Docker build works, not sure why. Probably a bug.
2. Use a Virtual Machine (VM)
If you have Proxmox or similar installed on a machine with enough resources, this is a great option. You can just give OpenClaw an entire virtual machine. Installation is also really easy, just run the one-line install script from openclaw.ai.
3. Use an old laptop, raspberry pi or similar
Similar to the VM approach, make a new user or just wipe your device and give it a fresh operating system. Then run the one-line install command.
4. Use a VPS (not recommended)
Using a VPS may be tempting but I don’t recommend it. Why would you store your “memory” in the cloud? There’s something nice about it being stored on a machine in your home. I haven’t tried the various VPS providers and their installers, but I have seen some stats that many users are accidentally exposing their OpenClaw to the open web. Not great for security. If you don’t know what you’re doing or value your data and want the most control, avoid this option. Everyone on youtube will recommend this path…probably because they get paid for the referral.
What about a mac mini?
Yes you can use one. I think it’s overkill. The only reason to go down this path is to use local models with OpenClaw. But you’ll need a lot of RAM. My experience with a 16GB M3 Macbook was not good. I could run the models for basic chat but OpenClaw wouldn’t work with them. The real power of OpenClaw is the ability to hot swap models. Just use the cloud ones for now, they’re great!
Tips
openclaw onboard: You can run openclaw onboard at any time to reconfigure your whole setup. I’d only do this once though. But while you’re setting up your node, rerun if you need to get things right. If you’re on docker it’s node dist/index.js onboard.
openclaw configure: run this at any time to tweak your setup. I did use it to install skills, but it seems most of the built in skills only work on macOS with homebrew. As mentioned above, I just build my own skills. You can of course, copy and paste the SKILL.md from the official OpenClaw repo. The agent can make any python or js code that the skill might require.
0 Comments