The Only OpenClaw Tutorial You Will Ever Need

This might be the only open claw tutorial you'll ever need. By the end of this video, you'll have your own personal AI agent, one that remembers who you are, manages your emails, books things on your behalf, and keeps working while you're asleep. It works 24/7, you can message it from your phone, and you won't write a single line of code to get there. Open claw is free and open source. The easiest way to think about it is this. Imagine chat GPT, except this AI tool actually remembers you over time, and it can go off and take actions on its own without you asking every single time. If you've tried setting something like this up before and hit a wall, I get it. The documentation is rough, and most tutorials skip the parts that actually matter. So, stay with me. I'll walk you through the whole thing in real time. We'll plan what yours should do, pick the easiest place to run it, get it live on Telegram, and by the end, you'll have one automation already saving you time every single day. Before we set anything up, let me show you what separates open claw from a regular chatbot, because this isn't chat GPT. There are three things open claw does that a regular chatbot can't. And once you see them, the reason you'd want to use open claw becomes obvious. First, it actually remembers you. Now, I know chat GPT has memory, too, but there's a meaningful difference. Chat GPT keeps a rough summary of your past conversations, fuzzy, compressed, easy to lose context. Open claw keeps structured memory files. Your name, your job, your family, your preferences, how you like to be spoken to, what you care about. These details are kept in actual files that you can read and edit. So, when you talk to it 3 months from now, it picks up exactly where you left off, not because a summary happened to remember, but because it kept a living record of who you are and what you've worked on. Second, it runs on its own. You don't have to open an app, you don't have to prompt it, it's just always there, 24/7 in in background. It can wake itself up on a schedule, check in every few minutes, or respond to something happening in one of your connected apps, which means it can reach out to you first. [music] And the first time that happens, it stops feeling like a chatbot and starts feeling like a real assistant. Third, it takes actual action. Real action. It sends emails, schedules calendar events, pulls documents from your drive, summarizes them, messages you back in Telegram. It searches the web, compares options, and books things. It's not just answering questions, it's doing real work. Put those three pillars together and you get things like this. Three examples, and none of these are hypothetical. These are the kinds of things people are actually using it for. Example one. You wake up. There's a Telegram message waiting. Here's the weather, your first three meetings of the day, the two emails that look important, and the one-line summary of each. Example two. You tell it, "I have an interview with someone at X company on Thursday. Put together a prep doc." An hour later, there's a Google Doc in your drive with the company overview, the interviewer's background, recent news, and likely questions. Example three. You mention once that you want to fly back to see your mom in March. It's been watching the price every day. The price drops, you get a Telegram message. It says, "The flight is at the lowest it's been in a month. Should I go ahead and book it for you?" Before we set anything up, open up a notepad. Could be notes, stickies, a physical journal, whatever. We're going to take a 60-second planning step, and the answers here drive every decision we make later. All right, three questions. Your answers tell us which integrations to focus on and where your privacy line is. Write your answers down in the notepad as we go. Question one. What do you want it to do every day? Pick two or three things. Morning brief, flight watching, [snorts] auto sort your inbox? Question two. What data does that touch? Gmail, calendar, drive, docs, your bank app. If there's anything that makes you uncomfortable, make a note of it now. Question three. What's your privacy floor? In other words, what will you not connect no matter how compelling the feature is? Decide here, not mid-setup. All right, next up, where to run this thing? This is where most people get stuck. Once you know what OpenClaw is and what you want yours to do, you still have to decide where it runs. And the right answer depends on who you are. There are three options. I'll walk you through them, and then we'll pick yours. Path one is your personal computer. It's free because you already own the hardware. The catch is that your agent only runs when your computer is running. Close the laptop and your agent goes dark. You also have to consider that it has access to everything on your machine, which becomes a much bigger deal once you start connecting it to real services. Path two is dedicated hardware. Common options are a Mac mini, a Raspberry Pi, or an old laptop running headless in a closet. You have the one-time cost of the hardware, there's no monthly fees. It's sandboxed, which means it's not sitting on your personal machine anymore, but there are trade-offs. If your home internet goes down or there's a power outage, you're offline. You're also the one maintaining it, and unless you have hardware lying around already, it's several hundred dollars up front. Path three is a cloud VPS. A few dollars a month for a small server in a real data center. It's always on, always reachable, isolated from your personal machine, and if anything ever goes sideways, you wipe it and redeploy clean. Hostinger has a one-click OpenClaw template, so you can quickly install it on a VPS. The server boots with OpenClaw already installed, and the rest of the setup happens in a browser. For most people just getting started, this is the right choice. Now, path three is the right answer for almost everyone watching. That's what I'm using, and that's what we'll install together. No terminals required at any point. The sign-up, install, configuration, all of it happens in a browser or on your phone. Now, if you'd rather run it on your own hardware, like a pie, a Mac mini, or an old laptop, that's a legitimate long-term home, and I'll cover the migration from VPS to dedicated hardware in the last chapter. Even then, I'd recommend starting on the VPS today to get the hang of it in an isolated environment with low startup costs, and then you can move once you've got the system figured out. All right, let's get signed up. Use the link on screen or the first link in the video description below. That'll take you to this page here or something very similar. And the link automatically adds an extra 10% discount on top of whatever Hostinger's current promo is. This is the cheapest way to get started. And if you want to start for even cheaper, you'll notice there's a drop-down at the upper right to select your VPS plan. Go ahead and change that from KVM 2 to KVM 1. That'll give you 4 gigs of RAM, 50 gigs of disk space, 4 terabytes of bandwidth, and this is plenty for the specs to get started. Most of the heavy lifting done for Open Claw is done by the third-party LLMs, which run externally. Once you've changed your plan, go ahead and click deploy. The next page you'll land on is the cart, and I'll quickly show you the most important things to check. The first thing is the period for your registration. You can pay for 1 month, 12 months, or 24 months upfront. Now, you'll get the best price per month if you choose 24 months. You can choose 1 month to get started, but unfortunately, they don't let you use our coupon for that, so the cheapest plan you can use to get our 10% off coupon is 12 months, and that gets you in for under $100 for your upfront cost. Now, you can get this even cheaper if you turn off the ready-to-use AI. I'll show you how to connect your own LLM later. The price now is even lower without that ready-to-use AI. So, you can get started for pretty cheap. If you scroll down, you'll see that there's a couple of other things you can add if you want to. For example, you can add instant web scraping, which is super handy. This is free, so I recommend just going ahead and adding this. You can also add dedicated email for your AI and daily auto backups. I do personally recommend using the daily auto backups because if something ever goes wrong with your configuration, you can easily roll back to a previous version. So, I'll go ahead and turn this on. But, if you want to get started for cheaper, you can leave this off. Finally, you want to choose a server location. Generally speaking, you just want to choose a server location that has the lowest latency or is closest to you. So, here I'm getting, let's see, 57 milliseconds of latency with Germany, which appears to be the lowest one. So, I'll use that server. Once you've got everything configured the way you want it, go ahead and scroll back up and then click continue. On the next page, you'll register an account. You can use Google, GitHub, or use an email address. I'll use an email On the next page, you'll see your order summary, which you can review. And you'll notice, too, that there's a 30-day money-back guarantee. So, if in the next month you're not happy with anything you've set up, you can always just ask Hostinger for a refund. Go ahead and enter your billing address and your payment information to complete the checkout. After that, you'll get this confirmation page, and then you'll be routed right away into the Open Claw configuration wizard. I want to point out right now that this might look different by the time you're following along with this video. Hostinger and Open Claw get updated regularly. So, if this doesn't look the exact same, don't panic. Just follow along and try to apply the general concepts I'm showing you. That's the most important thing. Okay, so Open Claw is open source. It doesn't ship with its own AI model baked in. You bring your own. So, we need to grab an API key from an AI provider. That's what actually powers the thinking part of your agent, and you pay the provider directly. Not Open Claw, not Hostinger, just the AI provider on the other end. You can choose between any of these AI providers to start. Anthropic, OpenAI, Gemini, X, Nexos. You can start with just one, or you can connect multiple. Any of them work just fine. For me, I like Anthropic, so I'll start with this one. But, if you want to start with a different AI provider, that's just fine. The process for getting your API key is similar for any of them. What you're going to do is copy the link below the field, open it in a new tab, log into or create an account. If signing up, follow the steps to register. Then you'll buy some usage credits to get started with your account. You can start for a lower amount like $20, but I recommend with Anthropic specifically to start with at least 40 or 50 dollars. Adding $40 to your account will get you to their tier two on their rate limit chart, which means you're less likely to get throttled and have your bot stop responding to you because you hit the rate limits. Pick however much you want to add, enter your billing information, and check out. Then get an API key. You'll click create key, name your key, I'll call mine OpenClaw. And then copy your API key. Now, this key is like a password. If anybody gets their hands on it, they can use your API account. So, keep it safe and don't share it with anyone. Copy it and then return to Hostinger and paste in your key in the relevant field. Repeat this for as many AI providers as you want and then scroll down to the bottom and find the Telegram bot token field. Now, we'll create our Telegram bot. This is the thing that you'll use to actually talk to OpenClaw using your phone. If you don't have Telegram yet, go ahead and install it for free on your phone or computer using the link in the description below. Once you've got it open, search for the BotFather. You can do that by typing BotFather in the search, or I've got a direct link to open a chat with the BotFather in the description below. BotFather is the official Telegram bot for making bots. Yeah, I know how that sounds, but that's the legit flow. Go ahead and click start to start a conversation. Then, we're going to either click or type the command new bot. I'll just click it here. Next, we need to give the bot a name. I'll call mine Flint. Next, you need to give your bot a username, and it has to end with the word bot. This also has to be unique in the Telegram ecosystem, so using a longer name is helpful if you're having a hard time finding something unique. For me, I'll do Flint the Wonder Bot. And just like that, Bot Father returns a token. That's this long string here in the middle. Again, this is like a password. Don't share it with anyone. Go ahead and copy this token. Return to Hostinger and paste that token in the Telegram bot token field. Once you've done all this, go ahead and click deploy. The next page you land on might have a survey. If it does, you can fill it out or simply scroll to the bottom and click skip. The next page we'll land on is the Docker manager inside of Hostinger's hPanel. And you can see that Open Claw is already running and Hostinger has installed a reverse proxy for us to help keep our whole setup safer. Now, if you don't know what Docker is right now, that's totally fine. Docker is essentially just a self-contained file where your application runs. Kind of like running an application on your home computer. To open your Open Claw dashboard, click to copy the gateway token, and then under the access column, click open. Paste in your gateway token here, then click login. Welcome to your Open Claw dashboard. In the center of the screen, you can see a chat window where you can start chatting with your assistant. Let's go ahead and say hello. When you send that first message, the agent kicks off a short onboarding wizard called bootstrap. It asks you a few questions, and in the background, it writes files like your soul.md and user.md files. Just answer the questions the AI gives you. There's no special preparation needed here. So, for me, it asks, "Who am I and who are you? Let's tell it. Your name is Flint. My name is Matt. Next, it'll ask for some more information about what the vibe is, what kind of bot it is. It asks for a signature emoji. So, here you can start to give your bot a persona. If you don't have a specific use case in mind yet, you can just copy what I'm doing here. I'm making a travel-focused assistant as a starting point. You can always rewrite these details later once you actually know what you want. Nothing you answer here is permanent. So, I'll tell it you're a helpful, casual, travel planning-focused assistant. You can pick your own emoji. It'll go ahead and write its files. It comes back with a message asking for more information. In this case, it says number one, what's your time zone? And two, how do you want to reach me? So, we'll say one, what the time zone is. And two, we need to get very specific about how it should set up Telegram. So, I've said, "Let's set up Telegram using the Open Claw pairing CLI commands. Read the pairing docs at docs.openclaw.ai/channels/pairing first. You have the Telegram bot token in your environment variables." Then, we'll hit send. Great. Now, it comes back saying we just need to message the bot on Telegram and it'll reply with a pairing code. Then, we'll paste the code in here. So, let's go ahead and pop over to Telegram again. And then, click the link that the Bot Father provided to start a conversation with our new bot. We'll click start. And then, it gives us this pairing code. We'll go ahead and simply copy this and paste it into our chat in the Open Claw dashboard. It comes back and says we can try messaging the bot on Telegram. So, let's go do that. We'll say, "Hello." And we have a message saying Telegram's working. We're properly connected. So, now you can talk to your Open Claw agent on any device that has Telegram installed and logged in. Now, let's pop back over to our Open Cloud dashboard for a second and I'll show you around this interface a bit more. Here at the top left of the chat is your session drop down. Think of each session as a separate conversation. As you have different sessions, different conversations, they'll appear here. Here in the middle, you have the model drop down. If you click this, you can change which AI model is answering your messages. And it'll pull the list from the provider you connected earlier. Leave the default for now. We'll cover model selection later. And then third, you have the thinking level drop down. Higher means it's slower with better reasoning, but it costs more tokens, meaning more money. Lower is faster and costs less. But, may not be as thorough with its answers. Let's stick with the default adaptive for now. If you ever want to attach something to a message, you can click the paperclip. You can click the microphone to provide voice input. And if you ever want to start a new session in your dashboard, you can click this plus button here. Now, the first time I saw this, I thought this was to add new files, so don't get that confused. The paperclip is the to attach files, and the plus button is to start a new conversation. All right, your agent is deployed, has a personality, knows your name, and can message you on Telegram. Before we teach it to send emails, manage your calendar, and take real actions, let's take 30 seconds to discuss some safety rules. A quick word on security before we go further. The one-click template ships with some sensible defaults already on, including sandboxed execution, so you don't have to configure any of that yourself. But, there are a few habits on your end that matter more than any default. So, let me just name them. First, start small. For your first week, keep your agent on Telegram only. Don't connect your banking app, don't hook it up to a password manager, don't give it your trading account. Let it earn trust first with low-stakes things. Second, never paste API keys or credentials directly into the Telegram chat. Think of API keys like passwords. They prove to a service that it's you and they don't belong in a chat log. They go in the hosting or panel as environment variables and I'll walk you through that before we build the first automation. Third, when the agent asks to confirm an action, read the confirmation. Don't just reflexively say yes. The approval gate is only useful if you actually look at what you're approving. All right, that's it. Now, let's make the agent actually useful. We're going to give it some skills. Out of the box, your agent can hold a conversation and browse the web. That's mostly it. To make it actually do things like send emails, read your calendar, or publish code to GitHub, you install skills. A skill is essentially an app for your agent. You pick the ones you need and the agent uses them. Skills live in a marketplace called ClawHub and I'll show you how to access it. In the Open Claw dashboard, on the left side, click skills. Here you can see your pre-installed skills. Under the all tab, you can see all the skills that are installed right now. Ready shows you the ones that are installed and ready to go and need setup shows you the ones that are installed but still need some configuration from you. To add new skills to your Open Claw, first you need to find a skill. Open a new tab and go to clawhub.ai. Here you can search for skills if you're looking for something specific or you can scroll through and see the featured and trending skills. If you find a skill you like, you can click on it to see more details. Here you'll be able to see details like what you would run at the command line if you're running this on a local machine or if you want to just prompt it to install via the chat, you can copy the prompt here and just simply paste it into your chat. You can also see the security scan. Now, this is important. This VirusTotal scan shows this skill as being benign, but other skills are potentially malicious. You see, because ClawHub is a community marketplace. Most skills are well-built and just fine, but because it's a marketplace, anyone can submit skills. Before installing, check the author, the install count, the rating, and if something looks off, it's better to pass on it and be safe. Be selective up front, so you don't have to clean up any messes later. In this video, we're going to start by installing Google Workspace. Once it's connected, your agent can read your email, manage your calendar, pull documents out of your drive, all really high-value stuff. This is the chapter that makes the whole thing feel real. GOG, the Google One Gateway, is a single skill that fires up Gmail, calendar, Drive, Docs, Sheets, and Contacts in one OAuth flow. Fair warning, this chapter runs a little longer than the others, but you only do it once, ever, and the payoff is huge. Let's get into it. In ClawHub, search for GOG. You could also search for Google. And the one we're looking for is the one from Peter Steinberger. That's this one here, @stypeet. Click that one. And then, simply copy the prompt for the install and setup. Let's go to our Telegram, and then paste in that prompt. It came back saying the skill was installed. Now, we're going to set it up. Google reasonably wants to verify who's asking before handing over access to your inbox. The way you prove it is by creating a Google Cloud project, generating what are called OAuth credentials, and pasting them to your agent. A few minutes, one time, and never again. I'll walk you through every click. OAuth is the mechanism that lets one app ask another app for permission to access your data without ever seeing your password. You approve the access on Google's side, Google hands back a credential, and your agent uses that credential. It's the same pattern as sign in with Google on any other their First, open a new tab and go to console.cloud.google.com. First, you'll need to sign in to your Google account. And next, you'll need to create a project. Uh in the upper left, click select a project. Then, click new project. Give your project a name. I'll call mine OpenClaw. And click create. When the project is created, go ahead and click select project. And then on the left side, hover over APIs and services. Then, click enabled Now, we need to go through and enable the APIs for each individual Google service that we want to use. So, let's start with Gmail. We'll click enable APIs and services at the top. Then, search for Gmail. Select it from the list. Then, click enable. Then here in the Gmail API page, we can see that the status is set to enabled. Now, we need to add the next Google service. Click the back arrow in the upper left. And then repeat the process. Click enable APIs and services and search for your next skill. Go through and do this for all of the Google Workspace skills you want to use. Gmail, Calendar, Drive, Sheets, Docs, People. And I'll be honest, this OAuth process is probably the most annoying part of the entire setup. You're creating a project in Google Cloud, enabling APIs one by one, setting up OAuth consent screens. And the whole process takes about 10 to 15 minutes, and it can feel really tedious. But, once it's done, it's done, and the payoff is huge. Once you're done adding all your individual APIs, next click OAuth consent screen on the left side. Then, click get started. Give your app a name. Again, we'll call it OpenClaw. Select your email from the user support email drop down, and click next. For audience, select external. Click next. For contact information, just add your own email. Click next. And then agree to the terms of service. Click create when you're done. Great. Now that our OAuth configuration is set up, click audience on the left side. And here we'll add a test user. Scroll down the page and then click add users under the test user section. Then enter your email address and click save. Now we need to set up our OAuth credentials. Click the hamburger menu in the upper left, then under APIs and services, click credentials. In the top, click create credentials and then select OAuth client ID. In the application type drop down, select desktop app and give it a name. I'll call mine OpenClaw and click create. And here we've got a confirmation saying that OAuth client was created. Now, the thing you'll want to do here is download the JSON file at the bottom of this pop-up. Then, return to your Telegram and send the JSON file to your bot. It may ask what your Gmail address is, so go ahead and just tell it. And then it'll provide you with this big long link to click. Go ahead and click that. Open it in your new tab. Log in to your account. And it'll say Google hasn't verified this app. You should only continue if you know the developer. Well, in this case, you're the developer, so you can click continue. And then continue again. Select all to give OpenClaw access to all of those different tools. And click continue. And then you'll get a error that says this site can't be reached. This is actually the expected behavior. Go ahead to your address bar and copy the full URL in there, then paste it back in the Telegram. Occasionally, you might run into an error where the state expires and it doesn't go through right away. Just follow the same process once more, clicking the link, authorizing, and then copying the URL and providing it to OpenClaw. All right, OpenClaw says we're connected to Google. Now that we're connected, let's run a quick test to make sure everything's working right. In Telegram, let's tell our bot add a meeting to my calendar with Joey Nobody on Tuesday the 28th at 4:00 p.m., 1 hour long. It might ask to confirm, like in this case here. That looks right. I'll say yes, good to go. And after a minute, you can see the meeting showed up on my calendar and OpenClaw's confirmed it. Now we know OpenClaw can add calendar events, but can it read them? Let's move this meeting to a different time and see if it can identify that. We'll ask OpenClaw, I moved the meeting. When is it now? And it quickly responds that it's on Wednesday from 2:00 to 3:00. That's exactly right. You can do the same thing with emails, Google Docs, Sheets, anything in your Drive. Google is now connected. Now you have a real agent with real access to real apps. Before we stack more on top, let me show you where the agent stores what it knows about you because that's what makes the agent feel personalized. Every agent has a personality, and right now yours is a little, hmm, generic. The bootstrap wizard made it for his past, but it doesn't know you yet. Four files drive everything about how your agent thinks and acts. I'll show you what each one is for. You won't hand edit them, you just ask your agent to update them and it handles the rest. To access your workspace files, on the left side click agents, and then check out the files tab. Here you can see your core files. This includes agents, soul, tools, identity, user, heartbeat, and memory. Agents is the behavioral rulebook. Things like always confirm destructive actions, use my calendar first, email second when scheduling. Never send email between 10:00 p.m. and 6:00 a.m. To update this, simply message your agent. In Telegram, you can say, "Add a rule to agents.md. Don't check my email between 10:00 p.m. and 6:00 a.m." Just like that, it added a quiet hours rule. And if we close this out, click refresh, click agents and then preview, now if you scroll through, you can find it added this quiet hours rule. This rule exists amongst all these other rules that come default with this workspace file. If you want to edit it manually, you can click editor and then directly edit this text content. Next up, we have soul.md. This is the personality file and most people's soul.md, yours and mine after bootstrap included, is weak. It says something generic like be helpful, not performatively helpful. That's generic. If you want it to feel like a real assistant, you tell it who it is specifically. You can scroll through and see what the default looks like, but you can see it's pretty short. It has a little bit of information from our onboarding earlier. Here you can see mine mentions the travel planning. But if you want to really supercharge the personality for your bot, you can simply tell the bot how it should be. So I might send it a message here in Telegram saying something like, "Update your soul. You are Flint, a seasoned travel agent style personal assistant. Dry, direct, a little sarcastic, and you care about saving me money. Default to short answers and use travel metaphors when explaining things." Again, if you prefer, you can always update this manually by clicking the editor button. Here it says it's updated the soul, and it says, "Consider this boarding pass stamped." Already the vibe's been updated. I love it. Next, let's talk about the user file. This is what it knows about you. Role, habits, family, preferences. It's initially populated during the bootstrap onboarding, but very lightly. You can see there's not much information here. It grows over time as you chat with the bot. You can update it by just mentioning things in conversation. You could say something like, "By the way, I love hamburgers." The conversation says, "I'll make sure every destination comes with a burger recon report." And again, if we refresh this page and then click preview, we can see in the context it now says, "Loves hamburgers." Great. The final file I'll show you here is memory. This is the interaction log. It's what you asked, what it did, decisions it made. It's essentially your agent's working memory. You won't really touch this one directly. It's auto written. This is the file that makes your bot remember what you talked about last week. Dreaming, a process we'll talk about in a bit, processes this file into long-term memory. But, we'll get to that later. You don't have to be a master of these files right away. I just wanted to point out that these exist and where you can find them, so you have a general sense for what they are and how they work. So, now your agent has a personality, knows who you are, and can touch your apps. But, before we put it to work on a schedule, there's one other quick setup piece that pays off if you need to connect external tools. And that's where API keys actually live. From our security discussion, you already know API keys never belong in the Telegram chat. So, where do they belong? They belong in Hostinger's panel as environment variables. That's the permanent encrypted home for every third-party key your agent ever needs. Database access, weather, finance, whatever. Set them there, and OpenClaw reads them automatically on startup. You don't have to think about them ever again. Head back to your Hostinger Docker manager, and then on your OpenClaw project, click manage. Scroll to the bottom to environment, and click to open the drop-down. Here you can see all of your secrets. You have your variable names on the left, and the secret value on the right. Hostinger stores the values encrypted. They're not visible in plain text after save. For a flight monitor automation that we're about to build, the agent is going to need a way to check live prices. A dedicated flight data API is faster and more reliable than web search for this. SERP API is a platform that offers a ton of different Google search API tools. For example, if you go to their dashboard, you can see that they have a ton of these options. And the one we want is Google Flights API. Here we can see the API endpoint, which we'll provide to our bot. So, I'll go ahead and copy this now. And start prepping a message to the bot in Telegram. We'll say, Google Flights API endpoint, and paste that in. And then, go to your account, and grab the API key from here. Head back to Docker manager, click plus environment at the bottom of the environment section and give it a name. Typically, you name environment variables with all caps and underscore instead of spaces. So, we'll say flight_api_key and then paste in our API key in the value field. Next, we'll click save and deploy and then Hostinger will restart the Open Claw project to deploy the changes. This takes about 20 to 30 seconds. After it finishes and it says Open Claw is running, let's finish the message in Telegram. We'll say can you check that you have access to the flight data API? Here it came back and said it wants to see serp_api_api_key. I could just tell it that I named it something else, but to make things simple, let's just name it the way it wants to see it. So, we'll click manage. Scroll to bottom, open environment and then change flight_api_key to serp_api_api_key. Save and deploy and then when it's running again, tell it, "Okay, check your environment variables now." And after a moment and a quick check, it says the flight data API is live and working. It confirms it can reach the service. So, just remember any future skill that needs an API key, follow this exact three-step flow. Sign up for your service, grab your API key and add it into the environment variables, then save and redeploy. Since our agent can now pull real flight prices on demand, let's point it at an actual flight route and have it watch. This is where always-on actually starts earning its keep. Before we build mine, think about yours for a second. What's something you check every day that a computer could just check for you? A stock price? The weather at your kids' school? A flight on a specific route? A YouTube channel you follow? A deal alert? Anything you reach for your phone to look at over and over. That's the first thing you want to automate. For this demo, I'm building a flight monitor since travel planning is my agent persona and since we just wired up the flight data API in the last chapter. Yours can be anything. The pattern is the same. There are two ways to schedule things in OpenGlue and you should know the difference before we build. Cron is run this at 8:00 a.m. every day. Specific times, predictable, cheap. Most of your automations should be cron. Heartbeat is check at a regular interval like every 30 minutes, see if something changed. That burns a lot more tokens because every check is a fresh AI call. Don't put everything on heartbeat. That's how people wake up to a surprise bill. Default to cron unless the task genuinely needs to be done constantly. In Telegram, we'll say using the flight data API we set up, monitor flights from LHR to CDG, flexible dates next month, and message me if the price drops below $150. Check three times a day on cron. And send. The agent will think and then walk you through what it's going to do. After a moment, it'll explain that it set the automation up, but if it needs more details, it might request some first. In my case, it's good to go. Now, if you want to test it to make sure it's actually working, you can say run this task once manually now as a test to see what data is being pulled in. And as it stands right now, it looks like there are already flights under $150. So, if I wanted to, I could update the task and have it look for flights under $100 if I wanted to be really frugal. But regardless, now my travel agent bot can monitor flight prices and will check them for me three times a day. I didn't have to spend a bunch of time coding automation. All I had to do was grab an API key for a specific tool, and then tell Openclaw to work with it. And now that task is automated. The most common mistake on a first automation is stacking everything on a heartbeat. Check my inbox every minute. Check the weather every minute. Check prices every minute. Oof, every one of those is a paid API call running in the background. Use cron unless there's a real reason the timing has to be minute by minute. The automation uses a specific AI model to do the thinking, and which model you choose is the single biggest lever on your bill. Different models cost different amounts. Some are cheap and fast. Some are expensive and smart. Use an expensive model for everything, and your provider balance disappears faster than you'd expect. Use a cheap model for everything, and the agent gets dumb the moment you need it to be sharp. The fix is smart routing. One model for casual chat, a different one for the work that actually has to think. The mental model is three tiers. Tier one is the premium stuff, like GPT 5.4, Claude Opus. These are smart, slow, and expensive. Tier two is the everyday horsepower, GPT 5 Mini, Claude Sonnet. These are fast, capable, and cheap. Tier three is the dirt cheap or free app, like Deep Seek or open models. They're perfectly fine for simple chat. The whole game is matching each task to the right tier, and Openclaw does most of that for you once you set one routing rule. In Telegram, say something like, "Add a rule to agents.md. Default to Sonnet for chat and simple tasks, and use Opus when I ask you to reason, research, or use tools." You can use whatever models you want, but I recommend using lighter weight, cheaper models for simple tasks. And generally, you want to default to flagship, top-end models for more complex tasks. Or agent orchestration. Go ahead and submit that. And Open Claw writes the rule and confirms it is done. This one line can save you 80% of your credit burn. Open Claw's workspace UI can also show you what your usage is and how many credits you've burned. On the left side, go to usage and here you can get a summary that shows you exactly how many tokens you used and what the cost is. So here's the cost for today is $1.10 across two different sessions. And you can always change it to a 7-day, 30-day, or custom time range to see exactly how many tokens you're using. Now while this is pretty good for getting a general idea of what your costs are, I've found it's not 100% accurate. So if you want to see exactly how much you spent on AI calls, the best place to find this is in your LLM provider dashboard. For example, if we go to platform.cloud.com to see the Claude console, you can see the credits that are available and remaining. And this updates pretty regularly, so you can even check this balance between messages on Telegram to see exactly how much each message is costing you. Any third-party APIs you have may also have their own additional costs, so you want to check the dashboards there too to get a sense for how much that's costing you. In addition to setting the model you're chatting with automatically, you can also change the model at any time in a conversation manually. You can do this in the Open Claw dashboard by clicking the drop-down in the chat window and changing the model here. This gives you access to change to any of the models you have connected. Right now, I just have Anthropic models connected so I have the various tiers of Claude. You can also do this directly from Telegram by typing forward slash to show the list of commands and then typing model and submitting that command. Once you do that, you can click browse providers and then it'll show you the list of all the different providers you have connected and you can change your current chat session to any of the connected providers, just like you can in the Open Claw dashboard. Since we're through most of the heavy lifting that I like to use Opus for, I'll go ahead and switch my chat to Sonnet to save a few tokens. And here we get a confirmation that this model will be used for our next message. Here's one troubleshooting note that belongs right next to this cost conversation. If your bot suddenly goes quiet, stops replying to Telegram, stops running its automations, the first thing to check is whether your provider balance ran out. Open your OpenAI or Anthropic dashboard and look at the remaining balance. If it's zero, top it up and the bot should come back to life without you doing anything else. But if the balance is fine, that's when you move on to the real troubleshooting. We'll cover that later in this video. There are three traps to watch for while you're getting comfortable. None of them are catastrophic, but each one will quietly cost you more than you'd expect. First, using a tier one model for everything. That's totally fine for testing, but if you wire a top-tier model up to something that runs every few minutes for a month, that's how you watch a hundred bucks evaporate over a weekend. Match the tier to the task. Second, retry loops. If a tool call fails and the agent retries five times in a row, that's five times you just paid for. This is more likely to happen with a cheap or free model than a tier one model, but it's getting better with time. If a number looks weird on your usage panel, that's almost always what happened. You can avoid this by giving your agent a rule not to retry something more than three times without checking in with you first to find a solution. Third, and this really is a combination of the first two. Heartbeats on top-tier models. It's the same problem we hit at the end of the automation chapter. Expensive thinking every few minutes running in the background while you're not even looking. If you ever put something on heartbeat, double-check what model it's using. All right, our routing rules are set and our bills are under control. There's one more feature worth knowing before we move on. It's a memory trick that keeps your bill from creeping up over time. A quick word on tokens first because it's the concept that makes dreaming matter. Tokens are the chunks of text that AI model actually sees. Every word in your message, every line of your agent's memory, every bit of context it pulls in. You get charged per token. So, the more context the model has to carry, the more each message costs you. That's the problem dreaming solves. Remember memory.md, the daily log? Every conversation gets written there. The file grows and pretty soon your agent is dragging thousands of tokens of old conversation into every new message. Dreaming is the agent going back through its own memory while you're not using it, compressing what matters and dropping what doesn't. So, it stays sharp about you without token cost creeping up forever. There are three phases. Light dreaming skims the last day, pulls out anything important, and updates user.md with new facts about you. REM looks for patterns across a longer window, things like everyday Matt asks about this. Deep rewrites big chunks of memory into a cleaner, more compressed form. The net effect is fewer tokens carried forward, but the tokens that remain are the ones that actually matter to you. All right, here's how to access dreaming. On the left side, scroll down to dreaming under the agent section and select it. Then, at the top right, you can see it's off by default. Just click the dreaming off button to toggle it to on. When you do this, it'll restart your gateway dashboard and this can take a few minutes. Dreaming is off by default because it burns tokens. Those calls show up on your provider bill, like OpenAI or Anthropic, the same place you're already watching for balance. And this is what it'll look like when dreaming is active. Now, I recommend leaving it off for your first week of usage and see if your agent feels useful enough without it. Just because it burns through extra tokens to run the process. I recommend turning it on once you have maybe a month of memory built up. That's when the compression starts paying off. Turning dreaming off does the exact same thing where it'll restart your Gateway dashboard. Sometimes you ask your agent to do something big, research three different CRMs, for example, and give me a comparison. Instead of plotting through one at a time, open claw and spin up a small copy of itself for each piece of the task. They all run in parallel and the main agent stitches the answers back together at the end. That's a sub-agent. Same quality of output, a fraction of the wall clock time. In Telegram or OpenClaw, you can say something like, "Research three scheduling tools, Calendly, SavvyCal, and cal.com. For each, pull pricing, key features, and one weakness. Give me a comparison table. Use sub-agents." OpenClaw will spawn several sub-agents working in parallel. You can see here it's spawning new sessions. Three sub-agents are researching this in parallel. And if you click here in the top left, you can see all the three sub-agents that have gotten spun up. Each sub-agent gets its own session. Look how quick that was. Because each sub-agent handled researching one tool, they were able to work in parallel, and then it came back with the summary super quick. You can also use sub-agents when a task needs to have several big items handled sequentially. You can have your main session be an orchestrator that runs sub-agents to complete smaller parts of the project. Things are working great so far, but something will probably break eventually. Maybe OAuth tokens expire, skills have bugs, Google rate limits you, a sub-agent gets stuck in a loop. All of this is normal and none of it is catastrophic. Here are a few tips for how to handle issues. The first move you have is to ask your agent when an error appears. So, here I got an error about the cron job failing. So, let's just write a message asking for clarification. I'll say, "Hey Flint, what is this error? How do we fix it?" And it comes back with a response having dug into the issue a little bit and providing a potential follow-up solution. You can resolve 80% of your errors like this. Another issue that you might run into is that OAuth or some other credentials that your agent is using have expired. Here I got a message saying that it can't pull information from my inbox because of an authentication error. So here it says to fix it just say re-auth Gmail and we'll walk through it together. So I'll say, "Okay, let's re-auth." And just like that, it'll initiate the process to reconnect the account. If you get an error and you want to drill deeper to try to understand what's going on, you can always go check out the Open Claw logs. On the left side in the dashboard, scroll down and click blogs. And here you can get a full list of everything that Open Claw has logged, every action it has taken, every warning, every error, every little bit of info ends up here as a detailed list of logs. If you're not sure what you're looking at here, they may not be immediately helpful to you, but I wanted to show you where they are because this is where you'd go to find all the information about every error that Open Claw encounters. And Open Claw has access to see these logs of course, so you can always just chat with the agent about the logs that appear to figure out the solution to any error that you're encountering. If something's truly broken and you have no way to fix it because the configuration got super messed up or something like that, you can always revert to the last known backup or snapshot. To do this, go to your hosting or Docker manager, go to the backups and monitoring tab, and then here you can see the list of daily auto backups if you activated the daily option. Otherwise, these will be backed up weekly. To restore from a backup, you can simply click the restore button. Click the checkbox to confirm that you understand this will replace all current VPS files with the backed up version and then click restore. I'm going to hit cancel though because my version's working fine. You can also take a snapshot of your VPS at any given time. On this same page, scroll down to the create a snapshot section and click create snapshot. Click create. And after a minute, you'll get a confirmation that says snapshot created. I recommend taking snapshots before making any big configuration adjustments that might mess up your system. You can restore from a snapshot just like you would from an auto backup. This way, if anything goes wrong when you do the configuration, you can always roll back to the last known working version. Back at the start, I mentioned that some of you will eventually want to run this on your own hardware instead of a cloud VPS. This is where we will cover that move. Reasons you might eventually want to move are you like tinkering, you want zero recurring cost, you prefer owning the machine outright. All those are valid, and it's a real migration path, not a dead end. Here's a quick tour of how the move works. Though my honest recommendation is start on the cloud VPS today, get comfortable, and migrate once you actually know what you want yours to do. In Telegram or the OpenCloud dashboard, type create a full backup. Once it's done that, you can simply say provide me with instructions on how to migrate to a new hardware system using that backup as the base configuration. OpenCloud will provide you with a detailed list of instructions, including commands to run on your local machine where you want to set up your OpenCloud instance. You can simply copy those commands and paste them into the terminal. When you're ready to actually move, the migration itself is pretty straightforward. The full walk-through lives at docs.opencloud.ai/install. Fair warning though, this is where the terminal makes an appearance, which is part of why I held off on this until last. Once OpenCloud is up and running on your new box, importing your backup is a single command from those docs. It takes about a minute. A few things won't survive the move though. That's by design and it's actually good security. You'll need to reissue the Telegram bot tokens through BotFather, reauthorize your Google account through the OAuth flow we did earlier, and your LLM API keys need to be added to a local environment variables file. Essentially, anything that's actually a credential to a third-party service has to be reissued. It's kind of like when you buy a new computer and you have to start by logging into all of your accounts. The computer doesn't have them saved yet. The agent itself though, that comes across whole. All of the workspace and memory files, every skill you installed, every rule in agents.md, the persona you tuned, and every automation you set up. Your agent on the new hardware is the same agent, it just woke up in a new house with new keys. All right, your agent is running. It's on your phone, it remembers who you are, it's running automated tasks, and it'll keep doing all of that while you get on with the rest of your day. We didn't write any code anywhere in the process. The next steps from here are simple. Give it small tasks, let memory build up, and when you're ready, come back and layer in another skill. Tune your soul.md until your agent actually feels like yours with weird little quirks, the way you like to be talked to, the shortcuts only you need. This is where it stops feeling like a product and starts feeling like a teammate. If you haven't signed up yet, use the link on screen or the first link in the description below to get 10% off your hosting or VPS and set up your OpenClout easily with their one-click deployment template. Thanks for watching. Now, go build something cool.