What’s OpenClaw?
In the unlikely scenario that you’ve been living under a rock, OpenClaw is a self-hosted AI agent framework with integrations for email, chat, and more. It promises to be your best personal assistant.
Installing and Configuring OpenClaw
Let me be honest: it’s not newbie friendly. The install command looks simple, but trust me, that’s misleading. After installation, you need to configure tokens and settings to get the agent running. Needless to say, I’m not writing a tutorial for it.
The overall experience needs to be smoother for democratization. No wonder companies are jumping on the hype train with “one click” installations—just Google it. However, you’re better off trying it locally via Docker or using a VPS to limit potential damage.
My Current Setup
Here’s my current stack:
- VPS Server – Pick your poison.
- OpenClaw – Installed via npm script.
- Telegram – Fantastic and works out of the box.
- OpenRouter – My LLM provider. Easy to onboard and start using, no issues so far.
- LLM Model –
MiniMax-2.7. Decent overall, though sometimes you have to correct the agent. - Tailscale – How did I not know about this until recently? Thank you to this tweet:
Model Experiments
I tried several models:
- Claude – Super expensive. Burns through money like nothing. Hard pass.
- Gemini Flash – Runs quickly and cheaper than Claude, but reasoning can be problematic. It broke my OpenClaw instance a couple of times by corrupting
~/.openclaw/config.json. - MiniMax – My default model. Pricing is economical and it works most of the time. I’ve noticed it’s lazier than other models and tends to take shortcuts when there’s too much work. For regular chats though, it works very well.
- Kimi – No success with this one, even though it scores high on leaderboards. 🤷
Tips
Here are a few things I learned the hard way.
First up: turn on verbose mode (/verbose) for the first few days. You can catch the agent early if it’s doing something unexpected or burning tokens unnecessarily. The bot has full system access—it can edit its own configuration and restart services. It still feels unreal that it can modify itself. Unfortunately, it’s not foolproof; you can brick the installation quite easily.
When it comes to personas, experiment and see what sticks. I started with a genderless bot, then made it female with a nice Indian name. I had it generate a self-portrait for its Telegram profile pic! One feature request for Telegram: I wish bots could update their own profile pics and descriptions—it would make them feel more lively. As for WhatsApp setup, I skipped it since it requires a separate phone number.
If you code, install opencode or claude and ask OpenClaw to invoke these CLI tools. It works surprisingly well. But be careful with Google account access—I haven’t and won’t trust an agent without hard guardrails. I’m considering setting up a shadow account for email forwarding, giving the agent full access only to that.
A few other things: use Tailscale to set up an alias for your VPS and enable Tailscale SSH. It’s safe and super easy to log in. Lock down your Telegram settings—make sure your bot isn’t open to random strangers messaging it. And don’t upgrade OpenClaw impulsively. It’s still rough around the edges. I did and paid dearly with hours of debugging. See here and this:
Summary
I’m still playing with my personal assistant, mostly via Telegram, and learning every day. Not sure where this will go or what it will evolve into. But one thing’s for sure: the world will never be the same.