Transcription
I didn't ask for any of this. My AI agent researched the news, checked my calendar, pulled my tasks, and sent me this briefing at 8:00 a.m. on its own. And it does this every single morning without me touching a thing. OpenClaw is the closest thing to having a personal AI employee. And after spending the last 3 weeks setting it up the right way, I'm going to show you exactly how to get yours running from scratch.
OpenClaw is an open source AI agent that runs on your own server and connects to the AI models you already know, like Claude and GPT. The difference between this and a regular chatbot is that OpenClaw can actually take action. It reads your emails, manages your calendar, runs commands, installs software, and does real tasks on your behalf through whatever messaging app you already use, whether that's Telegram, WhatsApp, Discord, or something else. And because it runs on your own server instead of someone else's cloud, everything stays private and under your control.
So, before we get into the features that make this thing worth using, you need to get it set up first. The safest way to run OpenClaw is on a VPS, which is basically a virtual computer that runs 24/7 in the cloud. The reason you want this instead of running it on your personal machine is because OpenClaw has full access to the system it's installed on. And that means it can create files, delete files, install software, and run terminal commands. In fact, some users who ran OpenClaw on their main machine have reported their agent wiping out system files, or exposing their API keys to the open internet, all because of a single bad prompt. A VPS keeps all of that risk in an isolated environment which you can delete and rebuild in two minutes.
The platform I use for this is Hostinger. And the reason I chose Hostinger specifically is because they have a one-click openclaw template that skips all the complex terminal setup. I've left a link in the description that takes you straight to the OpenClaw deploy page. Once you open the link, click launch openclaw and you'll be taken to the billing page. Here I selected the 12-month plan because the longer term significantly lowers the monthly cost. Then I turned on daily auto backups, which is useful since OpenClaw can modify its own server environment and backups give you a clean restore point if anything breaks. After that, type in the code parker to get 10% off the plan. Just note that the discount doesn't apply to the daily backups. Once everything is set, click continue and complete the payment.
Once that's done, you'll land on the OpenClaw configuration page. The first field is the gateway token, which is the password to your entire OpenClaw dashboard. Copy the token and save it somewhere safe, like a password manager, because you'll need it every time you log in. Next, you need an API key for your AI model. I'm using Enthropics Claude, so I'll head over to platform.anthropic.com, sign in, and create a new API key. You'll want to load at least $5 in credits so the model has funds to run on. Copy the key, go back to Hostinger, and paste it into the Anthropic field. If you prefer OpenAI or another provider, just follow the same steps with their platform instead. Scroll down and click deploy. Hostinger will start building your server, which takes about 2 to 5 minutes. When it's done, you'll see a green check mark and a purple link right next to it. Click that link, paste in your gateway token, and you're logged into the OpenClaw dashboard.
The first thing to do is configure your agents identity. Just say hello in the chat, and OpenClaw will ask you a series of questions about who it is, who you are, and what kind of assistant you want. Answer the prompts, and your agent's personality and role are locked in. After that, I'd set up a messaging channel so you can talk to your agent from your phone instead of being stuck in the browser. The easiest one is Telegram. Just tell your agent, "I want to set up Telegram." And it'll give you step-by-step instructions to do that.
Now that your agent is running and you can reach it from your phone, it's time to turn it into something that actually works for you. By default, OpenClaw processes everything in a single straight line. If you ask it to research five competitors and then write a summary of each one, it has to finish the first task before starting the second, which means you are stuck waiting while it works through the list. Sub agents change that completely. The idea is simple. Instead of one agent doing everything, you create specialized workers that your main agent can delegate to. One handles coding, another handles research, a third handles content writing, and your main agent acts as the manager who decides who does what. And the reason this matters beyond just speed is that each sub agent gets its own context window and its own memory. That means your research agent isn't cluttered with coding conversations, and your coding agent isn't confused by research notes. Everything stays clean and separated.
To create one, all you have to do is send your main agent a message. I'll paste in a prompt that tells it to create a new persistent agent named Atlas. Set it up as a dedicated research assistant and assign it a model. And if I go to the agents tab in the OpenClaw dashboard and refresh the page, Atlas is now listed right there as a separate agent with its own workspace. Now I'll create a second one. This time I'll paste a prompt for a coding assistant named Iris running on a cheaper model since coding tasks don't always need the most expensive option. And just like that, I have two specialized workers that my main agent can use whenever it needs to. So now when I ask my main agent to research competitors in a space and build a comparison spreadsheet, it doesn't grind through everything on its own. It passes the research task to Atlas and the spreadsheet task to Iris and both of them work at the same time. The results come back faster and the quality is better because each agent is focused on exactly one type of work. The best part is that you can assign cheaper models to sub agents that don't need heavy reasoning. Your main agent stays on Claude opus for complex thinking and your sub agents run on sonnet for routine tasks which cuts your API cost significantly without losing quality where it matters.
Now having a team of agents is already a huge upgrade but they're still only working when you tell them to. The next feature changes that. This is the feature that separates OpenClaw from every standard AI interface because you can set it up to do things on its own without you ever sending a message. Every afternoon at 3:45 p.m. my agent sends me a meeting dossier on Telegram. It scans my Google calendar for the day, identifies exactly who I am meeting with, researches their recent company news or social posts, and sends me a cheat sheet so I am perfectly prepared for every call. I set this up once, and it has been running every single day since without me asking. To build this, you send your agent a single prompt that describes exactly what you want. I'll paste mine in right now, and the next afternoon at 3:45 p.m., your phone buzzes with a full briefing. You can take this even further. I have a second automation that runs at 9:00 a.m. where my agent drafts content for me. I will paste that prompt, too. What this does is it looks at all the random voice notes and ideas I dropped into the chat yesterday, connects the dots between them, and turns them into three polished social media posts ready for me to publish. If you head over to the cron jobs tab in your dashboard, you'll see both of these scheduled tasks listed there. And you can edit the timing or disable them whenever you want. This kind of background work puts OpenClaw in a league of its own because no other AI tool on the market operates at this level of independence when you are not even talking to it.
And once your agent is working on its own, the natural next step is to give it more abilities to work with. Skills are basically plugins that teach your agent how to do new things. When you first set up OpenClaw, it can chat and run basic commands, but it doesn't know how to search the web, interact with websites, manage your smart home, or pull data from sources. Skills add those abilities. Claw Hub is the marketplace where all of these live, and right now it has almost 19,000 community-built skills that you can install in seconds. The easiest way to get started is to install the Claw Hub skill itself, which gives your agent the ability to browse, search, and install other skills on its own. I will tell my agent to install Claw Hub, and it will handle the whole setup. Once that is done, I can ask my agent to search for skills by describing what I need. I will type find me a skill that searches Reddit. Using that new capability, it scans the marketplace and finds the best options available. But if you already know exactly which skill you want, you can just grab it directly. For this example, I really like a skill called last 30 days. I will copy the link to that skill from GitHub, paste it to my agent, and tell it to install. Within about 10 seconds, my agent has a brand new ability it did not have before. Now, I will test it. I will type search Reddit and YouTube for the biggest AI news from the past week and give me a summary. And there's the result. It pulled live data from multiple platforms and gave me an organized summary without me having to visit a single website. One thing to keep in mind with Claw Hub is that it's a community marketplace, which means anyone can publish a skill. The platform does have anti virus scanning and a report system, but you should still check the download count and star rating before installing anything unfamiliar. If a skill has thousands of downloads and good ratings, it's generally safe. If it has 12 downloads and no ratings, you should probably skip it.
Now, with sub agents handling your work, automations running in the background, and skills expanding what your agent can do, there's one problem left to solve, and it's the one that trips up almost everyone who uses OpenClaw for more than a few days. The biggest frustration with OpenClaw is that your agent starts forgetting things over time and you'll start to notice it after about a week. Details you mentioned earlier, preferences you set, decisions you made together, they start disappearing from the conversation. And this happens because of how OpenClaw manages its context window. As you keep chatting, the context fills up. And to prevent it from overflowing and costing you a fortune in tokens, OpenClaw runs something called compaction. Compaction summarizes older conversations into a short compressed version and removes the originals. The problem is that summary often loses the details that actually mattered. The fix is a single prompt that tells your agent to save the most important information from every conversation to a memory file before compaction runs. I'll paste it in. Now, what this does is it creates a system where your agent writes key decisions, preferences, project details, and anything it flags as important to a persistent file that survives compaction. So, even after the older conversation gets compressed, the information you care about is still there because it was saved separately.
And if you want to take this a step further, there's a self-improving agent system that builds on top of this memory fix. You can either install it as a skill from Claw Hub or you can set it up with a single prompt. I'll paste the prompt version so you don't have to install anything from a third party. What this creates is a feedback loop. Every time you correct your agent or tell it you want something done differently, it logs that correction in a file and adjusts its behavior going forward. Over time, it builds a collection of patterns that are specific to you. Things like how you prefer your emails written, what tone you like, which formatting you want in reports, and what kinds of suggestions you actually follow through on. The result is an agent that gets noticeably better the more you use it. After about two weeks of corrections and feedback, my agent stopped making the same mistakes and started anticipating what I wanted before I asked for it. And that is the loop that ties everything together. By layering workers, schedules, skills, and memory, your agent becomes a self-improving system that gets smarter with every interaction. I have left a link to Hostinger in the description if you want to try these superpowers yourself. Thanks for watching, and I'll see you in the next one.