➡ Open Claw is an open-source AI tool that performs tasks, not just chats. It runs on your own machine, giving you control and privacy. To set it up, you need a dedicated AI machine, an AI model, a web search subscription, a messaging app, and Open Claw itself. The AI works by interpreting your prompts and executing commands based on them, giving the illusion of a smart, sophisticated system. It can perform a range of tasks, from sending emails to executing command line scripts, all while respecting your privacy and security settings.
➡ OpenClaw is an AI tool that can perform tasks like creating, reading, and executing files, even on remote servers. It’s not perfect yet, especially with graphical interfaces, but it’s great for command line interactions. While it’s not very exciting for personal use, it can be a game changer for businesses. For example, it can handle emails, book appointments, and even create databases for a small business, automating many tasks and potentially replacing some office jobs.
➡ This channel focuses on teaching you about technology and its privacy risks. We offer products like Braxmail for email privacy, Brax Virtual Phone for anonymous calls, and Bites VPN for hiding your IP address. We also sell Google-free phones, with our Brax 3 phone now available for pre-order. Your support on Patreon, Locals, and YouTube helps keep this channel going.

While Microsoft pushes Windows co-pilot and its agentic OS, an open source project called Open Club beat them to it with nearly 200,000 GitHub stars in explosive growth in just a few weeks. As a personal assistant, it’s underwhelming. Maybe even stupid for some. But for general computer work, mind boggling. This is AI that actually does things. Not just chats. Big Tech’s 650 billion AI infrastructure spent in 2026 isn’t mostly for training models. It’s for inference and agents like this. Demand. Massive. Privacy. You control it. No Microsoft spying like we called. Cyber security holes exist, but it’s big tech free.

Let’s dive in. Stay right there. Let’s first get our arms around the topic. It’s about AI that actually does things. Not just talk to you in chat. This AI tool is called Open Claw and it runs AI agents. Microsoft has been selling you this concept for years now with co-pilot. But suddenly we have an open source project where you can run this AI on your own machine where you can understand everything happening under the hood and no secret compilation of your data like what happens with Windows Recall. You don’t have to worry about privacy threats so much since you control everything.

My Open Claw setup. First, let’s discuss the pieces of the infrastructure for running Open Claw AI agents. Number one, a dedicated AI machine. You need a dedicated AI machine. I studied the options and I settled on running it on a computer I already had. It doesn’t have to be super powerful. It can be any kind of OS, but I’m running my AI machine on a Linux computer. Number two, LLM model. Aside from the machine, you need access to an AI model. There are cloud API providers like OpenAI, XAI, Gemini, Anthropic to mention the big players. But for the best privacy, I’m using Olama here which provides open source models and no learning from my prompts.

However, loading Olama models locally didn’t work too well which is too bad since local models are free. The reason is that OpenClaw is a very heavy AI user so it was way too slow to run locally. I tested first with XAI but after one day of use, I already consumed $25 of API time. So that wasn’t going to work. I instead went to Olama and they have cloud models you can access. I signed up for a pro plan which is $20 a month and the price is fixed regardless of how much OpenClaw uses it. For my test, I used the cloud model.

For my test, I used the cloud model GPT OSS 120B cloud. Number three, web search. In addition to the Olama cloud subscription, I also subscribe to Brave Search API which gives me the ability to do web search in OpenClaw. This was $5 for 1,000 searches and I’m nowhere near that. Number four, messaging app. To talk to OpenClaw remotely, you have a choice of different messaging apps but the original OpenClaw one was made using Telegram. I’m not a Telegram user so actually having a dedicated use of Telegram for AI was good. I have Telegram installed on my phone, other computers and my AI machine so I can talk to it from anywhere.

Number five, OpenClaw. Next, you have to install OpenClaw. I’m going to skip the installation process here. I’ll tell you right now that it is complicated and I had to repeat it multiple times to get it right. But quick tip, the only important thing really is to enter the information about the model being used and for Olama it really is just about specifying the cloud model that I’m using. The base configuration is actually in a single file which is OpenClaw.json. So if you back up this file in the workspace directory, you could install this over and over and not miss a beat.

Basics of how an AI agent works. Now let’s explain the very simple way that AI agents work. Once you understand this, you actually have to set aside preconceived notions of how you interact with the computer. Everything you do is based on natural language and that’s what the LLM is for. And most of you have good experience chatting with LLM, AI and some chat interface. OpenClaw intercepts this interaction with the LLM. Just a terminology thing for you when you ask an LLM a question that is called a prompt. The size of the prompt is referred to as the context window.

While you assume it is just a few sentences of your question, this can actually be quite large. The typical context window size today is 200 to 400,000 tokens or words. But to make it seem sophisticated, an OpenClaw passes large amounts of background information with the prompt. This can include memory of past interactions. It also includes a description of how the LLM should act with you. And then the secret sauce is that based on what you ask the AI, OpenClaw gives the LLM instructions that if you ask something from the AI that is a tool available to OpenClaw, then the LLM will insert some text in the response that OpenClaw can intercept.

For example, if you talk to the AI machine and you ask it to email the current weather to you, the word email is converted by the LLM to some trigger format like this. This will then be intercepted by OpenClaw to launch some particular agent. Again, the trigger words and their corresponding natural language equivalent are actually part of the context provided to the AI in every prompt. The ability to pull out trigger words from the LLM response is just referred to as a parser tool and it’s a very common algorithm in programming languages. This, though, is where OpenClaw inserts guardrails to ensure that commands are constrained according to the security rules.

But the other thing that happens is that these trigger commands are put in a queue, so it would seem like a single step but are just commands done in a particular order. This gives you the perception that OpenClaw is super smart and is doing sophisticated work. The actual secret is the amount of context provided to the LLM and its ability to spit out triggers. Large amounts of text are provided to the LLM to even issue just some simple command since it will include all the background context, pre-instructions, and tool definitions, and what was done before. And this is why AI agents are costly if you use token-based pricing because the prompts are huge.

In OpenClaw, it doesn’t track what you do without your knowledge. Instead, you teach it your preferences and then it can save those in files, which it will read and add to the AI context window. You can see what the files are and what it has stored about your interactions here. There’s a skill.md and a soul.md and a memory folder with all the interactions in the past. Again repeating, these files are made part of the prompt context. Memory is just sending all the history of session actions together with the latest prompt. So the LLM interacts with each prompt in isolation.

OpenClaw is the one providing the complete background. Now you understand that agents are simply commands intercepted by the OpenClaw parser from LLM responses. And when a trigger is received, it launches the appropriate tool and the AI will pass it applicable parameters. OpenClaw is smart enough to save parameters like credentials in another file so you can keep interactions brief. OpenClaw specifically listens in on the LLM via what it calls the Gateway, which is like a looping listener. And then every single time it loops and reads what happened before and adds it to the context and gives you the appearance of a continuous conversation.

And as I said before, triggers for agents are placed in an action queue. In terms of execution, OpenClaw agents are just launching script files or adding a new message to the loop by itself. On the script side, anything that runs in command line or CLI will work. CLI means command line interface. This could just be a .sh file or bash command scripts. Sometimes it is some built-in Linux command like curl, MySQL, PHP, sendmail, and so on. Anything you can run in a Linux terminal can be run by OpenClaw. As far as permissions go, OpenClaw is basically using your user ID on Linux and doing whatever your user is allowed to do on the machine.

This is super powerful when you understand it. It also provides you with the mechanism to set a security guardrail. OpenClaw can list files, create files, read files, delete files, set permissions, and execute files. It can also generate the content that goes into a file. But it does this in sequence and seemingly in a single instruction. So if you ask OpenClaw to send email, it will be able to save the credentials to a file, create a Python script to send the email, execute the script, and send the email with whatever message you want. Thus, certain agents are tasked to do these things without further bidding.

That is how it can send email and learn to do it on its own. No email script needs to exist in advance. But it can also execute existing commands in the command line. Let’s talk about crazy stuff. I tested OpenClaw’s full power. I added SSH to one of my other cloud servers, then run commands on that server, and then give me back the results on Telegram. OpenClaw has writes via SSH to create files, read files, execute files, just like it can do locally. So it could write a script in Bash or Python and execute it on the remote machine.

And here’s the incredible part. This is vibe coding. The LLM could write the code by itself. You don’t need to know any programming. Sure, there’s a lot of setup to allow all this because of security issues. So there’s an initial complexity in getting it configured. But after you get it done, suddenly you’re like thinking, what the fuck? What did I enable here? Currently, the limitations of agents are that it doesn’t work as smoothly when it has to interact with a graphical interface like a website. Normally, it just reads the HTML code and sees if it can interpret what the website wants to do.

So let’s be clear that this is the least developed skill in OpenClaw. And this is where the personal assistant uses a bound. This will improve in the future with thousands of developers addressing this. Likely many websites and apps can just have simpler text-based interfaces to allow agents faster access. But when it comes to command line or CLI interactions, this is huge. And enterprise servers are already CLI capable today. What do we use it for? So this is the big question. Frankly, I see so many videos of YouTube creators trying to make it big with broad statements of how fantastic this thing is.

And the main demonstrated uses are very limited. And the reason is that AI agents aren’t really too exciting for personal use. This has been the flaw in the Microsoft co-pilot push in the consumer arena. And OpenClaw is even more limited in the personal assistant side. Due to cybersecurity risk, OpenClaw is locked down so its base configuration doesn’t talk to websites well or logs in without your permission. It can’t go to Amazon and start buying toilet paper for you. At least not in its base state. Sure, in some limited way, it could likely make a reservation on open table.

And yes, it can check weather. And yes, with the proper tool loaded, it can even read, delete, and reply to emails. That one I haven’t tried yet. Sure, it has a crone job built in so it can repeat tasks if you ask it to. Like to test, I asked it to email me a weather forecast daily. That’s pretty much the summation of the personal assistant side of things. The effect of this is that you stop going to the apps or the websites, and instead the AI agent communicates with you via a single interface, in this case Telegram.

But is this a killer app use? No. I don’t spend hours looking at a weather app. Email is likely the only major pain point in personal use. Now let’s get our mind blown. The real game changer here is that the AI machine can create files, read and edit them, change permissions, execute them. These files can contain either data or code. I already said this earlier, I’m being repetitive. But maybe this is not sinking in. You can create an agent that can build out its own infrastructure. It can write its own code and it can connect to the internet and pass this data to other machines.

It can even control remote machines as I’ve already shown. The piece that I didn’t really make clear is that I gave the AI natural language instructions. Like in my test, I said, send an email about a particular thing. And the AI decided it should just write Python code to handle it from there and then create a recurring Chrome job to repeat it. It just did it. AI can just build out its own infrastructure from your instructions and your needs. So I asked it to send an email and it made this script. Now we know AI can write programming code and display it in your chat.

That’s nothing new. But when it can write it and execute it in the same context, it skips a lot of steps in between and then makes it reusable. I didn’t even go this far, but you could ask it to interface with a MySQL database using the MySQL command in Bash. CLI database access OMG. It can create a database on its own. It can query existing databases. Again, if it can be done by a command line, then OpenClaw can already do it. It can access, create and execute PHP. That’s a command line interface. If you want to see all this, ask me to make another video on it.

Let me give you a super simple example so you understand the implications. Simplify small business example. You have a small air conditioning installation company. An AI guy sets up OpenClaw on a Linux server. The AI guy has figured out that the workflow should be triggered by email, all natural language based. This AI machine then is given access to all the company email accounts, sales, service and so on from the website and also the private email addresses of individual employees. OpenClaw can have agents listening to the emails which are tied to the website. If it’s sales related, OpenClaw can book the appointment and send some basic promo materials.

Then OpenClaw can forward the email to the specific salesperson for in-person interactions. But now all the pre-sales is done. If it’s a service call, OpenClaw can automatically forward the email to the appropriate technician. There could be a database of technicians that OpenClaw can create. Communications can be handled by OpenClaw to talk to people on other apps including Signal, Telegram or even SMS. It can even talk to you. This is a small business, yes, simple. Then you can get advanced. You can start creating full sales pitch material so that the AI can actually start convincing the customer why your product is better.

When something is beyond the understanding of the AI, it can then respond that it will forward that question to my manager and then that reaches some higher level human. Sales can be recorded in natural language form and it will be recorded in the sales database without needing any programming or user interface. Orders for a supply of air conditioning units can be handled all by natural language similar to the old days before computers. Inventory can be done by sending packing slip emails to the AI. What I didn’t tell you here is that this could all be stored in structured databases just like some custom program system.

But here there’s no AI created, at least at the beginning. So then you can ask the AI to give you a report of sales for the month. No problem since it can query the database on its own. Suddenly we are automating small businesses that had none of these before. Unfortunately, the office staff that used to handle all of this would likely not be needed. Now this could get too big for one open cloud so you can launch multiple open cloud agents so that each one specializes in one task. The white collar job crisis.

Whether we like it or not, this will affect jobs as I already indicated in my example. Yes, the air conditioning installation company will have the same service staff and the same installation staff and likely the same sales staff but there’s less need for back office staff. We have to be strategic in thinking about what the effectiveness is on the job market. Those in jobs that need to develop trust in customers like sales people, especially for high end products, are very safe. Those that do physical work are safe until robots enter the picture in the job market.

Those that invent and create are safe. AI is not good at inventing innovative new things. For those looking at opportunities, especially those in IT related fields, it is not about being the best coder anymore. As my example shows, it is the practical side of understanding what a company needs and being able to configure an AI agent to perform that task. It’s more about domain expertise. I bet a prior manager of an air conditioning installation company could build an AI system. With a tool like OpenClaw, it shows that a person with domain expertise requires some ability to understand how the AI models work and how it triggers tasks via an AI agent.

Understanding this architecture is not complex, but it can be a very powerful skill. It is not dissimilar to my original role as a systems analyst back in the day in my first job. Basically, AI experts need to interact with real people and learn their needs and tasks and document them, then turn around and tell the AI what those needs and tasks are and have the AI do it. In this case, you teach the AI in OpenClaw based on conversations, but the neat thing is if you save the configuration files it creates, you can reproduce it without starting from scratch each time.

Everything is natural language. All documents are procedures and tasks are natural language, but the end products here could be code and databases. If you want me to dig deeper into these good examples, let me know and I can make you more videos. It depends on your interests, but I am personally intrigued. Let me know your thoughts in the comments. Folks, privacy is of course the main focus of this channel and I teach you technology so you understand the risks technology adds to your life. We have people who discuss these issues, including AI at my platform Braxme.

To support this channel, we have some products in our store that provide the toolkit to retain privacy. We have Braxmail, an email service with unlimited aliases and identity protection. Brax Virtual Phone, anonymous phone numbers. Bites VPN for anonymizing your IP address. The Google phones, phones free from big tech tracking. The Brax 3 phone is on its second batch and is open for pre-order right now at BraxTech.net. The first batch sold out shortly after release. Big thanks to everyone supporting us on Patreon, Locals and YouTube memberships. You keep this channel alive. See you next time.

[tr:trw].

See more of Rob Braxman Tech on their Public Channel and the MPN Rob Braxman Tech channel.