← Back to blog

The Ultimate OpenClaw Tutorial (2026): From Setup to Automating Real Tasks

| Weihao Ming

This is the OpenClaw tutorial I wish existed when I started. It covers both setup paths (managed cloud or self-hosted Docker), walks you through your first five tasks, explains the key concepts you actually need to understand, and finishes with real workflow examples you can steal. Whether you have three minutes or thirty, you will have a working agent by the end.

Part 1: Getting OpenClaw running

There are two paths. One takes three minutes. The other takes thirty. Both get you to the same place: an AI agent you can message from your phone that has a real computer at its disposal.

Path A: OpenClaw Cloud (3 minutes)

This is the fast lane. OpenClaw Cloud is managed hosting — we handle the infrastructure, the containers, the updates, the security. You just connect and start talking to your agent.

  1. Join the waitlist. Head to the Cloud signup page and enter your email. Invites go out in batches, usually within a day or two.

  2. Connect your API key. Once you are in, the onboarding bot on Telegram walks you through linking an LLM API key. You bring your own key — Claude, GPT, Gemini, whatever you prefer. This means you control your model costs directly and we never see your prompts or responses.

  3. Send your first message. Open Telegram, find your OpenClaw bot, and type something like:

Go to news.ycombinator.com and tell me the top 5 stories right now.

That is it. Your agent spins up, opens a real browser, navigates to Hacker News, reads the page, and sends back the results. If that works, everything is wired correctly.

The Cloud plan is $25/mo during beta, $39/mo after. The pricing breakdown covers what is included.

Path B: Self-hosted with Docker (30 minutes)

If you want full control — or you just enjoy running infrastructure — self-hosting is the way. You will need a machine with Docker installed (a VPS works great, a spare laptop works too) and about thirty minutes.

Step 1: Get an LLM API key.

Sign up at Anthropic, OpenAI, or whichever provider you prefer. Copy the API key somewhere safe.

Step 2: Create a Telegram bot.

Open Telegram and message @BotFather. Send /newbot, follow the prompts to name it, and copy the bot token it gives you. This token is how OpenClaw receives your messages.

Step 3: Install Docker.

If you do not have Docker yet:

# Ubuntu/Debian
curl -fsSL https://get.docker.com | sh
sudo usermod -aG docker $USER

On macOS, grab Docker Desktop. On Windows, the Windows setup guide covers the WSL2 path.

Step 4: Run OpenClaw.

docker run -d \
  --name openclaw \
  -e ANTHROPIC_API_KEY=sk-ant-your-key-here \
  -e TELEGRAM_BOT_TOKEN=your-bot-token-here \
  -v openclaw-data:/data \
  -p 8080:8080 \
  ghcr.io/openclaw-ai/openclaw:latest

Replace the placeholder values with your real keys. The -v openclaw-data:/data part is critical — it persists your agent’s memory and files across restarts. Skip it and you lose everything when the container stops. The Docker setup guide goes deeper on volumes, networking, and docker-compose.

Step 5: Set up HTTPS with Caddy.

Telegram requires HTTPS for webhooks. Caddy makes this painless:

# Install Caddy
sudo apt install -y caddy

# Create Caddyfile
sudo tee /etc/caddy/Caddyfile > /dev/null <<EOF
yourdomain.com {
    reverse_proxy localhost:8080
}
EOF

# Restart Caddy
sudo systemctl restart caddy

Caddy automatically provisions a Let’s Encrypt certificate. Point your domain’s DNS to your server’s IP, and within a minute or two you will have a valid HTTPS endpoint.

Now message your bot on Telegram. Same test as above — ask it for the top Hacker News stories. If it responds, you are live.

Part 2: Your first five tasks

Once you have a working agent, resist the urge to build a complex automation immediately. Start with these five tasks. Each one exercises a different capability, and by the end you will have a solid intuition for what OpenClaw can do.

Task 1: Web research

Research the top 5 project management tools in 2026. Compare their
pricing tiers in a table with plan names, monthly costs per user,
and standout features.

This makes the agent browse multiple websites, extract structured data, and synthesize it into a format you can actually use. You will notice it opens tabs, navigates between pricing pages, and handles the inevitable cookie banners and popups without you lifting a finger.

Task 2: Data extraction

Go to Hacker News, extract the top 10 stories with their titles,
point counts, and comment counts. Format as a numbered list.

Simple and fast. This tests the browser’s ability to parse a page and pull specific data points. If you have used web scraping tools before, you know how much boilerplate this normally requires. Here it is one message.

Task 3: File processing

Write a Python script that generates a CSV file with 100 rows of
fake sales data (date, product, quantity, price, region). Then
analyze the data: total revenue by region, best-selling product,
and average order value. Show me the results.

This is where OpenClaw being a real computer matters. It does not simulate code execution — it actually opens a terminal, writes a Python script, runs it, reads the output, and reports back. The CSV file exists on disk afterward. You could ask the agent to email it to you or upload it somewhere.

Task 4: Multi-step workflow

Check Google Trends for today's top 5 trending searches in the US.
For each one, find a relevant news article and give me a one-sentence
summary of what is happening.

Multi-step tasks are where OpenClaw starts to feel different from a chatbot. The agent plans a sequence of actions, executes them in order, handles errors along the way (a link might be dead, a page might load slowly), and assembles the final output. If you want to understand how the agent runtime works under the hood, that post breaks it down.

Task 5: Scheduled task

Every morning at 9am Pacific, check the weather forecast for
San Francisco. If rain is expected in the next 24 hours, message
me with the details. Otherwise, do not message.

Scheduling is one of OpenClaw’s most underrated features. This is not a cron job you have to manage — the agent handles the scheduling internally and runs the task at the specified time. It persists across restarts if you have your volumes set up correctly.

Part 3: Key concepts

You can get a lot done without understanding the internals. But knowing a few key concepts will help you write better prompts and build more reliable workflows.

The computer environment

OpenClaw is not a browser extension or a chatbot with web access bolted on. It runs on a real computer with three core capabilities:

  • Browser — A full Chrome instance via CDP (Chrome DevTools Protocol). It can navigate pages, fill forms, click buttons, handle JavaScript-rendered content, and download files. The browser deep-dive covers the internals.
  • Terminal — A real shell where the agent can run commands, execute scripts in any language, install packages, and process files.
  • File system — Persistent storage where the agent reads and writes files. CSVs, PDFs, images, scripts, databases — whatever the task requires.

This trio is what separates OpenClaw from “AI browser agents” that can only interact with web pages. When someone asks what OpenClaw can actually do, the answer is: anything you could do with a computer, a browser, and enough time.

Memory

OpenClaw remembers things between conversations. Not just what you said five minutes ago, but context from weeks or months back.

The memory system stores data as local Markdown files on disk. When you tell your agent “I prefer tables over bullet lists” or “my company’s fiscal year starts in April,” it writes that to memory. On future conversations, relevant memories are pulled in using hybrid search — a combination of vector embeddings (semantic similarity) and BM25 (keyword matching).

This is local-first by design. Your memories live on your machine (or your container’s volume), not on some vendor’s server. On OpenClaw Cloud, memories stay inside your isolated container with zero-access architecture, meaning not even we can read them.

Skills

Think of skills as plugins that extend what your agent can do. Out of the box, OpenClaw can browse, code, and manage files. Skills add integrations: Google Sheets, Notion, Slack, email, calendars, image generation, and thousands more.

There are over 31,000 skills on CrawHub, the community marketplace. Each skill is a SOUL.md file — a Markdown document that describes capabilities, API endpoints, and usage patterns. The clever part is semantic discovery: you do not need to manually install or activate skills. The agent finds relevant skills automatically based on what you are asking it to do.

The skills explainer goes deeper, but the practical takeaway is: if you need your agent to interact with a specific service, there is probably already a skill for it.

Model selection

OpenClaw is model-agnostic. You are not locked into one AI provider. The models you can use include Claude, GPT, Gemini, Groq, Mistral, xAI, and even local models through Ollama.

In practice, model choice matters more than people think:

  • Claude Sonnet — Best balance of speed, cost, and capability. This is what I use for 80% of tasks.
  • Claude Opus — For complex multi-step reasoning where accuracy is critical. More expensive, noticeably smarter on hard problems.
  • GPT-4o-mini or Claude Haiku — For simple, repetitive tasks where cost matters more than depth. Checking a weather API, formatting text, quick lookups.
  • Local models via Ollama — For people who want zero data leaving their machine. Performance varies by model, but it works.

You can switch models per conversation or set a default. The right model for the job depends on the task complexity and your budget.

Channel adapters

You control OpenClaw through messaging apps. Telegram is the most popular (and currently the only option on OpenClaw Cloud), but self-hosted setups can connect through WhatsApp, Slack, Discord, Signal, iMessage, and over 20 other platforms.

The hub-and-spoke architecture means adding a new channel does not change how the agent works. The gateway translates messages from whatever platform you use into a common format the agent understands. You could talk to the same agent from Telegram on your phone and Slack on your laptop, and it would maintain context across both.

Part 4: Real workflow examples

These are workflows people actually run. I have simplified them slightly, but the core patterns are real.

Competitive price monitoring

Every day at 8am, check the pricing pages of Competitor A, Competitor B,
Competitor C, Competitor D, and Competitor E. Record the prices for their
Pro plan in a CSV file called competitor-prices.csv, appending a new row
each day with the date and all five prices. If any price changed from
yesterday, message me with what changed.

This runs unattended. The agent opens each pricing page, extracts the number, appends it to a growing dataset, and only bothers you when something changes. After a month you have a clean price history you can chart. People in the community have been using this for SaaS competitor tracking, flight prices, and real estate listings.

Client report automation

I am going to send you a CSV file. When you receive it, do the following:
1. Analyze the data — total revenue, revenue by category, month-over-month
   growth rate, and top 10 customers by spend.
2. Generate charts for the key metrics.
3. Compile everything into a clean PDF report with the title
   "Monthly Revenue Report — March 2026".
4. Send the PDF back to me in this chat.

This combines file processing, data analysis, code execution (for the charts), and document generation. The agent writes a Python script using matplotlib or similar, generates the charts, then assembles the PDF. The whole thing takes a few minutes depending on the data size.

I use a version of this for a freelance client who sends me raw data every month. What used to take me an hour of manual spreadsheet work now takes one Telegram message.

Job listing monitor

Every 6 hours, check these job boards for new senior engineering
positions in San Francisco:
- LinkedIn Jobs (search: "senior software engineer" location: SF)
- Indeed (same search)
- Hacker News "Who is Hiring" thread

Deduplicate by company name. If there are new listings since last check,
message me with company name, title, and link. Keep a running list in
jobs-tracker.csv.

The deduplication is the key detail here. Without it, you would get the same listings every six hours. The agent maintains the CSV as its source of truth and only alerts you on genuinely new postings. People have extended this with filters (salary range, remote-only, specific tech stacks) and it works surprisingly well.

Part 5: Tips for getting the most out of OpenClaw

After months of daily use, here is what I have learned.

Be specific in your instructions. “Research competitors” is vague. “Visit the pricing pages of HubSpot, Salesforce, and Pipedrive, and create a comparison table with plan names, per-seat pricing, and free tier limits” gives the agent a clear target. The more precise your desired output, the better the results. The prompt writing guide expands on this.

Break large tasks into steps. Instead of “build me a complete market analysis,” start with “research the top 10 companies in the CRM space and list them with their estimated revenue.” Then follow up with deeper dives. The agent handles multi-step tasks well, but giving it a clear sequence reduces errors on complex work.

Use memory for recurring context. If you always want reports in a specific format, or your company has naming conventions, or you prefer certain data sources — tell your agent once and it will remember. “Remember: when I ask for competitor analysis, always include G2 ratings and Capterra scores” saves you from repeating it every time.

Create skills for repeating workflows. If you find yourself giving the same instructions every week, that is a skill waiting to be written. SOUL.md files are just Markdown — you can write one in ten minutes and your agent will discover it automatically on relevant tasks.

Choose the right model for the task. Do not use Opus for checking the weather. Do not use Haiku for analyzing a complex dataset. Matching model capability to task complexity saves money and often produces better results because lighter models are faster and less prone to overthinking simple tasks.

Check the error guide when things break. OpenClaw is powerful but not infallible. Pages load slowly, CAPTCHAs appear, APIs rate-limit. Most common issues have known fixes. The security guide is also worth reading if you are self-hosting, especially the parts about not exposing internal ports.

Where to go from here

If you made it this far, you have a working agent and a solid foundation for using it. Here are the natural next steps:

And if you want the managed path without the infrastructure headaches, OpenClaw Cloud handles everything — isolated containers, pre-installed skills, automatic updates, and a Telegram interface that just works. Bring your own API key, keep your data private, and let someone else worry about Docker volumes and SSL certificates.

← Back to blog