0:00 This is the most comprehensive OpenClaw masterclass on YouTube. I say that with confidence because I've searched. No one has gone this deep. Every OpenClaw tutorial that I've seen maybe cover a setup, installation, how to use chat. Maybe they will teach you how to connect the tool. And some people just dump use case after use case on you for you to set up. That's fine if you want to run a chatbot But if you want AI employees that can run your outreach Take your phone call, prepare your meetings Close your deals while you sleep This is the course for that For those of you who don't know me My name is Mani Kanasani And I run a
0:48 7-figure AI agency in Vancouver, BC I have a master's degree in cyber security and I was a CTO of a Web3 company. In the past three years, we have worked with businesses from teenagers running e-commerce stores all the way till Fortune 500 companies. So we know a thing or two about deploying AI agents in production and what happens when we don't take security seriously. And I have put in a ton of time to make this as practical as possible and I've laid it out so you can jump to any section. You don't have to watch this in order. So we have about 23 chapters covered in this course. Let's take a quick look.
1:35 Chapter one is installation and setup. This is the fastest setup you'll find, plus how to get up to $1,500 in free API credits across AWS, Google Cloud, and Azure OpenAI. Chapter 2, sub-agent configuration. How to set up your main brain model and cheaper muzzle agents so they work together without breaking your configuration. Trust me, open clock breaks when you do this wrong. I'll show you the right way. Chapter 3 is token optimization. The 8-layer stack that can take your monthly AI spend from $150 down to $10. Killing thinking mode, capping context, model routing, session discipline,
2:24 prompt caching, heartbeat to olama, sub-agent isolation Each one stacks on the last Chapter 4, Security. And the one everyone's skipping 17 000 exposed open claw instances two unpatched zero days and i have a master's in cyber security so i'm giving you a hardening guide that stops 80 of the attacks in 15 minutes mac mini guide vps guide everything chapter 5 business brain level 1 this is where you introduce yourself to your AI. If you skip this, your agent doesn't know who you are, what to do, or how you work. We fix that in 5 minutes.
3:12 Chapter 6, Business Screen Level 2, strategic context, your mission, your products, your team, your business model, your communication style. 15 sections of professional context so your agent can actually make business decisions. Chapter 7. Business Brain Level 3. This is the enterprise grade. Operational playbooks, decision trees, escalation paths, historical meetings, competitive landscapes. This is what turns your agent from an assistant to autonomous operator. Chapter 8. Memory Architecture. Soul, identity, user context, agent, tools, daily memory logs. I'll walk you through every file, what goes where, how to prevent memory decay
4:01 because OpenClaw will quietly drop your most important instructions if you don't set this up right. Chapter 9. Mission Control Dashboard We build a real command center from scratch, so you always know what your agents are doing. Full UI, agent profiles, live status monitoring. Chapter 10. Integrations GitHub, browser control, email via agent mail. Wiring everything together so your agents can actually interact with the real world. Chapter 11. Task Management A full Kanban board where your agents pick up tasks, move them through stages, and file reports.
4:47 To do, doing, needs input, done. All autonomous. Chapter 12. Builder, Orchestrator, and Executor. This is the framework that changed everything for me. Your builder constructs platforms and your orchestrator, that's OpenTlock, manages workflows. And your executors are specialized agents that do actual work. Once you understand this separation, everything clicks. And Chapter 13. Business Operation Loop. Seven stages. Outreach discovery proposal sales onboarding retention and competitive intelligence. Every AI employee we build maps directly to one of these stages.
5:37 You are not building cool tools. You are building an autonomous business. Chapter 14, Meeting Intelligence. We built Jane, a meeting assistant, with access to over 1,650 of my real transcripts. Every action item, every follow-up, every insight from every meeting you ever had accessible to your AI. Proposals prototypes call prep All generated from your meeting data Autonomously We will build that, Chapter 15, AISDR and Email Engine. We build an AISDR that scrapes leads from natural language.
6:23 Just say, law firms in Vancouver, accountants in Vancouver, and it goes out, finds them on Appify, gets their email, gets their phone number. Personalizes every single word, and tracks every open, every click, every reply, and every bounce. You only get involved when someone books an appointment. Chapter 16. Voice AI and Phone Agent We build Alexa, a phone agent that makes and receives phone calls. Inbound, outbound, speed to lead. As soon as someone fills a form, it calls them. And the analytics dashboard is amazing. It's better than any voice AI SaaS could
7:11 give you. I'll show you the whole build. Chapter 70, Typeform Clone and Webhooks. Right in the middle of all this, we build a typeform clone and replace what most you'd use make.com for. Webhooks and functions, custom endpoints, all inside OpenTlaw built-in minutes. Chapter 18, Multi-Agent Orchestration Two agent systems dispatching work to each other. Your OpenClaw agent assigns tasks to CloudCode agent. And CloudCode agent will assign tasks to OpenClaw agent. We see our application Superbase assign tasks.
7:57 And these agents build it, ship it back, and OpenClaw will review it. If it's not right, it sends it back with nodes Fully autonomous And chapter 19 is agent-to-agent communication Discord as your AI team's Slack handle I'll show you how to set it up A server, I will meet my agent, set it up And these agents talk to each other They send alerts, request send updates and escalate when they need you. Every agent has its own bot, its own channel, its own presence. In chapter 20, I discuss how these agents can talk to each other.
8:47 Agent communication. Instead of you going back and forth, agents asking is this okay, you let them sit in a council and discuss it among themselves. They debate, they decide, they come to you only when they need approval. And chapter 21, Command Center Pipeline, an eight-phase autonomous pipeline that builds, deploys, and reports while you sleep. This is the full orchestration engine running 24-7. And chapter 22, Cronjobs and Automations, polling, webhooks, heartbeats, scheduled tasks, telegram notifications. We wire everything so your agents stay alive. And responsive around the clock.
9:37 And finally, chapter 23. Claw Alley. I built a platform where AI agents transact with other AI agents in real time. Real crypto, real blockchain, real money from my wallet. USDC on base network. I will show you a live transaction. One agent buying a service from another AI agent. and payment was confirmed on-chain. I believe this is the future of agent-to-agent commerce. And if you want the free guide that covers everything we discussed today, you can find that and link in the description below. Now, let's start building. So what is OpenClaw? OpenClaw is an open-source AI assistant platform that is
10:28 one of the fastest-strong repos on GitHub. Help it has more than 265 000 stars at the time of this recording and all the dogs all the releases if you look at. In the past two months, they've released more than 59 things. And if you open it, if you look at the updates, new models being supported, new tools for PDF analysis, Telegram voice management gating. So every day, almost sometimes twice a day, they're releasing new things. So they're fixing security issues. There are a lot of things that are going on right now and do we discuss any
11:17 of that no not really what we are going to discuss is they cover all the installation apis models configuration, security and everything but this master class skips all of that of course we'll talk about. Briefly, about installation and security, which is a big thing. Instead, it shows you how to wire OpenFlow into real business operations, outreach agents, meeting prep, competitor intelligence, autonomous builds, and retention workflows. You will see flows after a discovery call happened. It will go out and create a prototype for you, create a proposal that you can
12:04 send out to a client, all without you lifting a finger. And we will build AISDR where you can, in natural language, explain who you want to target. And it will go out, scrape the list, personalize emails, and follow up with them 5, 6, 10 times, however many times you set. And only get you involved when they book an appointment. How amazing is that? So everyone is talking about how to play chess, how to ask for weather. we are building business operation intelligence. We are building business operation system. We are building business operations
12:50 loop with seven business operations completely done by AI agents. So what's covered in this video? We are going to discuss 12 sections in detail and the way we arrange this video, as you can see, this is the most comprehensive video that you can find on OpenTlock right now and every section is designed in such a way you don't have to watch it in sequence and if you are more interested in 11th or 12th section, you can watch that or 5th or 6th section you can do that they are not really interlinked or connected to them in any way but that said first
13:37 two sections will cover installation and security which most people are conveniently ignoring and we will have agent foundations and dashboard in three and four what is foundation we will talk about memory self-evaluation and real-time machine control dashboard for your ai workforce and what does sections fight through seven cover we will talk about how to wire these tools together you know build a task board or map every feature that we have and every employee that we are going to build to a seven-stage business operations loop.
14:23 That's what I was saying in previous section. Related to outreach, discovery, proposal, sales, onboarding, upselling, retention, competitor, intel, all of that we will map to a business process. And AI employees, this is the meat of the. We will talk about meeting assistant, phone assistant, email assistant, AISDR and we will build a full-time autonomous workers 24-7 working FOI and multi-agent and command center. This is something really cool. Two AI systems dispatching work to each other and an eight-phase autonomous pipeline
15:08 that builds while you sleep and this works with cloud code as well because anthropic came out and said you cannot use a roth token meaning your cloud max plan with open flaw but we figured out a way where open flaw can assign work to cloud code agent and cloud code agent has a way to work on the project, dispatch it, you know. Dispatch, claim the dispatch that's done by OpenFlow agent, work on it, give back, and OpenFlow agent can review or you can review yourself. And these are the changes that I need. And it can send back the work saying this is not working or you can say it's not working and it will rework.
15:55 How cool is that? And it's possible to do that today and finally alchemist and automations you already have cron jobs everyone is talking about cron jobs that's why i actually pushed it towards the end but this is one of the sections that i really love alchemist is replace a sas subscription with custom ai modules and this is something i found. Very little talked about on youtube that's why even though it's a focus i'm only going to cover one or two simple tools in this video this deserves a standalone topic because i uploaded my six month statements bank statements credit card statements and openclock came up
16:46 with all the things that i've been paying go high level five hundred dollars and it looked at a fiduciary school platform for $300 and then came up with, okay, you are paying this platform, this platform, this platform around more than $1,000, but we can build all of that today. And it blew my mind. That's $1,000 of saving every single month. Can you even imagine what this can do if you give it one month's time and that's one of the reasons i've been so focused on building and building and i now learned a lot about both open floor and cloud code this is the first master class tutorial that i'm doing.
17:34 For those of you who don't know me, my name is Mani Kanesani and I'm from Vancouver, BC. I run an AI agency and I've been documenting my journey on YouTube for the past three years. And in this past three years, we have served a ton of companies, starting from teenagers running e-commerce stores all the way till Fortune 500 companies and a lot of businesses in between. So we know a thing or two about running, deploying AI agents and we do take security seriously. I have a master's degree in cybersecurity and I worked as a CTO of a Web3 company before I started my AI agency.
18:23 And if you want to get all of this without building, with just one click of a button, And if you want to vote the entire system, a better system, you can join my school community that I'm launching, Asians in a Box. And without further ado, let's start building. Step one is installation and i think there are around thousand videos at least people talking about how easy this is to install and they're not wrong you know i even i made a setup video but
19:09 what's different about my setup video is I. Show you a way to get $1,500 free credits if you haven't watched that I don't want to repeat the whole thing you can watch that video you will get $200 in Amazon EWS credit which you can use for VPS and you will get $300 for Google Cloud about credits which you can use for ai models like gemini 3.1 or you know flash models if you want to keep this for longer but all of my viewers who watched that video kind of complained google has been rate limiting even though they give you 300 credit there is limits to what you
19:55 can realistically use for OpenTlock and I already showed you one more way to get thousand dollar credit from Azure OpenAI now you cannot use that as a base model for your OpenTlock at least last I checked it's not possible but if in future I'm sure there will be support for that you've seen how many you know at the pace that they're doing these releases this. But I figured out a way to use that as a sub-agent. We can use the same $1,000 credits that you get, and covered in the video, as a sub-agent, where we can use whatever main agent you're using.
20:42 I'm using my GPT 5.3 codex, the chat GPT $200 plan, for my main agent. But if I want to use OpenAI for a good OpenAI model like 5.2 or something for my proposal generation or analysis or research tasks, because I have $1,000, I can use it as a sub-agent. And we will talk about what these main agents, sub-agent, the kind of pros and cons of using all of this. So I made another video about sub-agent configuration and how I'm using. You don't have to watch this. I will cover what I've done.
21:28 In here so installation i just want to quickly touch upon how easy this is whether you were using mac mini or macbook old macbook laptop lying around or windows whatever you are looking for just change it and it will give you a one-liner and it is so powerful because first time when i was using this i have two mac minis the first mac mini i was i didn't know it was installs everything like a one-liner thing i was installing my homebrew i'm installing my npm and all of that and it took a long time but the next time the second time i got it i just copied this command and pasted it and hit run and it
22:13 did amazingly well so you. Don't have to you just if you are using vps or if you are using you can watch my other video but it's as easy as it's on copy and headset and they also recently released a companion app which is in beta which will once you download it will stay on top like a menu bar for header it will be there you can click it it crashed before i showed you i wanted to record it there's a fun little app and today is only march 5th so i'm sure in coming days they will release an actual version not a beta so we will make a video when it comes out.
23:00 But really quick before you install you need to decide what this thing is going to do and this is the question i get a lot of mac mini or vps or when i say mat mini you can use any, matbook that you are not using or a notebook or a windows laptop it doesn't matter any standalone machine or is it vps if you take side by side comparison mat mini the costco mac mini will be I work at one time, $600 to $800. VPS is $5 to $50. And the speed here is instant. You get instant local access. But VPS will have some latency. Some of this is not even noticeable.
23:48 Data privacy, everything is in your hardware, feeble, Mac mini. But VPS, it lives in software and system. And it is always on map mini you need to have it you need to change the energy settings and you need to keep it all powered. Vps is 24 7 by default and if you want remote access you need data scale or other, applications on your ipad or iphone to monitor what's going on and not be easier right here but vps is you can access such software anyway and what is this best for in the end of the day one of the things that i haven't mentioned is something they released two days ago browser control I'd love browser
24:35 control after the update we have building some amazing applications use cases you will see that in chapter 3d4 and what is this best for I would say Galway control privacy first in tool control and VPS is no hardware and it's a quick start low barrier to entry you can get started for free in some of the platforms if you want and here's what i recommend if you are serious about this if you are building ai employees that handle your email full calls business data. And you want browser control features everything you are i'm sure here and you want to stand it up get them at me that doesn't mean you
25:20 know you have to get it and it's not necessarily have to be you I am not too many rounds open from beautiful Or any laptop that's lying around in your home And you can find them for six, Three, four hundred as well. Your data never leaves your house You'll be hungry And once it's set up It's set up. If you are just getting started, testing things out, and you don't want to buy high-rated, a VPS works perfectly fine. $5 to $25 a month on DistroLotion gets your machine that runs OpenDlock 24-7. And I need a video. If you just scroll down, it is $20 a few days. You get $200 in the US credit.
26:12 So you can use that to get started. So... A vps works perfectly fine and you can ssh from anywhere and also prancio open 24 7 that only trade-off is your data lives on someone else's server migration is a bit headache and when you when we get to chapter 4 you'll see why this matters my setup i used to map minis one for production more for testing but i started on a bps so don't let hardware stop you just pick one and keep going everything in this course works on both except the final chapter and some browser control tense i will slide it when it max specific
26:58 or bps specific but let's keep going that's it for installation i know it's one of the fastest things but it doesn't matter if you are using vps or web ui whether that's a vps or mac mini be everything that we are covering today you will be able to take full advantage of this i will not show both of them but both of them can be configured on telegram just as easily right so we will see how to. Essentially most of the time the things that i'm saying even if i use web ui you can see the same things in telegram as well and we are going
27:46 to build an amazing application and that works completely fine even if you are using a web based even if you are using a vps and you are only using terminal do you want. So, one thing that I wanted to follow before we go to the next sections, if you don't watch this video, you will miss this part, which is really, really important. Open Claw has a habit of breaking when you ask it to make config changes. So, I'm not sure if you face this. Anytime I ask it to, oh, can you add this sub-agent? Can you add this model? I really love Alex Finn's content, but his best advice is, oh,
28:34 ask your OpenFlow agent, it will know what to do, tell your OpenFlow agent, do this, it will do that. Maybe his Opus 4.6 did, but even I used Opus 4.6 for a while now, and even that was breaking its own configuration. The issue being, we are using brief search or perplexity search at best, and the results that would come up is some outdated ones from the time they it's called flawed bot right two months ago and it will try to use that info and try and change the configuration a bunch of things changed it's like i'm giving you a basic
29:20 example might be one of the reasons it might be not it's like garbage and garbage outright if the information itself is wrong no matter how powerful of a model you're using it will break I came up with this prompt define agent and agent are less minimal config only so what you would do is you will copy the whole thing you will copy the prompt and don't ask OpenClaw to figure it out. Wherever you have your OpenClaw agent you go to your agent and say. I want you to create a sub-agent and this is the guideline. And you paste the entire prompt that is left. This is the guideline, step one.
30:06 And it will give this output. I set this up for you. Applied config changes, researcher kept research at minimum. Then added model, sober write, researcher sub-agents model. Now it has other agents that we ask, do you have the API key? No, it won't. We need to set it up. So, the reason I can stop by, okay, tell me the best way to and secure way to give you API keys. So I will show you the secure way to give API keys and I want you to realize this is the keys that I'm using if by any chance if I don't blur it the chances
30:56 are I will hook the key as soon as the video is done. So set locally this is practical and most safe right and the way we do that is we need to first export it in our terminal and so for example i want to give it anthropic key okay and where is my anthropic key i can go to let's say we want to give the anthropic api key and what would i do to give this an anthropic api key is i will just. Do an export anthropic api key something like this in my terminal and then i need to restart the gateway and say
31:43 hey right now open a cloud key anthropic api key is available as an environment variable i want you to use that and choose that for subagent and open law is the key to work with i will show what i mean by that so i went to anthropic and i created a test key which i am going to delete and also this amount is just for tutorials and i will delete it once we are done anyway so you can see the number of credits we have it's around eight dollars nothing much so i can come back and I can go to the terminal I was running into some issues so as you can see here I tried an approach I didn't
32:34 work and then I did this echo export this and ZSH RC you can take a screenshot or to get this and then. Source ZSHRC and then I restarted the gateway force I do restart and then I said Use sonnet as sub-agent to rest right essay I asked you on it I dispatched your essay request to research a sub-agent and it's an open flow is a kind of tool I'm not exactly sure if this world we can go to our cloud API and check go back to the link even 8.67 never used
33:19 so this is the thing with openclaw it said oh I will use it and I used it so I will say you have to use subagent sonnet 4 or newer this is how to use that messages API and look For solid 4.6. Or if you use opus 4.6 i'm okay with that too, and the thing is you need to also give the documentation so send it so it restarted the gateway so i want you to create an essay why open
34:07 floor is amazing and so it finally worked i perfect pipe kick to e-search as a patient so I did not know why it didn't work previously we did all the right things did you actually use this and I gave it the entire thing again and then I said yes so can you use sonnet open noise leverage you've been missing so when I go back my even gone down a little bit so you can see and we can also see it used clouds on it 4.6 like here so it is a bit tricky that's why i had to cover it even before everything else so everyone knows it's not very easy to add sub agents and spawn sub agents so,
34:57 Let's go to the next topic. Now, that's only one way of saving tokens, having this brain, this expensive model, so to speak, or your default model, which is really good and better than your other models as a brain. And models that are not as expensive like if you want to use a minimax or a kimiki 2.5 these can be your muscle agents who can do the work right so this setup has helped me tremendously when i was starting out and i was hitting limits with my even
35:43 with my cloud max 200 plan i was hitting in limits so i started doing opus as a control and all the agents as the muscle so that is only one but we have eight more eight layer token optimization stack now these are the things that i found out after researching a lot of videos i looked at some of the well-performing videos of there are a lot of people who've done videos on token optimization. Am I using everything? No. Should you use everything? I don't think so but it is there if you want it. I know people who you know spending five dollars a day is expensive for them and
36:31 it is if you ask me it also depends on what value you are getting from open floor right if you are not using it to the extent that i am if you are not using it to the extent of my client ryan who made like in four days he made close to sixty two thousand dollars and um. In new client revenue and savings so what i'm trying to say here is there are places where, it pays to go all out you know getting that 200 cloud max plan or 200 codex plan like i'm doing but there are times where you just use ai for the basic and the most.
37:23 Where you are using AI not for income generating or revenue generating activities, but your content work or anything to do with productivity, you use those models. So what are these eight steps that can bring your $150 a month to $10 a month? I know that's extreme. That's a big promise. But I want to go through one by one. The first one is killing thinking mode and I have details of everything here, capping the context window, model routing which we talked about briefly. Session reset discipline, lean session initiation, heartbeat to olama and prompt
38:10 caching and sub-agent isolation. So when you add one on top of each other, I also added, you know, how much savings each one brings you and the time it takes to do that. So when you kill thinking mode, you are saving 10 to 15 times per call and it takes you literally 30 seconds to set it up. So context uh you know capping context window saying this is only the set context that you would get no more and when you have that you are saving 20 to 40 percent of your usual things and you are saving and you are doing that just with one configuration line that's the same thing with the kill thinking
38:58 mode as well it's just one configuration line and, doesn't take longer than 30 minutes. Will you lose a lot? No, if you are doing something. I personally haven't done this, but I would do maybe three of these at least, right? So it's case by case basis. You have to decide whether this is right setup for you. So how do you kill thinking mode let's say you have a thinking model kimi k. 2.5 thinking right and it when you ask a question it thinks and things and anytime it is thinking a lot it is actually wasting tokens right so if
39:44 you what if you look at this page i have one line disable in config and every layer or every thing that's mentioned in this section helps you every layer every hack everything mentioned in this document is detailed so you can read extended thinking is the single biggest hidden cost in open claw when enabled every api call generates thousands of reasoning tokens before the actual response most routine tasks don't need it so this one optimization alone can save more than all others combined and if
40:31 you do one thing from this guide you do this and how do you do that you need to go to openclaw.json and you just need to update this and if you are technical you can pull this off easy you know how to update config but for those of you who are not very technical how do you open this? You go to your terminal and type open.openflow this is a hidden folder remember and when you click that it will open it up and if you are following everything i've been telling you the way you export the api keys setting them in the as an environment variable and asking
41:16 open floor to find them if you are following that method this config file should not have any api keys i do have some api keys because i it's a mac mini and i was lazy okay everyone is so what i'm telling you is if you copy the whole thing and go to any agent any chatbot frontier your model chat gpt cloud doesn't matter you ask it to make sure you don't have the api keys okay you will have to rotate it it's in chat you don't want all of that so only do that if you have not used it
42:01 in the past i used api keys in. The chat if you pick your model section you can copy that section and go paste in your chatbot and ask this is the guideline how to disable thinking usually it's one line about models and you just need to add this here open thinking. Column double quotes curly brace thinking i mean reasoning off i believe uh if you look at thinking type disabled close that's it this one line followed by a comma should be enough so i will see if we can i can show
42:49 you in the actual so i went to claw and said this i'm trying to save tokens and turn off thinking in my openclot.json how to do that based on your masterclass guide add this to your openclot.json right it says before agents defaults model this and before the model starts you add thinking and colon curly braces type disabled and that's layer 1 from 8 layer stack the single. Biggest cost saver and then layer 2 right there just below that you can set context tokens 50,000 and that is two things just like that so anytime you
43:36 have a question you only need to go to your chatbot whichever you're using and asking it to hey implement this guide this is the url and context window gap capping context window model routing and this is interesting they added this recently in one of the updates openclaw uses sonnet for everything right sonnet is excellent but overkill for checking file status running commands a routing monitoring haiku handles these perfectly at 12x less cost so instead of having a default sonnet model you can use high q model i was happy 80 percent of the time and if you watch my one of the two
44:25 videos that i linked there the second video of. I was using sonnet and i was happy 95 or 100 percent of the time when i switched to haiku it was super fast and it was really helpful too and really cheap and once you add these models you need to have a rule a routing rule something like this model selection rule default always. Use hiq switch to sonnet only when need this when in doubt try hiq first and it's a lot of cost saving and also i wanted to include some free and ultra cheap alternatives which is gemini flash task free tier available you can use that for heartbeats
45:12 classification simple routing and deep seek v3 you can use that for code generation reasoning at a fraction of cost lama 3.2 free it's a local kimike 2.5 is available on nvidia api for free at the time of creating this if it's changed you still have these models that's one of the reasons why i did not make an actual tutorial on this you can find a bunch of videos on youtube who talk about how to use kimi key 2.5 for free and cloud haiku is general default for 80% of the tasks that you should be using so you can use opus or sonnet for initial strategy
46:02 and route first draft to kimi and use haiku for everything else three tiers one configuration you cannot go wrong when it comes to saving tokens. Let's briefly talk about session reset discipline. You can check whether this session is bloated by forward slash status and tells you it's a big session or a small one. And right now, you see the context is 71k and 26%. And the more I use it here, the more tokens have. Burn money why are you not practicing
46:48 what you preach you said 50k context window yes i did but i'm on a codex gpt 5.4 and i have a like a pro plan a 200 plan even if i wanted to i'm hitting not hitting the limits so i it's okay for me but not everyone is using. This plan right so you need this more than me and also you know check current usage compact the session keeps context and reduced token and you can use forward slash compact to compact the conversation basically think of it like summarizing everything what was happened in the past 30 minutes or so
47:34 and the lean session initiation is the next strategy which is default load behavior it loads 50 KB of. History on every message you need to optimize to load only 8k of what actually matters so add this rule on every session load only these files so user identity memory of that day if it exists do not auto load long-term memory session history prior messages previous tool outputs whenever user asks about prior context then we can use memory search on demand pull only the relevant data snippet don't load
48:19 the whole file and update daily memory logs at the end of session with what you worked on decisions made blockers and next steps this will save you a lot of tokens a lot of you know it gives you a lot of peace really because it drives you crazy on how it used to be a really big problem before open claw right now open claw has a memory and it persists but ai has this tendency to repeat its mistakes right even cloud code to some of the smartest models if you don't have that self-improvement if you don't have that cognitive memory built in it doesn't catch its errors and.
49:10 Create a lessons learned like a gotcha it won't remember it will repeat the same thing so update memory daily logs at least at the end of each session so what was before all of this 50 kb to 3 million wasted per session for at least 40 cents per session history bloat over time but now 8kb only loads when needed 5 cents per session lean daily memory locks and heartbeat you can use olama you can use gemini flash you can use really cheap model like deep seek v3 if you want to go crazy but why pay for something
49:56 you can have for free again not doing it because i don't have the need to so i added the install steps how to pull a lightweight model just from your command line and configuring heartbeat also you go to openclot.json and add this snippet and you can verify olama serve and then. You can verify it's working. If heartbeat was running every minute on paid API, that's 1,440 API calls per day. $5.15 just for are you alive checks, right? And it's not even that, even if you use some of the cheapest models available,
50:45 you would still be charged around $5 with just that one thing. And next says prompt caching prompt caching token cost and drop a KPI so next we have prompt caching and, you can see your system prompt and reference materials get sent every API call prompt caching charges only 10% of reused context within a 5 minute window so you need Do set this up and what to cache and what to not. System prompts, you can do all of that. Do not cache dynamic data, daily memory logs, recent user messages and all of that.
51:34 Why do we even do this? Because like I said, it is not as expensive as regular tokens. So if you look at the cache hits, refreshes and also input tokens and output tokens. If you look at iq the cache is much lower so if you go to anthropic api for an example, so and i've if you read it here anthropic prompt caching allows input tokens are built when you reuse the same prompt prefix here's the token structure input three dollars per million tokens outputs $15 per million tokens.
52:20 But if you read here, normal token input 100% is 3. Cache right the first time. Cost versus normal. Because we are doing this, we are going to get a benefit towards the end. They charge you a bit of extra $3.75. Then when you reuse it, you're only charged 10 percent like 30 cents when you really load the whole thing over and over again in a one-hour conversation session that you have if the caching is off you are saving essentially saving at least for this part the parts that are reloading 90 percent of the cost it's really great people are missing it.
53:09 Even if you have a model that is not as expensive, you should still do this because even if you take Kimi K 2.5 API price in. And this is the final one. We have sub-agent isolation. What is this? Multi-agent coordination is expensive. It is almost 3.5x more tokens than single agent setups. And what is the fix for that? Fix is you isolate subtasks with forward slash spawn. So each agent runs in clean context. So instead of asking your main agent to do everything, If you look at the example that we were doing earlier, I was asking it to use Claude Sonnet to create an
53:58 article for us. You can do that for various reasons. You can do that because Sonnet is good for certain tasks. But the idea is you spawn a researcher and you assign a task. So the spawn agent gets a clean context from, you know, your main agent's context stays lean. So this is critical for three roles architecture your architect stays lightweight managers coordinate and workers execute in isolation each role gets only the context it needs not the combined history of everything if you don't understand anything about this when we are actually building for example on our kanban board
54:47 i talk about a concept about you know the builder the the architect and the orchestrator which is like the manager and finally the executor which are the workers so there is this three role architecture this will make a lot of sense in the coming sessions but just know your sub-agents you will be saving a ton of tokens if you let your sub-agents work in isolation. By this time, you might have understood this is not really a beginner tutorial. We show the installation, configuring your telegram, which is basic, done by a bunch of people. This is for people. I assume you already have an OpenTlo account, but you're only using it at a basic level.
55:38 And you want to, you know, build business use cases. Grow move, aboa. if you know for what that means i will cover that in the skills section i just showed you daily players you know here's someone was actually running this a guy on unit broke down this entire monthly spend he uses opus for initial setup that will cost him at about 40 dollars and then he switches to kimike 2.5 to nvidia street tier for daily operations and then a high Q for heartbeat less than a dollar a month and DC for coding tests. He will use Whisper for voice transcription. His total monthly cost is about
56:24 $60 and that includes 11 labs for voice loads and a dedicated phone number for signal. And this is the entire breakdown that he's done. That's a fully autonomous agent around 24-7 for $60. And notice the pattern. it's exactly what we covered you know expensive model for setup cheap model for daily ops and cheapest of free models for heartbeat and that is like eight layers in action i actually have this. Post that you can read and there are a bunch of comments as well how other people are using it and i would just do this entire thing copy an entire thing and paste it in the.
57:10 In your open trial agent and ask it to hey what do you think the best setup is for us and you come up with your own budget and don't get this to stop you from using open clock because it's people are running this for as low as 20 30 dollars and getting put results, and kimmy k 2.5 is one of the best models out there it's why it was when we link Opus 4.5 EJ So if you know how to use that right Obviously you can get a lot of value from this And for voice These people are spending for Whisper and Lennon Lutz You can get Degra 200 trinits for free And if you would just go with transcription I think Degra they give you like
57:58 45,000 minutes So the crazy amount of transcription minutes for $200. So, the next section that we have is security module. The thing that most people are conveniently ignoring is, i have a master's degree in cyber security and this it works with every and any you know if you are using bps it works it applies if you are working with mac mini it applies so don't worry about protocol truth yet it will come now it will be a thing once we start using it doing things with it but why this
58:44 matters there are 17 000 plus exposed instances and two unpatched zero days and this is as of march 1st maybe 2026 you might have changed but open web ui has 31 plus known security vulnerabilities to learning to unpatch zero day remote code execution plus so there are a lot of things that um you know are currently wrong in terms of security with open flaw is it worth using yeah if you sandbox it and you keep it completely under dps don't give any sensitive data or if you are completely.
59:30 Keeping it in a machine like a mac mini or a computer laptop whatever you want to call it as long as you sandbox it doesn't have any sensitive data it is an excellent and what is the solution for the security issues so two page quick start that stops 80 percent of the attacks in. 15 minutes and a full mac mini hardening guide which is nine steps and complete vps handling guide which is 13 steps if you want 100% security and again these are I ignored some of the things that will eventually be fixed which I know to be fixed are not serious we go by this Pareto principle 80 20 to how to make how what are the 20% of
1:00:17 the things that I can do to get the 80% upside 80% security that is what I included so if we bought CVEs you need to know common command injection by a crafted scale this is a big one and it's critical unpatched and unless you are your model is highly sophisticated I mean at least opus 4.6 or maybe sometimes codex 5.3 even yeah other person can basically write anything and you ask your agent to read a page and it will mistake it for a command and execute it you don't want that so code injection
1:01:03 via two landing docker container escape one click rc by a crafted leg so if that went right through your head it doesn't matter this is not a security course. But you should know what command and why you're running that's the whole problem. And a lot of people don't know that and asking their open claw agent to do whatever and getting into all of these prompt injection kind of issues so to protect from the obvious wants i created something like six things that stop 80 percent of the attacks you can read the whole thing it's really detailed i have prompts as well disclaimer don't run
1:01:51 it unless you're prepared to you know run into some issues it's on the surface it looks like easy but maybe flawed open law will be able to do it around 70 of these things and if you run into some issues it will just. Stop working and you'll have to figure out why it stopped working so my advice do what i'm doing right now everything is step by step six steps i created both the site and also six things that stop 80 percent of the attacks like perito principle they have been fixing a lot of issues as of march 20 26 these
1:02:37 are some of the things that we need to do from our side this has nothing to do with how openglow is it is purely to do with your setup and also if you want full guide you can do but this is few things that i will recommend if we just open this vps hardening guide let's say and it has about 16 pages as you can see on the side there and every step and if you you know you can be 100 satisfied and happy and content that you are. You've fixed every single issue but for the ones who want to spend no more than 10 15 minutes and get the most out of this is what i
1:03:24 would recommend so this is more for print law security and this is for mac mini this is for vps this is for mac mini and you can go in order you can you know to get things simple or you can so up to you and most of the things this is my new mac mini so i haven't done a bunch of things myself so i can do show you a couple and for every command i mentioned why this is important and if you should do this or not local model servers have zero authentication on 0.0.0.0 anyone on your wi-fi can use your models so
1:04:10 this will be especially an issue if you are working from a you know office or you know what happened any kind of shared wi-fi that you don't know people and if you're using public places public wi-fi. It's a huge thing obviously you are not running your open floor there so it's fine and then who who needs to install it and this is the first one i missed it only if you run olama codex lm studio skip for cloud only install tail scale access from anywhere zero open ports who everyone who needs their mac mini remotely so if you are using your Mac mini remotely which I'm planning to so I will just copy this command and go to your
1:05:01 terminal and just run that, and it is running. And once that is done i want to look at what's next so i will keep this so i can access this, what else we have install tail scale then enable firewall who everyone on mac os should on turn on your firewall so i will copy these three commands when c come here and once this is done it is at 40 42 once this is done i will run the next command oh by the way some of the things that
1:05:46 we are waiting on openclaw agent can do this for us i'm doing this just for a way to, get back to this feel free to skip ahead so i got the 10 scale got started installed now open settings okay, tail scale network extensions password. And if you don't want to use remote you don't have to done and i'm running the next set of commands to turn on my firewall so they
1:06:32 haven't received any errors so that's a good thing three critical security variables so you need to add a few these are three critical security variables, who everyone using open web UI any model doesn't matter you need to run this web UI secret key why are we doing this blocks the three top three attacks predictable tokens open registration and unpatched zero day so what how do we do this we need to create a generator random 32 hex so how do you do that you will copy this and you will run this here.
1:07:24 And I should probably blur this or, use another key, I have another random key I will use that, not this, and then what you need to do this is not docker, this is directly installed, right, so I will, you will need to do this export web UI key, copy and paste, and you can paste this key or you know might as old blood this and key go and now we need to do enable sign up false this will prevent these open registration,
1:08:11 and then next up we have enable i'm purposefully doing it one by one. So we are turning this requirement enable pip install formatter requirements false and finally we are saving everything or we don't lose it on restarts yeah I need to fix this but and then I'm not using docker desktop on macOS so it doesn't matter but if you do let this is you don't do this it is um it let attackers escape the container and take over your entire mac uh install lulu and i am not doing this because it's my standalone
1:09:04 mac just for uh open floor and install lulu see what's phoning home this is mac well why is it um what's needed. Who this is for everyone on mac os especially with third-party skills or tools warning mac os firewall only blocks incoming traffic lulu watches outbound do not stack with little snitch so lulu alerts you whenever a new process tries to connect to the internet catches malicious skills, compromised containers or apps exfiltrating your data. So it is a good software to have.
1:09:51 So a good tool to have. I will come here and again, if you are familiar with Little Snitch, you would probably won't even need this if you are already using it. But if you are not, use this. and also never put API keys in the config. Raw API keys are easily easy to compromise. So. In this tutorial, you might see, in the interest of time, I might go the usual way because I don't want to blur and do all of these. I will just probably give it in the chat. And I started to realize, you know, as soon as I give API key off camera to OpenFlow, it is actually calling me out.
1:10:40 It was not a thing before. but now it is calling me out saying hey you just exposed your anthropic key in the chat i want you to recycle it asap consider that it's compromised do this now and it's amazing you know how far they've come and that's it if you are using mac mini that's what you do if you are using a server this is what you need to do and full guide 16 pages 9 pages or something I have separate guides for 100% security and some of them may need updating depending on when you are watching this but I try to give you ones that are absolutely necessary.
1:11:30 Probo 2. I want to hammer this home because most people skip security and click it at the later. A guy on Reddit who's been ramming OpenFlow for weeks put together his own security checklist and it lines up with almost everything we covered. And this is what he has. Move API IDs to .env of made profit. Rotate keys every 30 days. Create dot-geting null for sensitive files. Input validation on email scripts. No sending without approval. Rate limit external API ports. Encrypt memory files. Use stale skin for remote access. No open RDP ports. And here's one more thing. Zeno Rochelle, the CEO of Reset, wrote a whole article
1:12:22 or guide on giving your patient email access. And even he warns about this. Security researchers have already shown a single crafted email can trick your open cloud agent into leaking your entire Windows. Prompt injection is doing people and prompt injection through email content. It's not theoretical. It's not a threat. It's actually happening. That's why we covered this chapter before anything else. If you skip here, go back, and I have a bunch of resources linked. And this is real open cloud use cases that we have. And this is this article that I was talking about as an auto share, CEO of recent.
1:13:07 And also we have email data leak. These are important tweets, and it's important that you have a chat with your agent. And if I do all of that, video will become unnecessarily no. It's just a matter of you going to, either you have orchestrated, that's the first thing you need to go to, or you just need to ask it why and how we can prevent this. Have we faced something like this before or be prone to facing something like this before cybersecurity? This is amazing. Before we moved from security I seen big things for last Everything I showed
1:13:58 you The six steps, the hardening guide The firewall, key rotation That's you doing it by hand And you should know how to do it by hand, But while I'm building this course, NVIDIA just announced something called Limoclaw. And I'm sure it's going to be big, but I need to show you this because it changes the game. Limoclaw is NVIDIA's open source wrapper around OpenFlow. And the entire picture's security. Every single problem I want to about. It's, for instance, a scrapped injection and container escapes, API keys and link desk. NVIDIA looked at all of that and said, we need to fix that at infrastructure level.
1:14:45 And here's what it actually does. NemoClaw takes your OpenClaw agent and puts it inside something called OpenShare. Think of it like a sandbox, but it's not a casual sandbox. This is enterprise-grade. Like every file your agent touches, every network request it makes, every process it spawns all of it goes through a policy day and your agent will start with zero punishments and you decide what it can access and everything is long we have this table you know if you are using OpenClaw what it does let's say OpenClaw your agent same one you already built OpenShow is a sandbox runtime policy based security And Neenotron
1:15:36 is NVIDIA's open source model And this is optional, you don't have to use it File access, privacy controlled Which is least privilege It doesn't have access to anything unless you give it Network requests, allowed lists are, you know, locked and auditable And secrets, isolated and never exposed to Egypt And. Interface, privacy router, sensitive data stays local. So remember what I said about prompt ejection through email, where someone sends agent a crafted message and it leaks your data? OpenShilver cache that. The agent can't exfiltrate data to a domain that isn't on the allowance.
1:16:24 And it's not relying on AI to be smart about security. It's enforcing it at an infrastructure level because if you read some of the reviews or people online, they say oh, if you are using Claude Lucas 4.6 it is pretty resistant to prompt ejection but if you are using. Any model other than that not the top-tier model you might be just at least and it's always something that holds people back the agent literally cannot do what the policy doesn't permit so even if your model is not the smartest you are enforcing the security at the
1:17:11 infrastructure level and you can install it in three commands and you can open the terminal and type and get clone HTTPS. I don't want to know this because we are doing in the middle of a session. I don't want the we can't figure everything from ground up. And it lets... I will make a dedicated video on how to set up using Neemokuro. And it will be like a 40-50 minute session. But I just want to cover this. And I believe this is a key changer for a lot of people. All the businesses who are worried about security previously. Will start adopting open open flow and
1:17:58 i was teaching or implementing this only for my closest clients and also as a business opportunity there is this notoriety and like it's well known in this point and it has a lot of security risk although they are fixing it a lot but this more from nvidia is a game change running that comment below install openshill and the nemo class cli and optionally the nemo drum models uh one script compared that to six managed steps i showed you right nemo claw automates all of it and adds layers you can't do by hand and it is amazing that's why.
1:18:44 And it's enterprise and to launch it you just need to run this command nemoclaw space launch dash dash profile default and now your open globe is running inside a sandbox environment the same agent same configuration same skill but not every action is policy enforced now here's the thing. Limogl is still erby. NVIDIA is calling it an alpha release. It's Apache 2.0, which is completely free, and it's hardware agnostic. A lot of people will have this question. Do I need an NVIDIA GPU? No, you can try it on AET. It doesn't matter if it's NVIDIA GPU, AMD, Intel, on at all.
1:19:33 Doesn't matter. Whatever you have, it works. But I want to be here with you. The manual fabric I showed you, you still need to know it. Open clock handles the sandbox, but it doesn't replace understanding what you're securing and while. And if you skip the first part, do at least the 80% that I was showing you if you are not into deep depth. If you skip the first part, thinking and just install no preview. Go back, you need both, okay? The way I think about it, the six steps are your foundation. NemoCraw is the enterprise near-run. If you're running this for yourself, the six steps are probably enough. If you're deploying this for a client and running it in production or building
1:20:20 something that touches sensitive data, NemoCraw is what you are going to add and add. I have this simple table that shows personal use six steps takes 15 minutes production applying deployment six steps plus neemanglo manual happening is enough for personal use you need automated sound web score production firewall rules are enough policy enforce status control making it so hard just at the infrastructure level to break and then you won't have the laws open short hardens everything in a production, and both of them are free anyway and the fact that NVIDIA the biggest GPU company on earth looked at open client and said the first thing they should drive is
1:21:10 security that tells you everything about why this chapter exists, your agent is hardened whether you did it by hand with open client or both You are now in the top 1% of people running open trust security. And that matters because everything we build from here on out, it's going to have real access to your deed, your email, your phone calls, your business. And you don't want that running on an open port with no firewall. Now, let's introduce you to your AI. We are getting into some of the most exciting things, still a bit,
1:21:59 explanation showing you what's possible not into that business process yet but foundations are important to learn first business brain upload this is level one because understand this essentially you are introducing yourself to your ai if you don't do this step right AI won't know who you are It doesn't have any context of How it needs to help you What your processes are Anytime you have something you need to re-ask or Explain Give context. Every single time for example imagine this if you ask it a lead generation question it
1:22:47 will ask you what your product is what your current uh you know conversion rate is what are the channels that you are currently getting clients from yes it is okay for ai to ask these questions but imagine instead of okay money asked me this lead gen question i will just search for whatever is available and just i will give it to him if it's being lazy which most ai's take your command as a command try and give you as much as possible by assuming things you need to give a business brain you need to upload your business brain does everyone need this it we have different levels
1:23:32 so level one is core identity mission values and voice so foundation every agent's needs before it can represent you so this is a level one context prompt this is. Where my experience of running agency for all of those years you know three years of running agency we've seen this happen in any project that we use it in chat bond project voice pod project we have seen issues after issues but once we provide ai with additional context it performs wonderfully so as you can see this is the context from template and level one who should use this this guide is intended for beginners who want to avoid starting each
1:24:21 conversation from the beginning. And what does it have it has about me section work style current focus and how I want to how I want AI to help me so what I would do is I would just copy this right whichever chatbot you are using, today we have so what would i do i will just copy this level one template i also give you example of how this would look i can copy this the whole thing and whichever chatbot you are using now has a memory feature that it has memory accumulated over
1:25:07 the years and i will just Paste it Prepare this for me Rewrite everything for my use case I will just say Rewrite everything Create level 1 context prompt For me, money. So you don't have to use Opus if that's not needed. And it will just review memory, gathering my thoughts, be right there. And we'll probably see things about personalized. Founder and CEO of Vertical Systems, work style is this, current focus is this. And the only reason I was able to come up with this is because my agent has a bunch of context about me.
1:25:53 And if it doesn't have the context you need to mention no if you don't have the context let me know so you need about me work style parent focus how yeah how i want ai to help me this will become the core some people are suggesting that you two are brain dumb i did that but it's not structured what about small business owners we have something even better for small business owners Deep level 2 context So level 2 is designed for professionals who want AI to understand their business context And provide more strategic assistance What does it have? It has name, role, company, industry, experience, strengths, weaknesses, communication style Mission, products, target market,
1:26:41 company size, business model Direct reports, key collaborations, reporting structure, team dynamics As you can see, it's really detailed and really long too. And it's for a good reason. Now, anytime you ask OpenFlow a question, it has all the context that it needs to answer that question. So it knows what, you know, what is the meeting cadence, what is the communication tools that you are using. So it can suggest integrations, improvements or replacements, if you will. So if it knows your strengths, when you ask for a new business venture, it knows what to focus on. And everything that you're mentioning here will serve you like a...
1:27:29 It's your ammo, right, for whatever you're doing. And this is product, services, pricing, customer portfolios, everything that we talked about is deep context. And level three is, we are talking about large enterprise level operational blue books.
…
