cborgia.me
← Back to the log

OpenClaw Tips and Tricks

OpenClaw is a self-hosted AI agent you can text from any messaging app — WhatsApp, Telegram, Slack, iMessage, whatever. You point it at a model (Claude, GPT, Gemini, or a local one via Ollama), and it becomes a personal assistant you can delegate real tasks to.

I've been running it for a few weeks. Here's what I've picked up.

Start on your laptop

You don't need a dedicated server to try it. Install it locally, connect Telegram, and start handing it tasks. Move to a Mac Mini or a VPS later if you want it running 24/7.

npm install -g openclaw@latest

Requires Node 22.16+. If you're on Ollama 0.17+, there's a one-command setup that handles everything.

Don't skip SOUL.md

SOUL.md defines your agent's personality — tone, verbosity, communication style. Five minutes configuring this saves weeks of the agent not sounding right. It's the equivalent of a system prompt you actually commit to.

Think delegation, not search

The biggest mindset shift: frame tasks as things you're handing off, not questions you're asking. Instead of "what's the weather," try "check the weather and if it's going to rain, remind me to bring an umbrella at 7:30am." OpenClaw is an agent, not a chatbot.

Give it its own accounts

Set up separate email, calendar, and storage accounts for your agent rather than handing over your personal credentials. This is both a security practice and an organizational one — you want a clear boundary between what you did and what your agent did.

Security basics

This one matters. OpenClaw can access everything on the machine it runs on.

  • Don't install on a machine with sensitive data unless you've locked down permissions
  • Bind the gateway to localhost only and use SSH tunneling for remote access
  • Enable explicit consent mode so it asks before running destructive commands
  • Never expose the web interface to the public internet without auth
  • Use a strong model — better models resist prompt injection better

Run clawdbot security audit --deep after setup to catch common misconfigurations.

First message trick

Before asking it to do anything, send this:

"Hey, let's get you set up. Read BOOTSTRAP.md and walk me through it."

This gets the agent to orient itself in your environment before diving into tasks. Much better than starting cold.

Use Threaded Mode on Telegram

If Telegram is your primary interface, enable Threaded Mode via BotFather. This keeps conversations organized — one thread per project or task instead of everything in a single chat stream.

Voice messages are underrated

With an OpenAI API key, OpenClaw uses Whisper for speech-to-text. Send voice messages while driving, walking, cooking — whatever. It transcribes and acts on them. This turned out to be one of the most natural ways to interact with it.

Just tell it to remember things

OpenClaw has persistent memory. Say "remember that my dentist appointment is Thursday at 2pm" and it stores it. No special syntax, no config. It just works — and it'll reference it later when relevant.

Start with what annoys you

Don't try to build a grand automation system on day one. Pick the one repetitive task that bothers you most — summarizing emails, checking a dashboard, organizing links — and automate that first. Then expand.

Explore ClawHub for skills

ClawHub is the skill marketplace — 5,400+ community-built skills. Browse it before building something from scratch. There's probably already a skill for whatever you want to do.

Lobster for workflows

Lobster is OpenClaw's workflow engine — it chains skills into pipelines. Think "every morning, check my calendar, summarize today's meetings, draft prep notes, and send them to me on Telegram." That's one Lobster workflow.

Useful resources