How to Get Started with OpenClaw and AI Agents
You have probably seen the headlines. AI agents that manage your inbox, monitor your competitors, draft your content, and run while you sleep. It sounds impressive. It also sounds complicated.
It is not.
I set up my own AI agent on a Sunday afternoon in Bangkok. After lunch with my daughter, French TV soap on one screen, terminal on the other. The actual work took about fifteen minutes. A few weeks later, it handles my morning briefings, monitors my industry, drafts email replies, and manages content on this website — all through Telegram messages.
I wrote about why every leader should try this a few weeks ago. This article is the practical companion. If you are curious but have not started yet, this is for you.
Why does this matter if you run a company, lead a team, or sit on a board? Because what I set up in an afternoon is functionally what enterprise vendors package as a six-figure automation platform. The gap between what one person can do with open-source tools and what organisations pay consultants and software vendors to deliver is closing faster than most leaders realise. Understanding that gap firsthand — not from a pitch deck, but from doing it yourself — is becoming a leadership responsibility.
First — What Is an AI Agent?
If you have used ChatGPT, Claude, or Gemini, you already understand the basic idea. You type something, the AI responds. That is a conversation. An AI agent takes it a step further: instead of just answering your questions, it actually does things.
It can send an email on your behalf. Check your calendar and tell you what is coming up. Monitor a competitor’s website and alert you when something changes. Draft a blog post and publish it. Summarise a 40-page PDF. All while you are in a meeting, on a walk, or asleep.
The difference is simple: ChatGPT is someone you talk to. An AI agent is someone who works for you.
The tool I use is called OpenClaw. It is open-source, free to install, and has a massive community behind it. You do not need to be a developer. You do not need to understand code. If you can install an app and send a text message, you can run an AI agent.
Getting It Running
There are five things you need to do. None of them are difficult, but I want to explain why each one matters — not just the what.
Give It a Home
Your agent needs a computer to live on. Not your main laptop — ideally a spare machine that can stay on. An old laptop, a mini PC, even a Raspberry Pi. The reason is simple: unlike ChatGPT, which only works when you open a browser tab, your agent runs continuously. It checks things. It monitors. It acts on your behalf even when you are not looking. For that, it needs a machine that stays powered on.
I used my Lenovo all-in-one running Linux. Nothing special, nothing expensive. Some of my team members at Xponential installed it on their MacBooks. One person used a Raspberry Pi. It all works. If you do not have a spare machine, you can also rent a small cloud server for a few dollars a month — but for getting started, any computer you have sitting around will do.
Connect It to Your Phone
This is the part that makes it click for most people. OpenClaw connects to messaging apps — Telegram, WhatsApp, Slack, Discord, iMessage. Once connected, your agent lives inside an app you already have on your phone. You message it the way you would message a friend. It responds the same way.
No special app to download. No website to visit. No dashboard to learn. Just open Telegram (or whichever app you choose), and there it is — your personal AI assistant, available 24/7.
I chose Telegram because I already use it every day. The key is to pick whatever app you open most often. I have seen people set up a “clean” dedicated channel for their agent and then never check it. The ones who connected to their daily-use app stuck with it.
Give It a Brain
OpenClaw is the framework — the body, if you like. But it needs an AI model to think with. You get to choose which one. This is actually a good thing, because it means you are not locked into one company and you control how much you spend.
If you have ever used ChatGPT, you have used an AI model (GPT). Claude, Gemini, DeepSeek — these are all different AI models from different companies. Each has strengths and a different price point:
| Model | What It Is Good At | Cost |
|---|---|---|
| Claude Sonnet (Anthropic) | Complex thinking, long tasks, using tools | Higher |
| Gemini Flash (Google) | Fast answers, web search, research | Low |
| Kimi K2.5 (Moonshot AI) | Strong reasoning, open-source, very long memory | Low |
| DeepSeek V3 | Surprisingly capable at everything, very cheap | Low |
| Nemotron (NVIDIA / Ollama) | Runs on your own machine, completely free | Free |
My recommendation for beginners: just pick one and start. Gemini Flash or Claude Sonnet are the easiest first choices. I started with Gemini for search tasks and Kimi K2.5 for reasoning — both on their free tiers — because I wanted to explore without spending anything.
One important thing to understand: because your agent runs all the time, it uses your AI model even when you are not asking it anything — checking schedules, running background tasks, what the community calls “heartbeat” requests. On a premium model like Claude Sonnet, my agent was costing about $25/day before I optimised. That is real money.
The fix is straightforward: set a daily spending limit on your AI provider’s account before you start. Every major provider has this setting. Use it. You can always raise it later once you understand your usage.
Once you have been running for a few weeks, you can get more sophisticated. OpenClaw lets you chain models — use a cheap fast model (like Gemini Flash) for routine tasks, and automatically escalate to a more powerful model (like Claude) only when the task is complex. You configure it once, and it routes automatically. In practice, this cuts costs by 70–80% with no noticeable quality drop for everyday use.
Teach It What You Need
Out of the box, your agent can have a conversation. That is useful, but it is not much different from ChatGPT. What makes an agent powerful is skills — small extensions you add that teach it how to do specific things. Think of them as apps for your agent.
Want it to manage your email? There is a skill for that. Want it to publish blog posts? Skill. Monitor competitors? Skill. The official marketplace — ClawHub — has over 13,000 of them.
That number is overwhelming. You do not need most of them. Here is what I actually installed and use — along with a few I have not got around to yet. I am being transparent about both, because most guides pretend everything is perfectly set up from day one. It is not. That is normal.
What I Use Every Day
Google Workspace (GOG). Connects your Gmail, Calendar, and Drive. The morning briefing alone makes it worth it — before I pick up my phone, my agent has already pushed me a summary of today’s meetings, emails that need attention, and flagged tasks. That used to take me fifteen minutes of scrolling through notifications. Now it takes fifteen seconds of reading one Telegram message.
Self-Improving Agent. Lets your agent learn from its own mistakes. When something does not work, it logs what happened and adjusts. Over time, it gets noticeably better at understanding what you want. Install this early — the longer it runs, the smarter it gets.
Proactivity. Without this, your agent only responds when you ask. With it, the agent starts doing things on its own — checking your schedule, flagging things that need attention, surfacing information before you ask. This is what turns it from a chatbot into an actual assistant.
Competitor Analysis. Monitors competitor activity and surfaces what matters. In my role overseeing digital strategy for a major retail group, staying current is essential — but doom-scrolling is not a strategy. The agent gives me a curated brief, not a firehose.
Content Generation and Research. For anyone publishing content. Agents research the market and key topics, and delivery content for this this web site, axelwinter.com, and now I manage content through Telegram messages.
What I Have Not Activated Yet (But Should)
Being transparent: OpenClaw comes with several useful built-in skills I have not turned on yet. Most guides pretend everything is perfectly configured from day one. It is not.
Model Usage Tracking (should have been the first thing I enabled) — tracks which models you are using and what they cost. Healthcheck — runs security audits on your setup. Summarize — auto-summarises long documents and threads. Coding Agent — a full Codex-style coding assistant. BlogWatcher — monitors competitor blogs for new content. Whisper (Voice Transcription) — send voice notes, get them transcribed and acted on. All built in. All just need switching on.
The point: you do not have to set up everything on day one. Start with two or three skills. Live with them. Add more as you discover what you actually need.
What Does Not Exist Yet
LinkedIn. No reliable skill exists yet. Given that LinkedIn is where most professional engagement happens, this is the biggest gap. For now I post manually. Consolidated cost tracking across multiple AI providers is also still manual work.
A Word on Security
Not all skills are safe. In early 2026, a campaign called ClawHavoc planted malicious skills on ClawHub using names similar to popular ones — typosquatting, essentially. Before installing any skill: check the download count, check the author, and read the security scan. ClawHub flags suspicious skills automatically — do not override those warnings. Start with well-known, top-ranked skills.
Just Start Talking
Here is the part most guides miss: you do not need to read the documentation. Just ask your agent. It will walk you through everything conversationally.
On my first day, I asked the agent to learn the computer it was running on, complete its own setup, and secure itself. It did. I asked it to connect to my WordPress site. It did. I asked it to set up a daily industry brief. It did. The experience felt less like configuring software and more like briefing a new team member.
What Using It Actually Feels Like
The setup is interesting, but this is the part that matters: what happens after? What does it actually feel like to have an AI agent running?
It feels like texting an extremely competent assistant who never sleeps.
Here are real things I do with my agent through Telegram — the same way I would text a colleague:
“Check my calendar for Thursday and tell me if I have a gap between 2 and 4.”
Done. It checks Google Calendar and responds in seconds.
“Schedule a meeting with [name] next Tuesday at 10am, send the invite from my email.”
Done. Calendar event created, invitation sent. I did not open Gmail or Calendar once.
“Summarise the emails I got overnight and flag anything urgent.”
Done. A clean summary lands in my Telegram chat before my first coffee.
“Draft a reply to the email from [name] about the project timeline — keep it professional, say we can deliver by end of Q2.”
Done. Draft sits in my inbox, I review it in ten seconds, hit send.
“What are our competitors doing with loyalty programmes this week?”
Done. A curated brief based on what the social listening and competitor analysis skills have picked up.
“Publish today’s article draft on axelwinter.com.”
Done. WordPress post created, published, no dashboard required.
None of these individually are revolutionary. Any single one takes a few minutes to do manually. But multiply that by twenty or thirty interactions a day, and it adds up to hours saved every week. And because it all happens in the same Telegram window I already have open, there is zero friction. No app-switching. No logging in. No remembering which dashboard does what.
The morning briefing is probably the single most valuable thing. I wake up, and before I have touched my phone, my agent has already told me: here is your day, here are the emails that need you, here is what changed in your industry overnight. That fifteen minutes of morning scrolling and context-switching — gone.
What We Built Together — In Real Time
Here is something worth sharing: this article itself is a live example.
While writing this, my agent — Andy — scheduled 6 tweets, posted this article as a draft to WordPress, ran a GEO intelligence scan across Bangkok’s top malls (25 Gemini queries, in English and Thai), emailed a branded HTML report to a client, set up a weekly automated scan as a recurring job, updated OpenClaw to the latest version, ran a security audit, and sent me a morning briefing. All via Telegram messages. No dashboards. No terminals. No copy-paste.
The agent did not just help me write about this. It was doing the work while I wrote.
That is the point. You stop managing tools and start having conversations with an assistant that gets things done.
My Setup Today
For the curious, here is what my agent looks like a few weeks in:
- Hardware: Lenovo all-in-one desktop (3 years old), running OpenClaw as a background service.
- Messaging: Telegram — my interface for everything
- AI Models: Currently Claude Sonnet as primary (powerful but expensive at ~$25/day unoptimised). Moving to: Gemini Flash as primary → Claude as fallback → DeepSeek V3 → Nemotron (free/local). OpenClaw routes automatically once configured — expecting 70–80% cost reduction.
- Active Skills: Google Workspace, Self-Improving Agent, Proactivity, Competitor Analysis, SEO Content Writer, AEO Analytics, ClawHub browser
- Not Yet Activated: Coding Agent, Voice Transcription – no need as of now.
- Missing: LinkedIn (no skill exists), consolidated multi-provider cost view
It is not a perfect setup. It is a working one. And that is the difference that matters.
That is it. An afternoon. Five decisions. And you have a personal AI assistant running 24/7 on your own hardware, connected to the apps you already use, doing real work for you in the background.
Every organisation I work with is evaluating automation platforms, AI copilots, and digital transformation roadmaps that run into millions in licensing and implementation. Most of those decisions are being made by leaders who have never actually run an AI agent themselves. They are buying based on vendor demos, not lived experience.
That is like approving a factory automation investment without ever having walked the factory floor.
Before your next board discussion on AI strategy, before you sign off on the next enterprise software contract, spend a Sunday afternoon doing what I did. Not because the open-source tool will replace your enterprise stack. But because the experience will fundamentally change the questions you ask, the budgets you approve, and the timelines you accept.
Start simple. Live with it for a few weeks. Then bring that understanding into the room where the decisions get made.
If you want the backstory — and the deeper strategic argument — read Why I Spent My Sunday Installing an AI Agent. Whitepapers on this topic are coming.
If you want the backstory on why I did this — and why I think it matters strategically — read Why I Spent My Sunday Installing an AI Agent.
Questions? Want to compare setups? Find me on X (@AxelWinterBkk) or LinkedIn.
2 Responses
[…] That’s what I mean by hands-on. A CEO building a working prototype, not requesting a feasibility study. If you’re curious how to get started yourself, I wrote a practical guide on how to set up your first AI agent in an afternoon. […]
[…] There is no single “AI” in my workflow. There are multiple models and one memory system, each with a specific role, and the whole thing is orchestrated by OpenClaw. […]