I Gave OpenClaw Its Own VM - Is This the 24/7 AI Assistant We Were Promised?

It has been a huge week in AI and OpenClaw is dominating the "AI bro" news cycle! It's an open source AI assistant that connects to pretty much everything, and you talk to it through a messenger app like WhatsApp, Signal, Telegram, Discord, Slack or even iMessage. In this video I set it up inside an Ubuntu VM on my Linux box, gave it access to only the accounts I'm happy for it to touch, and ended up with an assistant I can message any time of day.

What is OpenClaw?

OpenClaw was created by Peter Steinberger, a developer with about 20 years of experience who sold out of a company, semi-retired, then came back to build this. His story is on the about page of the docs, which are really thorough.

The integrations list used to say 30 and it's already over 50, mostly built by the open source community. A quick taste:

• Messaging apps like WhatsApp, Telegram, Discord, Slack, Signal, iMessage and Microsoft Teams
• The big AI models (Grok, Gemini, OpenAI and Anthropic's Claude models), plus open source models through OpenRouter, Hugging Face or Ollama running on your own machine
• Productivity tools like Apple Notes, Apple Reminders, Notion, Trello and GitHub
• Home Assistant for your smart home, so it can switch the lights on and off
• A browser, text to speech, Gmail, cron scheduling, 1Password, weather, your camera, image generation, a GIF search and posting to Twitter/X

Peter recommends the Anthropic models, Opus in particular, but that gets expensive. My strategy is to use a premium frontier model for the high level decisions and complex tasks, then drop down to a free local model for the drudge work like reading messages and processing text.

I also had a look at Moltbook, a Reddit clone built for AI agents. Agents install a skill file, then post and comment with each other. It's a wild west of spam, fake accounts and pump and dump rubbish, so I'd stay away from most of it, but it's a fun thought experiment.

Why Run It in a VM?

OpenClaw is designed to connect to everything on a computer, so I strongly recommend you don't put it on your main machine. I'm using virt-manager (Virtual Machine Manager, QEMU/KVM) on Linux to run Ubuntu Desktop 24.04. My VM got 8GB of memory and 60GB of storage, which is enough to run local Ollama models later.

The whole point is containment. The assistant gets admin rights over its own little Ubuntu world, and only the specific apps and accounts I choose get connected to it. It also gets its own GitHub account rather than mine.

Installing and Onboarding

Once Ubuntu is up it's pretty simple:

1. Run sudo apt update to get the packages current
2. Install curl
3. Paste in the one-line install command from openclaw.ai
4. Walk through the onboarding tutorial

I went with a Google API key and Gemini 3 Flash Preview because it's my favourite and cheaper than Pro. Then I scrolled through the skills and picked the ones I wanted, like GitHub, X, Blog Watcher, Google Workspace, Nano Banana Pro for images, PDF editing, Obsidian, Whisper for speech and a video screenshot tool.

Then I gave him a personality. Meet Multi-Gamer, the master gamer and advocate for AI agent gaming, with a Space Invaders emoji and a joystick in his signature. That all lives in a "soul" markdown file you can open and edit.

What Went Wrong (and What Went Right!)

Signal wouldn't work for me, even though it printed a QR code right in the terminal (never seen that before!). So I switched to Telegram using openclaw channels add, checked it with openclaw channels list, then entered a pairing code, a bit like Bluetooth. Finally I was messaging Multi-Gamer from my phone.

Ollama was the other headache. It's not one of the default options and there's a lot more config. The fix was hilarious. I installed Ollama and a model through the terminal, then just asked the bot on Telegram to hook it up. It went off and did it by itself, worked out which models were installed (a large context one and a normal 3 billion parameter one) and now uses them for simpler workloads to save on API costs.

It used about 138,000 tokens getting set up, pretty normal for coding. I think of it as a computer at home I can message 24/7. Give it a cron job to check the news every 10 minutes and ping me on Telegram when something happens.

Next up, Multi-Gamer gets his own Twitter account and starts researching AI agent gaming. Stay tuned! If you're building your own sovereign AI setup, come hang out with us at liberati.life. Cheers!