Michael Limberger

Using OpenClaw

You do not need WhatsApp, Telegram, or any messaging platform. The built-in Web Dashboard and Terminal UI work without external services, which is what you want for a local-only setup.

The Web Dashboard

Fastest way to start chatting:

Show me

openclaw dashboard

That opens your browser to http://127.0.0.1:18789/ with an authenticated URL (the gateway token is embedded). You get a full chat UI: type messages, see responses, and watch tool calls in real time.

For demos this is the best option. Visual, easy to follow, works in any browser, and needs no extra setup beyond what you already did.

The Terminal UI

If you prefer staying in the terminal:

Show me

openclaw tui

Interactive chat in the terminal, with slash commands:

/status       Check connection to Ollama
        /model        Switch models (e.g., /model ollama/qwen2.5-coder:14b)
        /new          Start a new conversation session
        /history      View past sessions

Prefix a line with ! to run a shell command directly:

!git status
        !ls -la

The TUI stays light and keeps everything in one window, which is handy when you are already living in the terminal.

One-shot CLI

For scripting or a quick question:

Show me

openclaw agent --message "List the files in my home directory"

One message, one response, then exit. Useful for automation or when you do not want an interactive session.

Add --thinking high when you want more detailed reasoning:

Show me

openclaw agent --message "Explain this error: <paste error here>" --thinking high

What you can do with it

Once you are in a conversation (Dashboard, TUI, or CLI), the agent can act on your behalf through tool calling. The model is not only generating text. It can invoke tools that touch your system.

Try these in order, escalating complexity:

Basic response test:
        "What day is it today?"
        Confirms the model responds at all. If you get nothing, jump to
        the troubleshooting section.

        Tool-calling test:
        "List the files in my current directory"
        The agent should invoke a shell command and return real filesystem
        results. If it only describes what it WOULD do, tool calling is not
        working. Check that your model supports tools (qwen2.5-coder:14b
        does) and that the "api" field in openclaw.json is correct.

        Multi-step test:
        "Create a file called test.txt with 'Hello from OpenClaw' inside it,
        then read it back to me"
        Tests file I/O and multi-step reasoning. The agent should create the
        file, read it, then report the contents.

        Code generation test:
        "Write a Python script that downloads a web page and counts the
        word frequencies, then save it as wordcount.py"
        Tests writing code and saving files.

+----------------------------------------------------------+
|  If the agent describes actions instead of doing them,   |
|  tool calling is broken. Check model compatibility       |
|  and the "api" field in openclaw.json.                   |
+----------------------------------------------------------+

Optional: connecting Telegram

If you want a more impressive demo, Telegram is the simplest messaging platform to add. It uses long-polling, so you need no public URL, no webhooks, and no port forwarding. Everything works behind your NAT.

1. Open Telegram. Message @BotFather. Send /newbot.
        2. Follow the prompts to name your bot. Copy the bot token.
        3. During openclaw onboard (or manually in the config), select
           Telegram and enter the token.
        4. Message your new bot on Telegram. You'll receive a pairing code.
        5. Approve: openclaw pairing approve telegram <CODE>

You can then chat with your local AI from Telegram on your phone. Messages travel through Telegram's servers (encrypted), but the AI processing stays on your Mac.

Important caveat: Telegram, WhatsApp, Discord, and Signal all need their cloud servers for transport. For a truly air-gapped setup with zero cloud contact, stick to the Web Dashboard or TUI.

Session management

OpenClaw keeps conversation sessions locally. Each session has its own context and history, stored at:

~/.openclaw/agents/<agentId>/sessions/

If Ollama crashes mid-conversation (it happens), do not panic. Session state lives independently of the model process.

Restart Ollama, send another message, and OpenClaw reconnects. You may need to re-warm the model:

Show me

ollama run qwen2.5-coder:14b "ping" --verbose

The conversation should continue where it left off.

Checking the logs

When something feels off, watch logs in real time:

Show me

openclaw logs --follow

In another terminal, send a message through the Dashboard or TUI. Watch for errors, warnings, or references to cloud providers (which would mean local-only config is not sticking).

Log files live at:

/tmp/openclaw/openclaw-YYYY-MM-DD.log

They are JSON Lines (one JSON object per line). For verbose debugging:

OPENCLAW_LOG_LEVEL=debug openclaw gateway run --verbose

That produces a lot of output. Useful for stubborn issues, overwhelming for everyday use.