➡ Today, we’ll learn how to set up an AI agent using OpenClaw and Ollama Cloud. This process, which was once complex, has been simplified for regular users. We’ll install OpenClaw on a Linux system, interact with it via Telegram, and use Ollama Cloud for AI models. This setup can be done on any computer, and the AI agent can help with tasks like programming and customer support.
➡ OpenClaw is a powerful AI tool that can be paired with Telegram for easy access. It can be set up to monitor your email, respond to messages, and even identify important content. It can also be programmed to perform tasks like writing code, planning vacations, or rearranging work schedules. However, it’s important to remember that the AI isn’t perfect and needs guidance and guardrails to function effectively.
➡ Braxme is a platform that offers various privacy-focused products and services, including Braxmail for secure email, Brax virtual phone for anonymous numbers, and Bytes VPN for IP address protection. They’re also launching Brax SIM Unlisted, a private SIM solution, and have developed hardware like the Brax 3 phone and Brax Open slate tablet. These products are available on their website and through crowdfunding on Indiegogo.com. Brax Technologies, the company behind these products, will soon be open for investment.

Today we will install an AI agent and actually use it. I waited a long time before I thought about doing this tutorial because it was too complex before and not within the realm of the regular user to do. But I figured out a way we will install OpenClaw and we will keep this simple. Trust me, it’s doable. I just want to get out of the theory stage where people say I just talk about AI agents but I don’t actually demonstrate. This is the reason I delayed. I didn’t want to bury you in complexity, but I figured it out and rearranged the instructions so it is easier.

I’m going to go fast with these instructions, but you can always pause the video so you can see the actual screenshots and you can copy my steps. The basic premise of what I teach is always privacy. Safe first While the safest option is a local LLM machine, as I showed in my local AI video, it is an expensive proposition right now, so we will compromise for now by using Ollama Cloud as our LLM source. Based on what you learn here, you may decide at a later point to invest in an actual AI machine, but you can push that decision off to a later date.

Then, after installing the basic OpenClaw agent, we will start interacting with it and I’ll show you how. I let it configure itself and then we can discuss the many potential use cases and I will give you some simple examples. In the end, the use cases will come from your imagination. Right now for me, my Open Claw agent does programming on my platforms and helps me manage my cloud infrastructure. It also does customer support on the Braxme platform. You’ll figure out what it can do for you. Stay right there. Prerequisites While this setup can be installed on a Windows or a Mac, I’m going to focus specifically on installing this on Linux, which is what I want my specific audience to use.

If you are running something else, you can install this on a VM running Linux and then the instructions will be the same. The instructions should work unchanged on Ubuntu, Debian, Mint POP OS and Zorin. It should mostly work on Arch and other distros with minor changes that you’ll know about. If you’re familiar with these distros, the installation is in a script so you don’t have to understand what what’s in the commands. Once the Open Claw agent is running, you will be interacting with it via your messaging app, which will be Telegram in this example. So the actual computer used for Open Claw itself will not matter so much and you will not usually need to access the computer except when there’s a problem.

It can be installed on any computer without any really specific requirements. Any 5 year old laptop would be more than sufficient. Installation will be easier if it’s a regular machine and not some headless device without a GUI like a Raspberry PI. While you can test it on your regular computer, I personally would isolate it to a separate computer as once you get it fully running it can do unexpected things so learn to limit its access. Also, as I said, we will be using Ollama as our source of AI models and the reason for this choice is privacy as we don’t want our agent prompts be captured as surveillance data or even learning data in the future.

You can easily turn this into local AI just by pointing it to a local Ollama model, but the installation is identical to local AI. Again, we will not use Frontier cloud models here, but you will be surprised how capable it is installing the main programs. The script I will lay out here will be just to install the programs for openclaw and Ollama. We will not be running anything related to actually configuring openclaw till the end. Cut and paste this into Linux terminal and press. Very important after running this close the terminal before proceeding. If you ever want to reverse this so you can reinstall again, use these commands to uninstall.

When there are problems. The best solution is often to uninstall and then repeat the process and always close the terminal after running the script. Ollama Sign Up Next you will need to set up an account on ollama.com to access the cloud models. Do a and sign up. Then go to the pricing tab and you will see the different plans they offer. It is possible to just start with the free plan but it will be very limiting enough to be disruptive. So I recommend you sign up for the Pro plan which is $20 a month. There will be quite a bit of AI resources for this $20 so it will be worth it.

It will be hard for an individual to even max out on this plan in heavy use. It is definitely the best deal out there for AI inference. After you create an account with Ollama and you logged in, go to the terminal and enter this Ollama sign in. This assumes Ollama was already installed in the prior step and you will be asked to connect the device to Ollama. Then to just test that Olam is installed. Run this on terminal. This should present the prompt and you can do a test question with the LLM Chatbot Telegram setup. Next, we’re going to integrate Telegram as a channel for us to communicate with the AI agent.

First, install Telegram on your phone and you have to provide it a phone number. And don’t worry, no one will see that number if you’re not a Telegram user because you will really be just talking to yourself. Although you can adjust the privacy settings so no one sees it. But this is important since you will link the mobile Telegram app to all the desktop instances. Next we need to install Telegram on Linux. Depending on the distro. In Ubuntu you just look for Telegram in App center or whatever the equivalent software source is for your distro. In Zorin it is just called Software.

On Mint it is called Software Manager. On POP OS it is called Pop Shop or in Terminal you can just enter sudo snap Install Telegram Desktop. After Telegram is installed on the desktop, run it and it should present a QR code. Then using your mobile Telegram app, go to Settings and select Devices and then link device and it will give you an option to scan the QR code and then point it at the QR code displayed on the desktop. Then the two devices will be linked. Next step is to search for Bot father in Telegram, then start a chat with it.

This special Telegram user provides the API token to connect to OpenCloud in your bot father conversation. Enter new bot and this will then prompt you to the process of giving the bot a name ending with bot and then it will give you an API token. I would take that API token and keep it available in text editor as we will need to collect all these APIs in advance. Later on you can refine your bot by typing my bots in the bot father conversation and then you can edit the bot, add a picture and so on so it’s easily recognizable.

We’ll get back to Telegram as one of the final installation steps as this is not done yet. Brave setup Next we need to add a web search capability. This is simply going to be done by going to brave.com and sign up to create a login. Once you’re in, go to Search API and go to the section API tokens. Now add an API token and you can call it openclaw. Again record the Brave API key in text edit as we will be using that later. I paid $5 to set up this account, but that gives you 1,000 searches a month.

I don’t use that much and brave gives you 1000 credit each month. So basically the $5 is a one time fee for the average person you you will not likely pay more in the future. Openclaw configuration the next step is called openclaw Onboarding. Enter Open Claw setup, then openclaw Onboard and now this will go through the various onboarding steps. This will prompt you through the process. Say yes to Personal by default, then select Quick Start mode. Select your model which will be Ollama Provider. Select Cloud and Local. Enter the model manually. I manually type Ollama minimax M2.7 cloud.

Then select your channel which will be Telegram. Then enter your Telegram bot token which I have in text editor on the side so I can cut and paste. Then I enable the dmpolicy pairing if it prompts for that. Next I enable web search by selecting Brave. Then I enter the brave API key under configure skills say no Then I do skip for now on the final step. And then I select hatch in the terminal to exit I type exit. Now make a note by scrolling up to get the local control UI which is this link above with the token.

This gives you access to OpenClaw via the browser. Paste that in your browser and then on this computer you can access the openclaw gateway UI at any time so you can bookmark this. You can already chat with your assistant using this browser link, but in real life use I don’t even use this since I can access everything through Telegram. But we haven’t done the final step with Telegram yet which is to pair it pairing telegram with OpenClaw the final setup that requires Terminal is to pair with Telegram. To do this go to Telegram, search your bot by name and type Start.

Or there might be a start button and then it should respond with a pairing token and that is something you also run on the command line. From here Telegram should now be active and you should be able to say hello to your agent. Setup is not yet complete, but from here we will have the agent do the rest of the work. Important OpenClaw commands Even though we don’t usually interact with OpenClaw via the command line, if things go wrong, be prepared to run these commands. Commands Gateway restart. You do this every time you change some configuration.

Open Claw Gateway restart Open Claw Configuration. This is the same as the onboarding, but you can go specifically to the section you want to modify. Open Claw configure OpenCloud Dr. If you make manual changes to the OpenClaw JSON file and there’s some problem, you can use OpenClaw to figure out what the problem is. Or you can restore from a backup or of openclaw.JSON openclaw Dr. All the important settings in openclaw are in your home folder under dot openclaw and there you can see your actual configuration file which is openclaw.JSON. this can get corrupted so I always back this up every time I make a change so that I always have a running version.

All of its actual running files are kept in.open claw/ workspace initializing openclaw the first thing you likely have to do is respond in telegram and tell OpenClaw who you are and what the agent’s name will be. In my case I’ll call him Braxbot. From this point on, OpenClaw can actually help you with installation and I will show you how it does that. Memory via Vector Embeddings One of the biggest features of AI is when you can use Retrieval augmented generation or rag. Well in open Claw this is very easy. You just tell it to install a vector embeddings model and alarm happens to provide a small one called nomic embed text and it runs as a local model.

Then I tell openclaw that its memory will be stored in these vector embeddings and instead of just text files. This way it is easily retrieved and can be indexed. Other than just telling it to do that, there isn’t much else I have to do and I’ve just given it long term memory. Just watch this conversation. I have to install embeddings. Set up email link with Himalaya plugin Next I want to demonstrate a very practical use of OpenClaw and that is to monitor your email. So in this example I set up the email with nothing more than giving it my email address and password and it took care of all the details.

Sa. Then I proceeded to ask it what’s in my messages today and then I asked it to reply to one of the mess. Sa. And I will show you the contents of the actual reply made by the agent. You can see quite a detailed answer here and you can see my trigger instructions were very short. Then I asked it to read the email to give me security codes and also to scan the email for subject matter that might be sales related and that I need to prioritize and it is able to do identified the security code and gave it to me which could be a 2fa code sent by email when you’re not on the computer.

And then it identified a potential sales email based on request for consulting information. The potential of an Agent the value of an agent is that it is independent and has long term memory. So this goes way beyond what a chatbot can do. In this example I’m talking to it via telegram and giving it instructions, but I could give it a schedule and say check my email every hour and act on it in some way. It could do this automatically or just forward the email to various parties. Depending on the content, this could be a way to limit the mailbox traffic.

For example, once you give instructions to the agent, it will be in its long term memory again. Using just conversation, you can give a sort of a user’s manual and a rule base to check every time some scheduled task is run and that can be indexed. If you tell it to store that content in vector embeddings, you can fine tune the behavior of the agent using the various markdown files in workspace which are Agent Heartbeat md, Soul md, Identity MD, Tools MD and Memory md. Although we already superseded Memory MD by telling it to look in vector embeddings, you can document what the agent can do via a skills file and it can populate the skills file itself.

For example, you can give it rules to define what prompt injections are so it doesn’t respond to suspicious email that are actually prompt injections. You don’t have to edit these files by hand, just tell the agent to handle it for you and you can invoke a skill like checking email by skill name so you can provide detailed instructions. You can get as sophisticated as I get. My OpenClaw has SSH access to my servers and it can look at the source code of my application and write code. The thing though is that I limit the write access to a designated directory and that’s the only place it’s allowed to read and write, although it can read anything and I have IT programming in PHP, Python, JavaScript, Bash or whatever else I need.

One of the main skills of an agent is to do research for you, whether it is something like planning out a vacation in Greece day by day or rearranging your work schedule. OpenClaw can do that or it can write a game for me as it did here. Agents aren’t perfect the biggest thing to remember as you learn to interact with an agent is that the agent isn’t perfect. It is limited by its context size, which is typically 128k, and the capability of the model. Frontier models like Claude 4.7 and ChatGPT 5.5 are much more sophisticated, but for typical use you will find that the Ollama open source models are pretty effective and can do the job very well.

Each model reacts differently to your instructions. So as you get more sophisticated in your instructions, you have to test how each particular model will interpret it. Don’t assume they all work the same way. The smaller models tend to act like grade school kids, while the frontier models are like college professors. The rest of the models are in between. In this case, I’ve tested quite a bit with GPT OSS120B and Minimax M2.7 and know they work pretty well. But you can change models very easily even on the fly. Just type the command on Telegram model, then the model name and as long as the model is predefined in OpenClaw JSON you can call it or have OpenClaw added.

From here on, OpenClaw can do the setup for you, but don’t trust it completely, it will screw up. So back up the OpenClaw JSON or tell OpenClaw to make a backup before touching it. What you have to learn from here is that the agent’s capability is decided by you. It will not decide it itself and without your guidance and guardrails it will screw up. But just like any student, it can be trained and even the basic move of giving it long term memory is enough to guarantee that it will be learning on its own. The New Interface One of the things that OpenClow has taught me is that I can actually be quite productive even without my laptop with me.

I was programming changes to braxbot on braxme while just talking to it via Telegram. In essence, if it has access to it, there’s no need for any special interfaces to get information. Just ask the agent and it will do it for you and the agent can be sitting somewhere else. This has triggered us to come up with a redesign of a flip phone concept device where the main interface is just a chat messaging system and this will be powerful enough to do research, acquire information and perform tasks without requiring any apps. Summary the next step is to imagine some other uses of this As a powerful assistant or employee, whatever your preconceived notions are about AI, this agent use now installed at the level of a normal person, is a significant productivity tool.

Whether you’re in tech, sales, marketing or customer service, there are already so many potential use cases for increasing productivity. For me, this has increased productivity likely around 5x, like having five extra employees. And in my case I invested in a local AI computer so my AI access is even free after the one time cost. Currently the cost of local AI is the only major drawback, so this use of Ollama is the least dangerous the AI alternatives. Hopefully in the future everyone can afford to install a local AI. Enjoy folks. I’m continuously searching for products and companies that can help us in our privacy quest to provide a platform for discussion where you can debate or discuss privacy issues.

Please join us on braxme and share your knowledge about privacy or learn from others at a more advanced stage than than you are to support this channel on the store. On Braxme we have products that are available that we have created to provide privacy solutions. We have braxmail for identity protected email. We have Brax virtual phone for anonymous numbers. We have Bytes VPN to protect your IP address from location tracking. I’m also introducing you to Brax SIM Unlisted, a privacy focused no KYC SIM solution. Watch out for this on the braxme site. I will provide a link to the sign up page for this service when it is ready, which is soon.

On the hardware side we have been creating hardware solutions. These are being crowdfunded on Indiegogo.com and managed by the sister company Braxtech.net the Brax 3 phone is now on a second batch and is now shipping it is in stock so get yours today if you missed the first batch. And the new Brax Open slate is the new Android and Linux tablet which is expected to ship in the fall. If you want to learn more about these products please go to braxtech.net and they are sold on indiegogo.com another announcement is that Brax Technologies which builds the hardware products will be accepting investors very soon via Regulation CF shares, so stay tuned for for that.

Thank you once again to those who support us on patreon locals and YouTube memberships. Your kindness is appreciated and encourages me to carry on. See you next time.
[tr:tra].

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