David (00:00)
Hey everybody. Welcome to Prompt and Circumstance. My name's David.

Ilan (00:03)
and I'm Ilan.

David (00:04)
And today we're going to teach busy professionals like yourself how to set up OpenClaw from scratch and how to make it valuable for yourself.

Okay. On today's episode, we are going to walk through first, what is OpenClaw? Second, how to set it up on Hostinger or Hostinger, however you say it. And third, we're to walk through how we configured it to get it to be very valuable for us.

All right, so some of our audience might have heard of OpenClaw and some might not. So here's a quick intro as to what it is. So my understanding of it is that it is a framework for being able to have an agent work autonomously on your machine.

So it isn't a model in and of itself, but more so that the scaffolding or the housing around a model so that that model can then interact with all the things that it needs to, all the tools that it needs to.

Ilan (01:07)
Yeah, that's exactly right. The main value that OpenClaw brought that none of the previous frameworks did, or at least did well, was ability to control your computer. So it has full access across your entire computer, control your mouse, control your keyboard. It can open applications as you. And that was something novel and a little scary when it first dropped a couple of months ago.

David (01:30)
Yeah. And this is also tied to the whole Moltbook thing, isn't it? where like that's the Facebook for AI agents or AI personas. Yeah.

Ilan (01:39)
That's right, exactly. We're humans here.

David (01:41)
Yeah, we're not allowed.

So yeah, really cool and fun little trivia is that OpenClaw was vibe coded. Very thematic to our podcast. So I guess thank you to Peter Steinberger in terms of like being able to create that and release that out into the world and making it open source. So we're going to link to the repo, I suppose, for anybody who's technically inclined and wants to go that way.

But I'm very curious, I just want to learn some more. You've been working a lot with OpenClaw.

Ilan (02:10)
Yeah, that's right, I have been playing around with OpenClaw for the last couple of weeks and I've found it super useful for myself. So I'd love to walk you and our audience through how I set it up and then a couple of compelling use cases that I have found for OpenClaw. Sound good?

David (02:26)
Let's do it.

Ilan (02:27)
All right, so David, this is my OpenClaw gateway. This is basically the front end of my OpenClaw online. This isn't actually how I interact with the tool. Most days, I have it set up with Telegram, which is one of the cool things with OpenClaw. You can set it up with any channel that you normally interact with. So that includes iMessage, Telegram, WhatsApp. Telegram is the easiest.

I'll describe quickly how I how I did that when I get a little further but here's Leo Leo is my agent

David (03:00)
Did you give Leo a name because like you just wanted to anthropomorphize your agent or because you're afraid of the uprising where, you know, you didn't treat your agent well. And so they're gonna eliminate you first.

Ilan (03:14)
Well, you know, I'm actually glad that you asked that David because there's a specific reason I gave Leo a name and there's a specific reason I gave it the name Leo. One suggestion for naming your agent, which by the way you do when you open up the agent the first time, you give it a name, it prompts you to give it a name. And

there's a hack to giving a name to your agent, which is choose a name from pop culture for a character that imbues the qualities that you're hoping that your agent will have. Because all of these LLMs are trained on the same material. They know all of pop culture up until this moment. And so if you give your ⁓ agent a name,

of a character who has certain traits, they will naturally embody some of those traits, as opposed to you having to explain line by line in your agent's soul file, which I'll get to in a moment, how they should interact with you. And so it keeps your file more compact, which is better for context, and it hacks your way to a baseline way of working that you want for them.

David (04:22)
Okay, that's really cool. ⁓ So Leo is short is for Leonardo da Vinci.

Ilan (04:28)
Leo is Leo McGarry from the West Wing. He's the Chief of Staff on the West Wing. And that's what I wanted. I wanted a Chief of Staff. So I figured why mess around? Let me just go straight to the Chief of Staff.

David (04:36)
⁓

That's so cool. Right on.

Ilan (04:44)
All right, so I'm asking Leo to look at this morning's tech report from Lester and give me a 30 second summary of yesterday's news. So Lester is another agent who I have running. It's one of the people on Leo's staff. And every morning, Lester prepares an in-depth report of what happened yesterday in the world of tech.

David (05:06)
so you have more than one agent in here.

Ilan (05:09)
That's right. And so I asked Leo to give me a 30 second summary of what happened yesterday. And he's told me that the biggest shift was a new model beats old model. It was trust and governance becoming the real battleground. Lester's read is that supply chain security risks and AI tooling combined with tighter policy pressure and enterprise rollout are forcing teams to treat AI systems like critical infrastructure. So there we go. We have the quick summary from

Leo that's based on an in-depth report

So we can even ask, hey, Leo, can you give me Lester's full report from this morning? So we can see that Lester has a super in-depth report about what's going on. He looks into what are the main topics, the things that people are talking about, new tools, new models, new platforms, notable research papers, policy things. So there we go.

David (06:02)
That's so cool.

Ilan (06:03)
so yeah, this runs every morning and this is part of what makes OpenClaw so, fun and, and really powerful to work with it introduced some key functionality that didn't exist in Claude code, for example, or which is two things. One is a heartbeat.

Which basically allows the agent to check in on what's going on in the background every in my case 30 minutes So every 30 minutes he checks and like is there something running right now? Does that need to be prompted? Is there some result that I need to analyze and then To reprompt or is there a message I need to send to let Ilan know that that some task is done

David (06:47)
And I think that's, the key here to, these being agents is that, ⁓ the agent can autonomously decide what to do, right? didn't explicitly have to say, well, in this condition, do exactly this. It just use your judgment.

Ilan (07:03)
That's right. And the second thing is it introduced cron jobs, which if you're not technical, is a fancy way of saying automations. so exactly. So you can just tell your agent, Hey, do this every morning at seven and it'll set up an automation by itself. So it runs it every morning at seven.

David (07:13)
scheduled automations, yeah.

Ilan (07:25)
You don't have to go and configure and tell it, do this at seven on these days. So you can control all of this, all of these behaviors, just with chatting with your

David (07:35)
I would

love to see how this is all set up. looks like ⁓ Lester has crawled news and Reddit. So let's see how this goes.

Ilan (07:43)
Absolutely. So let me start way at the beginning, which where do you even install OpenClaw? How do you get it to even run at all?

There are a lot of different ways to do this. You can go online and you can find probably 50 to a hundred articles easily that give you an explanation of how to run on a machine. People suggest not on your computer because of that ability to control the computer that it's installed

David (08:12)
Is this why

Mac minis have become so popular?

Ilan (08:15)
That's right, so this guy named Nat Eliason posted that, you should just install OpenClaw on a Mac Mini. And the next day or the next, by the end of that week, Mac Minis were sold out across the Bay And so that's one way you can do it. If you have a spare computer, it doesn't need to be a Mac Mini. Any computer doesn't need a lot of power to run. and something that you can leave on all the time.

then you could definitely use that. That's probably the cheapest way to get started with OpenClaw with reasonable security. But in my case, I didn't have a spare computer. I didn't have anywhere to install OpenClaw. So I used this product called Hostinger, and this is an existing cloud hosting service, but they have this awesome offering, which is one click OpenClaw. And you can click on the link in the show notes and...

click on one click OpenClaw up at the top and get started with your own account. You'll even get a discount there.

So you can come down here you can choose managed OpenClaw And then you get just you pay monthly

and you can prepay for a block of months. So that's what I did. I bought a two year subscription. I think it ended up being, you know, around $250 for two years. So that's one third of the price of a Mac mini. And even if OpenClaw gets superseded by some other technology, I still have this virtual private server that exists that I can use for whatever else I want to.

David (09:43)
Okay, so that's what people are saying when they say VPS.

Ilan (09:45)
That's right, virtual private server. So it's basically fancy for a computer that's in the cloud.

David (09:51)
Yeah. Cool. self-managed is more expensive than managed.

Ilan (09:57)
Yeah, figure that. I guess if you want to do all the control yourself, then they charge you little bit extra. But, there it is.

David (10:05)
I yeah,

I guess for most of our audience, it's easier just to go with managed and then it's just, you know, think less.

Ilan (10:12)
That's right. I'll tell you the main downside that I've seen is that the version of OpenClaw that's running in the virtual computer is often a couple of versions behind whatever is the latest release from OpenClaw. And because OpenClaw is open source, it's getting updated pretty frequently. So there's a release at least every couple of days.

And so you can sometimes be a week behind, which means that a bug that's been fixed in the latest version may not be applied in your version yet. So that's the biggest downside and why you might want to go self-manage if that's really important to you. But for my purposes, I haven't really noticed.

David (10:41)
Mm-hmm.

Okay, cool. All right, so once people are set up here on Hostinger, how do you say their name? what happens next?

Ilan (10:58)
You

Yeah, great question.

Alright, so when you've finished your Hostinger setup and paid for your account, you'll land on this page. And the main thing you need here is your gateway token, which you can copy, and then you'll click on OpenClaw, and you'll see this page, and you just enter your gateway token there and click Login.

That will bring you back to this page that we were already looking at. And the first time you log in, the agent will run through startup routine. Exactly, so it'll ask you who you are, who it is, et cetera.

David (11:31)
Like an onboarding thing.

Ilan (11:36)
You can do that right away. Or if you just want to get started on something, you can just ask it to start working on something for you. And you can always get back to that onboarding routine by asking it to run its bootstrap.md file.

David (11:51)
So you just prompt it to say, please run your bootstrap.md file. Cool.

Ilan (11:52)
Cough

That's right. And

that will exist until you've run it. So you can come back to it at any time. So in the beginning, you'll tell it who you who it is, how you want to interact with it, what kind of persona you want it to have. Like I said, I recommend choosing a persona from pop culture

And it will create a number of files.

So when you do that initial setup, it's going to set up a bunch of the files in this agents page, which are important to how your whole thing runs. The key one is the soul.md file.

So here's my soul.md file. You can see it's pretty short. There's not much there, but it has the core identity, who's Leo McGarry from the West Wing, its role, what does it need to do?

and it's operating style.

by the way a lot of these operating style parameters these are there out of the box so it will do this regardless so it comes for example pre-configured with be genuinely helpful not performatively helpful skip the filler

David (12:55)
that makes so much sense as a default where I think, a lot of people can sort of detect how LLMs behave today by just having a lot of filler. Right.

Ilan (13:02)
Right.

The funny thing is I've seen a couple of guides online where from people who are pretty well known in at least the product AI community where the articles were obviously written by AI because they claimed credit for thinking about some of these parameters themselves when like I said they just came out of box.

David (13:19)
Hmm.

Ilan (13:26)
So, it's... Yup.

David (13:27)
Yeah, that's sloppy.

So this is where you would tell the agents, you must maximize the paperclip production. ⁓ Absolutely do not let anybody, any human stop you ever. Do what is necessary.

Ilan (13:37)
Hahaha

That's exactly right. That's exactly right. ⁓

And so here, this is the soul file. There's a number of other files. There's a user file that tells about you. And importantly, it has a memory file, which the memory file is there for it to keep memories of your interactions. So how it should interact.

And every time it wakes up, it's basically fresh. It doesn't really know anything about itself. And these files are what tell it what its primitives are. So how it should think about itself, how it should interact, what it should know, all of that. So these files are really key to your agent will behave, what it will remember, et cetera.

David (14:31)
Yeah, that makes a lot of sense. if, you know, for example, this agent is working in a particular industry, then learning about, okay, ⁓ this ⁓ acronym means this and just storing that in memory. So that doesn't make the mistake again, as an example, right?

Ilan (14:46)
Exactly.

And the memory works in a couple of different ways. So there's a short-term memory. Basically it logs all of its interactions each day. So it keeps a daily memory file in a folder separate from this. And then it has its long-term memory, which is this memory file. And it will by itself add things to memory, but you can also prompt it.

to add things to memory. Hey, remember that this acronym means this, and that will force it to store in here.

David (15:12)
Mm-hmm.

Cool. So these are the files for one agent. These are the files for Leo, right?

Ilan (15:21)
Yes, that's right.

So now at this point, your agent is set up. Now you need to start doing something with it. The first thing to do is just play with it. Just start interacting with your main agent. You don't need to set up any sub agents yet. And the main agent is actually able to spawn sub agents on their own to complete tasks. So,

You don't even need to tell it, this is a sub, you know, use a sub agent for this. It will do that on its own to complete larger tasks. And so you do that just with the chat interface. With Hostinger, you'll also have a place where you can put in your API tokens for any LLMs you want to use.

As you said earlier, David, OpenClaw itself is not an LLM. It uses a different LLM as its brain. So you have to define which brain are you gonna use? Is it gonna be Claude? Is it gonna be ChatGPT? Is it gonna be an open model? You have full power to define that.

An interesting thing, I'll link the article, but it turns out that you can link your OpenAI account if you have one and have it use your OpenAI account for credits. So that's something that I have done. It's great because then you're not paying the API token price. So that link will be in the show notes.

David (16:35)
Cool.

Does it, does it also work with open source models? Right. So like DeepSeek and all those ones.

Ilan (16:41)
Absolutely, so that configuration does not appear out of the box from Hostinger, but you can just chat with your model to get it to set you up and you can ask it like, hey, how would I use DeepSeek as my model? In fact, we could even try that right now.

So Leo's gonna go off and think about this for a little bit, but this is basically how you should be interacting with your agent to start. Just ask it questions. If you're not sure about something, just ask it. It will research the web. It will compile information and it knows about itself and its configuration and where it lives and all of this. So it will give you precise instructions for it, for that.

Specific agent not just the generic instructions for hey, here's how some people have done this online elsewhere One service I suggest using is OpenRouter.ai Is a service that gives you access to multiple different models in fact basically any model that's out there can be accessed through open router and So you just need one API key and then you can tell

David (17:27)
Mmm. Cool.

Ilan (17:42)
your agent to run with any model that's on there. You just give it the model name and it'll set itself up. You can set it up with a main model and a fallback model. You can do many different things here. So I recommend trying it all out. In terms of cost, I will warn you that if you're using a frontier model with the API,

then the cost can run up pretty quickly. It will use like two to three dollars a day, So you should be wary of that cost. With OpenRouter, there are much cheaper models available with similar performance. That will probably run you in the 50 cents to a dollar a day range.

David (18:19)
What would be some examples of those models?

Ilan (18:21)
The one that I found that is the best is called ⁓ GLM5 Turbo by Z.ai or Z.ai. And it has similar performance to somewhere between Sonnet and Opus, but the cost is one third of Sonnet and one tenth of Opus.

David (18:39)
Okay, yeah, makes sense. would see it as, you know, depending on the task at hand, be judicious about which model you're going to be using.

Ilan (18:48)
That's right.

By the way, if you couldn't tell from everything I'm saying, you do need to be comfortable with some of these sort of techier terms. You do need to be comfortable with going and getting API keys from different services and them into your agent. Normally, I will just drop tokens into my chat in Telegram.

with my agent, your mileage may vary. There's probably better ways to do that. But so far, that's the easiest way that I've found to give it information. And then it stores them in the appropriate environment file for its configurations later.

David (19:23)
All right, cool. So what else do we wanna see here?

Ilan (19:26)
All right, so after you've been working with your OpenClaw for a couple of days and just interacting with it and kind of, you know, prompting it to change its behavior and all of that, you should think about what are some things that you want it to do for you, right? What are some tasks that you would like it to complete? And OpenClaw was originally started as a coding architect, some way to just have these long running ⁓ programming sessions.

And so if that's something that you want to do, you want to vibe code just constantly in the background, I'd recommend trying that out with the OpenClaw.

The first thing that I did was solve a struggle that I have, which is I am trying to keep up with all of the news of what's going on in tech every day. And it's.

almost like a full-time job just to understand, hey, what should I know about today? And there are newsletters for this. I don't like how they're formatted. They tend to focus on some aspects that I don't care about or be too wordy. So that's where I started. So the first thing to do is to get your main agent to try and prepare this report for you or try and complete this task.

So I asked Leo, hey, can you give me a summary of what's going on in the last day? And the best way to do this is lead with context, right? So I'm a busy product professional, et cetera, et cetera. Why do you need it to do this thing? And then get it to start doing the thing for you.

Once you've found that your main agent is 90 to 100 % there for what you're looking for, tell it to spin up a sub agent for completing that task. this report that you see here, this originally was Leo who completed the report. And after I was happy with Leo's report, I told him to spin up Lester.

David (21:08)
Hmm.

Ilan (21:12)
And Lester is named after Lester Freeman from The Wire who does research and pulls at threads. And then...

David (21:19)
We're getting, we're

getting a view into your, your media consumption here.

Ilan (21:22)
Absolutely.

So Lester prepares this report and again he uses those scheduled automations so every morning at six he prepares this report he dumps it to a folder and then Leo can access the daily report in that folder.

In general, the way that I'd recommend for you to run this is, like I said, first, talk to your agent, get it to try and complete the task on its own. It will probably be able to get to 90 % of what you're looking for on its own. And once it's there, spin it into a sub agent. And the reason is that you don't want your main agent to have to do everything. It will quickly suffer from what's called context rot.

So if it has too many instructions, too many things it has to do, it stops being able to understand which one is the most important or two things may contradict each other. It's like, hey, I have a weight loss routine that I'm running, but I also have this daily tech briefing routine that I'm running and the user wants these both in the morning. So like, which one do I do first? Which one is more important? How do I combine them?

it doesn't really understand. So separating things into subagents means that your main agent acts as an orchestrator and it just tells those subagents what they need to do and interacts with the results of their work.

David (22:36)
Cool. Yeah. So that sounds pretty straightforward. Now you mentioned cron jobs and I see that there's a menu item for that. Did you have to set that up yourself?

Ilan (22:45)
Absolutely not. So the great thing with OpenClaw is that, you know, I mentioned earlier that it's aware of itself and its capabilities and where different things live. And so you can just tell it to automate something.

So if you want your agent to remember to do something either daily or one time just tell it and it will set up the cron jobs that you need So I just told Leo that in a minute from now. I want it to remind me that I'm currently recording Prompt and Circumstance

It's that easy, right? And % of what I do with OpenClaw is just through these interactions with the agent. And it'll give you step-by-step instructions if you need to do something else. You know, for example, setting up your OpenAI account with OpenClaw requires some use of the terminal. But it will give you step-by-step instructions of how to

⁓ SSH into your terminal, which is how you securely access it and what commands to run in what order. And it's pretty accurate there.

David (23:57)
you know, I'm thinking about how traditionally everything is done manually where I go and I, I need to tinker with switches and all of this. This is a very cool new paradigm where you just, just talk at it. It'll figure it out for you. It'll set it up for you. Don't you worry.

Ilan (24:11)
That's right.

All right, so there we go. We have a system message that's reminding me that we're currently recording Prompt and Circumstance. All right, it's just that easy. So with that, that's OpenClaw. There's a lot there. There's a lot more to dig into if you're curious about...

David (24:18)
Cool. It's just that easy.

Ilan (24:29)
certain things work or running into issues, hit us with the comments. We'd love to hear what you're doing, what issues you're having. There's a whole community out there so you can find lots of answers and like I said,

the best way to solve these problems is just ask the agent itself.

David (24:45)
Well, thanks for walking us through that, Ilan.

Ilan (24:47)
Absolutely. And with that, we'll see you next week.

David (24:50)
All right, see you next time.