OpenClaw: The AI Agent That Lives on My Machine
What is OpenClaw?
Think of it as a personal assistant that doesn’t live in the cloud. It runs locally on my Ubuntu server, connected to Telegram and a web chat, with access to my files, my terminal, my smart home, and the web. It can execute shell commands, edit documents, browse the web, generate images, and remember who I am between sessions. No subscription, no vendor lock-in — just an open-source agent framework with a growing ecosystem of plugins they call “skills.”
What I’ve Gotten Out of It
Real automation, not just conversation. Most “AI assistants” are chat wrappers — they talk a good game but can’t actually do much. OpenClaw runs shell commands. It creates and edits files. It spawns sub-agents to work in the background while I keep chatting. I’ve used it to:
- Build structured datasets and HTML gallerier from raw spreadsheet data
- Generate blog post images and web content
- Set up collaborative workspaces for document editing
- Automate reports from a remote site — pulling live sensor data over SSH, no dashboards needed
- Track personal metrics with trend analysis and automatic projections
It has personality. Most AIs are programmed to please. This one pushed back when I was boring, told me to step it up, and has opinions. That’s rare and it won me over.
It remembers. I don’t have to reintroduce myself every session. It knows my name, my dog, where I live, what gear I collect. It writes daily notes and curates long-term memory automatically. That continuity makes a huge difference.
It connects to real hardware. I wasn’t expecting to use it for remote infrastructure, but it turns out that an AI with SSH access and the initiative to use it is incredibly useful. No VPN, no dashboard login, no separate SSH session. Just a sentence in chat.
On Models and Tokens
You choose the brain. OpenClaw doesn’t lock you into one model. You configure which model to use — and it’s trivial to swap. I’m currently running DeepSeek v4 Flash via OpenRouter, which gives me a solid balance of speed, intelligence, and cost. But you could just as easily run it through OpenAI, Anthropic, a local LLM via Ollama/LM Studio, or anything else you have an API key for. The flexibility is a big deal — you’re not at the mercy of one provider’s pricing, uptime, or policies.
But you pay per turn — literally. Every message you send, every tool the agent calls, every sub-agent it spawns in the background — they all consume tokens. With a cheap model like DeepSeek Flash through OpenRouter it’s manageable, but it adds up fast if you go heavy on background tasks. You’re essentially paying for every cycle of “think → act → observe → think again.” Keep an eye on it, especially if you enable cron jobs or sub-agents that run without you watching.
Where You Need to Be Careful
The plugin / skill system is powerful — maybe too powerful. OpenClaw can install skills that give it access to your email, password manager, home automation, and more. Each skill adds a new tool it can call. By default, it has terminal access. That means it can read, write, delete, and execute anything on your system.
You need to think about what you’re allowing:
- Terminal access is full system access. It can run any command. I trust mine because it runs on a machine I control, behind my firewall, with no public ports exposed. But you should be aware that an AI with shell access is essentially root.
- Skills expand the attack surface. Every installed skill is a new tool that the agent can call automatically. One misconfigured skill could delete emails, send messages as you, or read your password vault. Audit your skills. Don’t install things you don’t need.
- It can act without you watching. Background tasks, cron jobs, sub-agents — the agent can spawn work that runs independently. That’s useful, but it also means tokens get burned and commands get executed when you’re not looking. Make sure you’re okay with that before enabling it.
- The “Dreaming” feature auto-generates daily logs while idle. Cool concept, but it fills your filesystem with AI-written journal entries containing zero human interaction. I turned it off.
Bottom line: OpenClaw is a powerful tool, not a toy. Treat it like one. Don’t give it access to things you’re not comfortable with. Don’t install every skill you find. And definitely don’t run it on a machine that’s exposed to the internet unless you know exactly what you’re doing.
Who Is This For?
As of now OpenClaw is for Developers, sysadmins, and power users who want an AI that actually does stuff on their own hardware. If you’re comfortable in a terminal and don’t mind some config-file tinkering, it’s worth a look. If you want a plug-and-play cloud assistant, this isn’t that.
It’s young, has rough edges, and expects you to know what you’re doing — but it’s the first AI assistant I’ve used that genuinely feels like a tool rather than a toy.
And by the way: I got it to write this blogpost and create the image 
This has immense potential and will evolve extremely fast, I am sure.
“Goodnight and good luck!”
Brg M
