Your First Conversation

From Chatting to Commanding

🔄 Quick Recall: In the last lesson, you installed OpenClaw in a hardened Docker container with five security layers. Your agent is running at http://127.0.0.1:18789/ — isolated, locked down, and waiting for its first instruction.

Now comes the moment of truth: your first real conversation with an AI agent.

This is different from anything you’ve done with ChatGPT or Claude. When you talk to a chatbot, the stakes are low — the worst that happens is a bad paragraph. When you talk to an agent, the stakes are real. It can move files, send messages, and change things in your world.

By the end of this lesson, you’ll be able to:

Connect WhatsApp or Telegram to your OpenClaw instance
Understand how agent communication differs from chatbot prompting
Have your first safe conversation with your agent

Connecting Your Messaging App

You have several options. We’ll cover the two most popular.

Option A: Telegram (Recommended for Beginners)

Telegram is the easiest to set up because it uses an official bot API — no workarounds needed.

Step 1: Open Telegram and search for @BotFather (the official Telegram bot manager).

Step 2: Send /newbot and follow the prompts to create a bot. BotFather will give you a token that looks like 123456789:ABCdefGHIjklMNO.

Step 3: In your OpenClaw control panel (http://127.0.0.1:18789/), go to Channels and select “Telegram (Bot API).”

Step 4: Paste your bot token and save.

Step 5: Open a chat with your new bot in Telegram. Send “hello” — if OpenClaw responds, you’re connected.

Option B: WhatsApp

WhatsApp integration uses the Baileys library to connect through WhatsApp Web.

Step 1: In the OpenClaw control panel, go to Channels and select “WhatsApp.”

Step 2: A QR code appears on screen.

Step 3: On your phone, open WhatsApp → Settings → Linked Devices → Link a Device → scan the QR code.

Step 4: Wait 30-60 seconds. You should see a confirmation in the control panel.

Step 5: Send a message to the linked number. If OpenClaw responds, you’re connected.

Multi-channel bonus: You can connect both (and Discord, Signal, and more). All channels share the same context and memory — your agent knows who you are regardless of which app you use.

✅ Quick Check: Why is Telegram recommended for beginners? (Answer: It uses an official bot API, making setup simpler and more reliable than WhatsApp’s web protocol workaround.)

How Agent Communication Differs from Chatbot Prompting

With ChatGPT, you write prompts like:

“Write a professional email declining a meeting invitation”

The AI gives you text. You copy it, paste it into your email app, edit it, and send it.

With OpenClaw, you write instructions like:

“Decline tomorrow’s 3pm meeting with Sarah. Tell her I have a conflict but suggest Thursday at the same time.”

The agent doesn’t give you text to copy. It sends the actual email. It opens your email integration, composes the message, and delivers it — all from that one message in Telegram.

The key shift: Stop telling the agent what to write. Start telling it what to do.

Here’s the mental model:

Chatbot Thinking	Agent Thinking
“Write me a summary of this article”	“Read this article and save a summary in my Notes folder”
“Create a to-do list for my project”	“Look at my project deadline, break it into tasks, and add them to my calendar”
“Draft a reply to this complaint email”	“Reply to John’s complaint email, apologize for the delay, and offer a 15% discount”

Notice the difference? Agent instructions describe the outcome, not the output.

Your First Safe Task

Let’s start with something low-risk. Here are three progressively bolder first tasks:

Level 1: Read-Only (Safest)

“What time is it right now, and what’s the weather forecast for [your city] today?”

This is a good connectivity test. The agent needs to access time and weather — no personal data involved.

Level 2: File Organization (Low Risk)

“Create a new folder on the desktop called ‘Test Folder’ and put a text file inside it with today’s date and a haiku about getting started with AI agents.”

This tests file creation in a controlled location. Easy to verify, easy to delete if something goes wrong.

Level 3: Information Synthesis (Medium Risk)

“Read the three most recent PDF files in my Downloads folder and give me a one-paragraph summary of each.”

This requires file access but is still read-only. The agent reads files but doesn’t modify, delete, or share anything.

Start at Level 1. Seriously. The temptation to jump straight to “manage my inbox” is strong. Resist it. Build trust incrementally.

✅ Quick Check: Why should you start with read-only tasks instead of jumping to email management? (Answer: Read-only tasks let you verify the agent works correctly without risking real-world consequences. If it misreads a weather forecast, nothing happens. If it misinterprets an email instruction, it could send embarrassing messages.)

How Memory Works (From Day One)

Every conversation you have with OpenClaw gets recorded in its memory system. Understanding this is important because what the agent remembers shapes what it does next.

OpenClaw uses a file-based memory system with two layers:

Daily logs: Every conversation goes into a file named by date (memory/2026-02-11.md). These are append-only — the agent writes to them during the day but the content is ephemeral. Yesterday’s log and today’s log are loaded at the start of each session.

Long-term memory (MEMORY.md): Over time, the agent curates important information into a persistent file. Your preferences, recurring tasks, important contacts — anything it thinks will be useful long-term.

Why this matters: The agent gets smarter over time. On day one, it might ask “What email address should I use?” On day thirty, it knows. On day one, it gives generic weather reports. On day thirty, it knows you bike to work and mentions rain risks.

The privacy implication: That memory file contains everything you’ve discussed. If an attacker accesses your machine, they get your conversations, preferences, and any sensitive information you’ve shared. This is why the Docker hardening from Lesson 3 matters — it limits what an attacker could reach.

Reading the Agent’s Actions

One habit to build from day one: watch what the agent does, not just what it says.

In the control panel at http://127.0.0.1:18789/, you can see:

What commands the agent executed
What files it accessed
What external services it contacted
How much it cost (in API tokens) per interaction

Make a habit of checking this after every interaction for at least the first week. Think of it like checking your credit card statement — you’re looking for anything unexpected.

Tasks to Avoid (For Now)

You’ll get to advanced workflows in Lessons 5 and 6. For now, don’t try these yet:

Task	Why Wait
Email management	Lesson 6 covers the security guardrails you need first
Installing skills from ClawHub	Lesson 7 teaches you how to vet them for malware
Connecting financial services	Too much risk before you trust the system
Running shell commands on your computer	Only through Docker — never give host access
Sharing access with others	Each person needs their own hardened instance

Key Takeaways

Agent communication describes outcomes, not outputs — “do this” instead of “write this”
Telegram is the easiest messaging integration for beginners (official bot API)
Start with read-only tasks and build trust incrementally (weather → files → summaries)
Memory accumulates from day one — the agent gets smarter but also stores more sensitive data
Monitor the control panel after every interaction for the first week
Avoid high-risk tasks until you’ve learned the safety protocols in later lessons

Up Next

Your agent is alive and talking. In the next lesson, we’ll build your first real automated workflow: a morning briefing that delivers weather, calendar events, and news headlines to your messaging app at the time you choose. Every single morning. Automatically.