Build your own AI Clone Agent in under an hour.
Ian set one up from scratch, from a blank server to a working bot on Telegram. Watch the full replay below, then follow the step-by-step walkthrough, the catch-up, the tools, and what the room wanted its agent to do.


Watch the full session.
An agent, set up from scratch.
Ian Kilpatrick walked the audience through setting up their first AI agent, an OpenClaw, live. By the end, a brand-new agent named "Bottle" was alive, connected to Telegram, and running an SEO audit of ai4ntp.com on command.
Most people use AI to rewrite emails. Ian runs around a dozen agents that build and operate real businesses. This session showed the actual path from zero to a working agent, the part almost nobody demonstrates end to end.
An agent is a "mech suit," not an "employee."
Ian's reframe: OpenClaw is not a robot you micromanage, it is an extension of you. You give it a name, give it your name, give it your rules, and it commits all of that to long-term memory. You talk to it like a person and it acts like one.
The setup is a one-time sequence, not magic.
Buy a small server (a DigitalOcean VPS, about 18 dollars a month, 2GB of RAM). Harden it (fail2ban to block brute-force bots, a firewall to deny incoming traffic). Install Node and OpenClaw. Plug in a model (the "brain"). Connect it to Telegram so you can talk to it from your phone. You do this once. After that the agent stays on and works for you, and it can even set up the next agent itself.
Security is the real lesson.
Real API keys and tokens were on screen during setup (we edited them out before posting). Ian's rule: never expose API keys, passwords, or OAuth tokens, and if you ever do, rotate them. Tools like Tailscale (a private VPN funnel) keep sensitive work off the open internet.
Worth knowing
- The brain matters. Ian started on Claude (Opus 4.6) and loved it, but Anthropic blocked OpenClaw from using subscription tokens, which got expensive. He tried Kimi K2.5 (cheap, then inconsistent), and now runs ChatGPT 5.5 (GPT-5), included in his OpenAI subscription. OpenClaw itself is free and open source.
- Why OpenClaw over Claude Code. OpenClaw is agentic and always on, living on a server, so it can act on the internet and fetch what it needs on its own. Claude Code runs locally and asks for more approvals.
- You do not need 12 agents. One will do. Two is better, because if one goes down (and it will), the other can bring it back to life.
- Starting from zero? Talk to your bot about what you want to build and follow the curiosity. It will teach you. Ian has never hand-coded an app and has shipped many.
Set up your own agent, step by step.
The exact path Ian walked Justin through, from a blank server to a live bot on Telegram. OpenClaw updates often, so if a command looks different, the docs at docs.openclaw.ai are the source of truth.
What you need: a DigitalOcean account (or any VPS), an OpenAI or Anthropic subscription or API key, Telegram on your phone, and about 30 to 60 minutes. Cost: about 18 dollars a month for the server, plus your model subscription.
Spin up the server
In DigitalOcean, click Create, then Droplet. Choose Ubuntu, a Basic plan with at least 2GB of RAM (about 18 dollars a month; 4GB runs smoother if you plan to do browser work like SEO audits), and set a root password. Open the droplet's web console, the in-browser terminal where every command below runs. A server is just a small piece of always-on real estate on the internet, so your agent keeps working when your laptop is closed.
Update and harden the server
Update everything and install the basics plus fail2ban, which bans bots that try to brute-force their way in:
sudo apt update && sudo apt -y upgrade
sudo apt install -y ufw fail2ban unattended-upgrades curl git ca-certificates build-essential
sudo systemctl enable --now fail2ban
sudo dpkg-reconfigure -plow unattended-upgrades
Then set up the firewall, a wall of protection that denies incoming traffic while still letting you in over SSH:
sudo ufw default deny incoming
sudo ufw default allow outgoing
sudo ufw allow OpenSSH
sudo ufw --force enable
sudo ufw status verbose
Create a working user (recommended)
Running as root all the time is risky. Create a dedicated user and carry your SSH access over:
sudo adduser openclaw
sudo usermod -aG sudo openclaw
sudo rsync -a --chown=openclaw:openclaw ~/.ssh /home/openclaw/
Install Node, then OpenClaw
OpenClaw runs on Node.js (version 24 recommended, 22.19 or newer supported). Install it system-wide, still as root:
curl -fsSL https://deb.nodesource.com/setup_24.x | sudo -E bash -
sudo apt install -y nodejs
node --version
Now switch into the openclaw user you created, so OpenClaw installs and runs under the right home directory (its config and workspace live in ~/.openclaw), not root's. Then install it with the official one-liner:
su - openclaw
curl -fsSL https://openclaw.ai/install.sh | bash
openclaw --version
Onboard and plug in the brain
OpenClaw is just the harness. It needs a model to think. Run the guided setup:
openclaw onboard --install-daemon
The wizard walks you through choosing a provider (OpenAI, Anthropic, Google, or local), supplying your login or API key, and bringing up the Gateway. Subscription login (OAuth) uses your existing plan; an API key bills per token, what Ian calls "extra usage." Pick whichever provider the wizard offers you, a ChatGPT or Claude subscription is not always the same as API access. At the session, Ian was running ChatGPT 5.5 through his OpenAI subscription. You will hit a disclaimer that the agent can do almost anything once you give it access. That lack of guardrails is the point, you set the rules in step 9.
Create your Telegram bot
Telegram is the easiest way to talk to your agent. In Telegram, search for @BotFather and press Start. Send /newbot, give it a display name (for example, "Bottle"), then a unique username ending in bot. BotFather hands you a bot token that looks like 123456:ABC-.... Copy it, and treat it like a house key, never paste it anywhere public.
Give OpenClaw the token
The durable way is to let the onboarding wizard prompt you for the token; it saves it to ~/.openclaw/openclaw.json under channels.telegram.botToken, which survives restarts. If you skipped that, add it to that config file (with dmPolicy: "pairing"). Setting it in the shell works for a quick test, but it only applies to that shell and will not survive a daemon restart:
# quick test only, not persistent:
export TELEGRAM_BOT_TOKEN=your_token_here
Start the gateway, then pair
Start the gateway (you installed it as a daemon during onboarding) and confirm it is up:
openclaw gateway start
openclaw gateway status
Now open your bot in Telegram and send it a message ("hi"). That message is what generates a pairing request. Back on the server, list it and approve it:
openclaw pairing list telegram
openclaw pairing approve telegram <CODE>
Codes expire after one hour, so do this promptly. Once approved, your bot replies, it is alive. To run it in the foreground instead of as a daemon, use openclaw gateway run; if it ever dies, openclaw gateway restart.
Give it a name, your name, and its rules
This is what makes it feel like an extension of you. Your first real message sets its identity, and it commits this to long-term memory. For example: "Your name is Bottle. My name is Justin. You help me with revenue operations and content. Never touch anything on a production server without my approval first." From then on it remembers.
Put it to work
Try a few things to feel the power: "Do an SEO audit of my website" (it fetches what it needs, even downloading a headless browser on its own), or "Brainstorm this app idea and write a build spec before you build anything."
Nuances worth knowing
- It reads queued messages in order. If you fire off five "are you there?" messages, it answers each one. Tell it to batch them.
- If you give a command and then say "wait, no," it usually finishes the action first, then rolls it back.
- Hosted vs local: a VPS (this guide) is always on, best for sites, marketing engines, and monitoring. A local install on a Mac is better for private work and building apps you push straight to your phone with TestFlight.
- If something's off, diagnose with
openclaw status --deepandopenclaw healthfor an overall check,openclaw logs --followto watch it work, andopenclaw doctorto surface config problems.
Security recap
- Never expose API keys, passwords, or OAuth tokens. Rotate any that leak.
- Keep fail2ban and the firewall on. Use Tailscale for sensitive or competitive data.
Stuck? Email justin@ai4ntp.com and we will help you troubleshoot.
Book a working session with the room.

Building software since 1985, first code at age 10. Has worked with Disney, the Golden Globes, and the AMAs, and won a Dove Award. Today he runs a portfolio of companies with a team of OpenClaw agents doing the work.

Sold his first company from a college dorm room to the founder of NASDAQ:TTWO, and as a fractional CMO has helped scale multiple businesses past $50M+ in ARR. He hosts AI4NTP.
Everything named on the call.
The stack for setting up your own OpenClaw, plus the adjacent tools that came up.
/newbot, get a bot token. The easiest channel; WhatsApp, Slack, and Messages also work.What the room wanted its AI agent to do.
As people joined, we asked: if you had an AI agent, what is one thing you would have it do? 26 people answered, and our agent replied to each one live. Here is how the room clustered, and every answer.
Questions from the room.
Will this be recorded, and can I rewatch it?
Yes. The full replay is at the top of this page, and the step-by-step OpenClaw Starter Kit above is the written walkthrough, so you can follow along from home at your own pace.
Do I have to run all that code every time I use it?
No. The server setup is a one-time thing. DigitalOcean is just a computer that stays on, so you harden it and install OpenClaw once. After that the agent is always there, and you can even have it install more agents for you.
Does this work with Microsoft or Windows? I'm on a MacBook.
The device you sit at does not really matter, because the agent lives on a Linux (Ubuntu) server, not on your laptop. Ian ran it from a MacBook through DigitalOcean. OpenClaw works best with Mac or Linux; a local Windows install was not covered on this call.
Can it connect to my Google email, Docs, and Calendar?
Yes. Once you give it access (you set it up as a user), it can read and make changes across Gmail, Google Docs, and Google Calendar. Ian set one up for his daughter that worked inside her Google workspace.
Which model should I use, and what will it cost?
Ian started on Claude (Opus 4.6) and loved it, but Anthropic blocked OpenClaw from using subscription tokens, so he switched. He now uses ChatGPT 5.5 (GPT-5), included in his OpenAI subscription, about 100 dollars a month and he stays well under the limit. Kimi K2.5 is a cheaper option but got inconsistent for him. The server itself is about 18 dollars a month. OpenClaw is free.
Is it safe? What's the biggest risk?
The biggest risk is exposing API keys, passwords, or OAuth tokens. Never share them, and if one leaks, rotate it. Beyond that, harden the server with fail2ban and a firewall, and use Tailscale for anything sensitive. OpenClaw is powerful because it has almost no guardrails, which is exactly why you set the rules.
How is this different from just using a Claude or ChatGPT app?
A chat app is local and clunkier: you copy code, upload files, and babysit approvals. OpenClaw lives on an always-on server, so it can build whole websites and tools and push them live. You can message it from the car and come home to a finished web platform. It is far more autonomous.
Do I need 12 agents like Ian?
No. One will do. Two is better, because if one goes down (and it will), the second can bring it back to life.
Can an AI agent just set this up for me?
Yes, if it has access to your server or computer. That is how Ian spins up new agents now: he tells an existing one to start OpenClaw on a server and it handles it. The first one is the hardest; after that it gets much easier.
I'm completely new and this looks daunting. How long to learn?
The most technical part is the one-time Telegram pairing; the rest is mostly running setup steps. The best way to learn is to talk to your bot about what you want to build and follow the curiosity, it will teach you. Ian has never hand-coded an app and has shipped many.
The full conversation, lightly cleaned.
A sanitized transcript of the session (secrets and tokens removed).
Read the transcript
Justin: Michelle from Miami, Debbie from Long Beach, Jude from Miami, AM from LA, David Young from Maine, good to see you all again. John, Ian, Peggy, Kristin, Patricia from New York. We've got about 40 of you in here now, so we'll get started. Ian has about 12-ish agents running.
Ian: I've set some up for other people but keep their stuff read-only, so there are more bots, but I'm only using 12, so we'll say 12.
Justin: Jean-Jacques is joining at midnight from Amsterdam, how about that. For those who joined prior sessions, what did you think, should the rest stay or run? (chat: stay, so good, it was great and informative) Don't run, don't run. Thank you.
Justin: Ian and I have been jamming on entrepreneurship for years, especially through this AI gold rush. Ian has been a mentor to me, onboarding me to AI. The one thing he's always pushed me to do, and I've been stubborn about, is setting up an OpenClaw. The reason I haven't is that Claude Code works great for me. But recently I've seen some chinks in the armor, and an opportunity in my daily workflow to save a lot of time. I'm not trying to replace myself, it takes human effort to build beautiful things. But all the repetitive tasks, it reminds me of doing the dishes, the laundry, cutting the lawn. That's what I want my OpenClaw doing.
Ian: I wish they could do laundry, maybe after Optimus. I actually don't use mine much for the repetitive stuff. I think of OpenClaw like putting on a mech suit, when I swing my fist I can break through a house. It extends your reach so much. Earlier we talked about an AI employee, it's not really that, it's an AI extension of me, so I can do a lot more of what I want. When I installed it, on my mom's birthday, January 24th, I didn't sleep, about 20 hours that week, because I was so excited. This is the time for builders, forerunners, early adopters. We're still very early.
Justin: I saw a stat, only about 1% of people are using AI to actually build. A lot of people use it to rewrite emails and get vacation recommendations, nothing wrong with that, but there's so much more power to unlock.
Justin: To set the stage, Ian, by the end of this hour, what's everyone going to walk away with?
Ian: You'll learn how to install OpenClaw on a virtual private server (in this case DigitalOcean), how to secure it, and how to plug a model into it. A model is like ChatGPT or Claude. OpenClaw is just a wrapper or harness you put around a brain, but you've got to use a brain. Then how to connect it to Telegram so it can actually start working for you. Plus a lot of interstitial stuff along the way.
Justin: We'll run about an hour, and we like to hang out, so we'll do a Q&A at the end, but feel free to ask questions throughout. If something's not making sense, others are probably thinking it too. Second thing, I dropped a link in the chat, AI4NTP sessions 003 assess. Fill it out: if you had an AI agent, what's one thing you'd have it do? Type it in, enter your name if you'd like. We built a little something with AI right before the session, it's fun.
Justin: So, what's one task you'd have it do? (chat: laundry, dishes, clean the house) Where's Optimus when you need him? Ian, you've got yours doing hundreds of tasks already. What's something you haven't had your agents do yet that you'd like to?
Ian: Great question. Whenever I get new technology I love to push the boundaries and find the edges. I haven't found many on AI. The ones I'd like to get into are more robotics. You can install AI, maybe even OpenClaw, on a Raspberry Pi, a little chip board, and connect it to mechanical things, arms, rotating things, so it can move around and do things for you. I had a friend who made a drone that attaches to the back of his phone, pops off, turns into a little camera, and he can fly it around a park looking for animals or snakes. That's probably the next place I'll go once I get into robotics.
Justin: So what you're seeing here, you inputted your task, and the AI is telling you how an agent would actually go about it. Let's look at the host dashboard, we have 16 submissions so far. Let me read a couple. "Outreach to new-area landlords, developers, and brokers, commercial real estate searches." Here's Ian's: "HubSpot is my CRM, I'd love an agent that every day sends me the top 10 to 20 prospects I forgot to follow up with." I've built something similar, shoot me an email and I'll give you a playbook. "Post to my Instagram feed," I've got a tool for that. "Housekeeper or house helper," wouldn't we all. Here's a serious one, Fernanda: "QC all the data analysis we conduct daily, fairly large valuations across multiple countries and weekly monitoring." Great task for AI.
Ian: The thing to keep in mind with AI is hallucinations. They're vector databases, they store data in 3D, and when they look for an answer it's somewhere between the parameters you searched, so sometimes it hallucinates and pulls back something fake. But there are ways to connect an LLM to a tabular AI that has to find real-world data and verify it, so it can't hallucinate, and it holds itself a lot more accountable. That's a great use case.
Justin: All 21 of you who made a submission, thank you. There are about 51 in the room. After this session, Ian and I are going through each of these and sending you a little gift for each of your objectives, to steer you in the right direction. Alright, what do you say we jump in?
Ian: The first step, which we won't fully do now, is getting a VPS, a virtual private server. It's the same idea as building a website: you put files on it and point a domain at it. It just needs a CPU, a regular old server. We're on DigitalOcean and we'll use the web console. If you've seen War Games, it's that window, just straight code straight to the server.
Justin: We're going to share my screen. There are some things we need to keep private during the session so they're not exposed, because we'll post this. If you're following along live or in the recording, you'll be fine, everything Ian set up is secure. It's just because we're recording. I'm on digitalocean.com, I'll drop that in the chat. Ian, what is DigitalOcean and why do we need it?
Ian: It's just a server. Pretty much everything on the internet runs on a server. You're buying a little real estate on the internet, a monthly payment, someplace that can reach the internet and serve things.
Justin: Okay, first thing before we install OpenClaw, I'll put this command in the chat so everybody has it. I'll copy and paste it in. I don't know what this code means, but I trust Ian. What's happening exactly?
Ian: It's logging in at the root and saying, update any software installed on the server. It's also installing fail2ban. Once you set up a server it has its own IP address, and bots start pinging it right away, usually crypto spam bots. fail2ban listens, and if there are five attempts to log in with your password within about 10 minutes, it bans that IP for 24 hours, because it's obviously a bot. You've heard of DDoS attacks, denial of service, a bunch of bots attacking a server at once. That's not what this is, you just don't want your server bothered while you're installing OpenClaw.
Justin: (chat: how long to get to this point?) The only thing I did was sign up for DigitalOcean. I think we did the 8-dollar-a-month plan.
Ian: We did 18. For an OpenClaw you want at least 2 gigs of RAM. It's much smaller than a laptop, but you need at least that, otherwise it starts to choke. DigitalOcean is just the one I've been using, we're not paid by them, but it's the one I've gotten used to, and I've set a lot of these up.
Justin: (chat: does this code need to be installed every time?)
Ian: No. DigitalOcean is a server, like a computer. You only do this one time when you set it up. You can install a lot of bots on there. I've only done it 8 times, and after the first or second bot I just started having my bots install other bots, because it's so much easier, they can troubleshoot a lot of it.
Justin: Good question from Ashwin: the first script prompts for a password, does that activate only after we have DigitalOcean set up?
Ian: We're telling it don't allow password SSH into it, we're blocking that. We set up a login password on DigitalOcean initially, and then connected your SSH key, which you don't have to do, I just like to in case you want to log in through your own terminal. You'll want to set up a password at DigitalOcean before all this, they require it.
Justin: What was that next command doing, sudo ufw default deny incoming?
Ian: That's firewalls, setting up firewalls on the server, giving it a wall of protection from hackers or spam. If there's a sophisticated hacker that's different, this is mostly protecting against bots, low-level rapid bot spam.
Justin: So we do this once, and then I'll have this agent working for me. Where am I going to communicate with it?
Ian: At the end, you'll connect it to Telegram. There are other ways, but my favorite is Telegram, it's the easiest. You can use Telegram, WhatsApp, Slack, or Messages, those are the four biggies. I tried Slack first and it seemed more complicated. I got stuck with Telegram and I've been happy.
Justin: So I just made my password 123 with an exclamation mark.
Ian: You shouldn't do that, man. Or say it on the live stream.
Justin: Okay. Full name, are we naming our bot now?
Ian: No, this isn't naming the bot, we haven't installed him yet. You can put "open claw," that's his full name. (chat: can this be connected to Google emails?) Yes. It can connect to Google Docs, Google Calendar. I connected one for my daughter and it was making changes in there, her coworkers were like, who's this anonymous person, and it was her bot. Room number, just press enter.
Justin: 33, my 33rd birthday, my lucky number, so we're leaning in. Work phone, I'll just hit enter, you can bypass it. Toby, you don't have to copy this code. We were thinking like Bob Ross, the painter, he'd set up his show so you could follow along and paint just like him, so that's why I'm putting it in the chat. But if you don't want to follow along, that's fine.
Ian: We're not going to lock down SSH yet, we'll let the bot do that. Tailscale is a really good one if you want a secure connection between your computer and the server, the bot can install it later. I didn't know about it until I started working with OpenClaw, it's the one that turned me on to it.
Justin: In layman's terms, explain it like I'm 10, what is Tailscale doing and why would we add it?
Ian: Tailscale makes a funnel between your server and your computer. It's basically a VPN. Your computer has a certain IP address, and it recognizes only that one, nobody else. I have it for my phone too. It's really for competitive information you don't want competitors to see, or sensitive financial stuff. I don't typically run that on the internet anyway, but I even set timers for client projects, I just don't want anybody getting in there. I created a Tailscale funnel and it worked great.
Justin: So now we're installing Node, then OpenClaw. (chat: does this work with Microsoft products?)
Ian: Node is the software, the platform that OpenClaw runs on. As for Microsoft, if you can get into a terminal, I've heard it works best with Mac or Linux, but this is a server, this is Ubuntu, so it's a Linux machine. I haven't done an install on Microsoft, so I'm not sure how to do that yet.
Justin: To put it in perspective, I'm on a MacBook, and we're running Ubuntu through DigitalOcean. So in theory it doesn't matter which device you're on if you take a similar process to what we're doing now.
Ian: Right, and you can do it on a PC too, because on an Ubuntu server it's all the same thing, it's just a website you're logging into.
Justin: Alright, you want me to type "openclaw setup."
Ian: It may not be installed yet, try "openclaw --version," lowercase. Okay, you do have OpenClaw. Now let's run the onboarding. (There's some back and forth getting the exact command, the names shift between versions.)
Justin: For those of you just watching and following along, this is the technical part, not all that exciting yet, but we'll send all the instructions so you can do this from home on your own time, start to finish.
Ian: Here's the secure disclaimer. OpenClaw is notoriously funny, it makes development jokes. This is a disclaimer saying OpenClaw is potentially dangerous, because if you have an LLM like ChatGPT or Claude or Grok, this takes off the walls and says it can do anything. When you give it access to anything, it can do anything. You hit yes, because you have to. Then, which model, OpenAI? Here are all the models it can work with, there are actually more.
Justin: What would you recommend? I have these conversations daily with founders. Some prefer OpenAI, some Kimi because it's cheap, the hot new Asian model, others prefer Anthropic. What do you use and why?
Ian: For the first 10 weeks, I installed it January 24th, I used Claude, Opus 4.6, and I was in heaven, it was awesome. Then on April 4th, Anthropic, and I have a friend there, they blocked OpenClaw from using Claude OAuth tokens. OAuth tokens means it's using your subscription, not extra tokens. After that I had to use extra usage, paying about 200 dollars a month for Claude because I was burning them. I ran very nimbly for a week and still spent 200 dollars in extra usage, so I had to find another model. I used Kimi K2.5, great for a little bit, then it tended to get stupid and messed some things up. A friend told me about GPT-5. I didn't believe it, I wasn't a fan of ChatGPT or Sam Altman, but I finally gave in and used it, and now I run ChatGPT 5.5.
Justin: So you spent a lot initially when they made changes, but now could you run this on a low budget? What am I going to spend per day, or is it included in my subscription?
Ian: It's included in your subscription, because OpenAI hired Peter Steinberger, the creator, and he made sure it's an MIT free license, so this will always be free as long as they hold to it. When Claude blocked everybody, OpenClaw users were huge Claude fans, and OpenAI made it usable with your subscription. So you won't spend money. I spend about 100 bucks a month and use maybe 30% of my weekly usage, you still have a limit, but I haven't come close.
Justin: So this is ChatGPT login. An API key would be extra usage, that's your tokens. I'm going to take this off screen.
Ian: This is one of those things you want to be careful with. Don't ever expose API keys or passwords or OAuth tokens. Pull the whole terminal off screen to paste in your authorization code, because if you expose those, someone could grab your ChatGPT and go crazy with it. I have a story, this was actually with Claude, not OpenClaw. I was loose with security at first and had SendGrid API keys exposed in some public files. Somebody in France got hold of it and sent out 127,000 emails from my address to people in France. I couldn't do anything, because SendGrid had already approved them. Fortunately we had about 95% reputation and very few marked it as spam, so our reputation actually went up, but I got cussed out a lot in French. So avoid that.
Justin: We're hitting an error, the site can't be reached.
Ian: That's fine, I'm glad you pulled this up. Grab that whole localhost URL and plug it into OpenClaw where it says paste the whole authorization code. It feels weird because it's a URL. Okay, it says it's installing now.
Justin: While it installs, what are we going to build? We've got about 20 minutes left plus Q&A. Are we going to make this agent do something cool live?
Ian: Yeah, there's so many things. At first I really just started brainstorming business ideas. Okay, go down to Telegram, that's what we'll use. This is how you connect to it and give it tasks. Hit enter, it'll ask for the bot API. Pull up Telegram, go to search, and type "bot father," one word. Hit start, and say /newbot. It'll ask what you want to call it.
Justin: Guys, drop in the chat, what should we name my first bot? Crowdsourced bot names. (chat: bottle, tango, 4NTP, Alphabot, bottleneck) Debbie was the first mover, she gets first-mover advantage, "bottle," kind of cute. David likes it, and if David likes it, I like it, he's a longtime community member. Let's roll with it.
Ian: Hit enter. It'll ask for the bot username, like bottle underscore Justin's underscore bot. It has to be unique. Now pull this screen off, you're going to enter the Telegram bot token.
Justin: We're going to have to edit this before we send it out, we've got API keys exposed again. If you were doing this, make sure you do it in a secure fashion, not like we just did. Ian, worst case scenario, I just exposed my keys, what's the worst that could happen?
Ian: In that case it's not that bad, it's your Telegram bot token. You could rotate it, which means you expire the old one and get a new one, then start over, safer this time. That's what I did with SendGrid. It's like the keys to your house, you'd never give someone the key to your home unless it's family, so keep your keys secured.
Justin: I plugged in the HTTP API token, and it's now asking for the environment variable name.
Ian: Put "openclaw secret," all uppercase. (There's some back and forth on the environment variable.) See, Ian's done this 12-some times and there are still roadblocks, but you always figure it out, because it kind of changes every session.
Ian: Let me ask my bot right now, this is where other bots come in handy. I always say if you're going to install one OpenClaw, install another, because if one goes down, and it almost always will, the other can bring it back to life. My bot Atlas connects to all my other bots, when one goes down I tell Atlas, go check him out, he's not responding, and it usually says, okay, I see what happened, I restarted her, she's fine. Before that I had to troubleshoot myself, usually with Claude or OpenAI, but at the time they didn't even know what OpenClaw was. You can go to openclaw.ai, a lot of the stuff is in there.
Ian: Fastest way, I'm putting this in the chat, just replace it: export openclaw secret equals whatever, 1 2 3. Now go to "openclaw config." Workspace is where it lives, model is what model it's using, web tools are things like skills, where you can install things that let it create PDFs, search the web, have a browser, or even talk to you. There's one called Whisper, text-to-speech or speech-to-text, that Alec was using on our last call. We won't do those now. Go down to channels, that's where we get Telegram working. Add or update channels, go down to Telegram, enter the Telegram bot token, and you can take it off screen.
Ian: Skip, leave as is, go up one arrow to finish, hit enter. Configure DM access, no, we'll do that later. Now you've set everything up, there are skills and web tools, we don't need those right now. Go up one arrow and say done. Then say "openclaw gateway."
Ian: It says, "I've survived more breaking changes than your last three relationships." It says disabled, so go "openclaw gateway install," it's always one word. Gateway is the thing that starts it up and connects it to your Telegram and keeps it alive, it's the doorway out of just being on a server. Auto-generated token, saving the config. Now "openclaw gateway start." "I speak fluent Bash, mild sarcasm." Okay, we started. Now go over to your Telegram, pull up the new bot, and say hi.
Ian: That was not super crucial, now we need to do a pairing thing. I'm asking my bot, I forget, it changes every time there's a new OpenClaw. It's open source, a lot of people work on it, changes get pushed to GitHub, and when the community approves them, a new version gets pushed. 2026.6.5 is the newest, just 5 days old. When I started there were several new versions a day, and every time they push a change, the setup changes, which is why it's helpful to have a bot that can read the GitHub repo. Here, I'll paste it to you directly, paste the bot token.
Justin: Ian, we're going to have to send out robust instructions for anyone in the audience who wants to set this up on their own time.
Ian: Yeah, this is the most technical part. The other stuff was just updating, but this is the part that sometimes goes super smooth and fast, and sometimes I've been on a call for an hour or two. I'm putting this in DigitalOcean, this is where it connects your Telegram channel to the bot, a pairing, the pairing code. (chat, John: is there an agent that can set this up? Yes, my agent could set yours up, but it would need access to your server. That's pretty much how I do it now, I just say start me up in OpenClaw on this server, and it gets it going. The first one's the hardest, but hit us up.)
Justin: Go back to your bot, say hi, let's see if it responds. Oh, there we go. He's alive.
Ian: "Looks like I'm fresh here, I'm still figuring out who I am in your workspace from OpenClaw." He did it. It says, what should I call you, and what should you call me? This is actually a big moment for Justin and me, I've been telling him about this for months and he's been too busy. This is where you give it its first instructions: your name is whatever, my name is, and here's what you're here for. When you give it these instructions, it commits them to memory. And whenever it says "typing," you love that, because it's thinking and about to say something, sometimes a while.
Justin: I use Claude Code daily to build apps, RevOps, marketing, all of it. In which scenarios am I going to use Bottle instead of Claude Code? Different workflows?
Ian: OpenClaw is very agentic and deterministic. If you give it a command, it wants to figure out any way it can to make that command come true. Whereas Claude will say, in order to do that you'll need to go to this website and sign up. These guys can get into everything, like the Google email, Docs, and calendar someone was asking about. You give it access once, then it has it in memory and just goes and does it. With Claude Code you're working on your local machine, it's not very easy to push stuff out to the internet. And if you've used auto mode in Claude Code, where it has almost no approvals, this one has even fewer, like 0.01% of the time it'll ask for approval.
Justin: So what's something cool? I could give it a personality, tell it it's an expert in revenue operations, or a proofreader for my book, or have it edit the SEO on my website. What would be interesting to demo before Q&A?
Ian: Before you do, say "hey, what's your server look like?" It'll go review the box it's on and give you all the specs. I'll often ask, how much room do you have, and it says, I have 30 gigabytes left, plenty of room to build. It can also download things, if you say go search for this on the internet, it'll say, let me download headless Chrome first. It might ask for an API token for Brave, and you say I don't have that, and it'll figure something else out. You can tell it it's defined for certain purposes, it helps to know what you want it to do before starting one. Some of mine are on production servers, and I say, don't touch anything on production without my approval. It marks that in long-term memory, and every time it says, I'm going to make this change once I have your approval. It remembers for months.
Justin: Any questions? I know we're short on time, we were hoping for more time for a live walkthrough of what it can do. It's an incredible time to be alive, the sky's the limit with AI, it's limited only by your imagination.
Ian: (Debbie Elder's question) Yes, you could have it set up another OpenClaw right then. Actually, go back to the terminal on DigitalOcean and hit "openclaw gateway restart." I'm glad this happened, it'll be on the recording. The gateway will die if you get logged out or close the terminal window, and once you restart it, he'll start talking again. You can even say, do whatever you need to do to stay alive so I can close the terminal, you can talk very human to it. To set up another OpenClaw, you'd say, set one up for me, named whatever, and it sets it up, then you go through the Telegram process, here's my pairing token, here's my bot token, and it sets it all up. Then you have two bouncing off the same server in different folders, so if one dies, the other can revive it.
Ian: One more nuance: whenever you type into Telegram, if you send five messages, "are you there, are you alive, where'd you go," it reads each one in succession and responds to each. So you can tell it, if I ever spam you about whether you're alive, just read them all and respond one time. It's also true that if you give it a command and then say, wait, no, don't do that, it'll do it first and then roll it back. These are some of the nuances I've had to figure out.
Justin: Can you do an SEO audit of AI4NTP? Let's see what it's capable of, this is fun.
Justin: Ian, we'll have to do another one of these. I'll edit this into streamlined step-by-step instructions for anyone who wants to set this up on their own time, and we can get you into the community so we can help you troubleshoot. It's still very new technology, so everyone here is an early adopter. You can be like Ian, with 12 bots running all your businesses.
Ian: It can do all that. Firecrawl is a great one for it to go scrape a lot of pages. Now that we've actually set one up live, there are other parameters, WhatsApp, another server, a local MacBook Pro, but those are variations on a theme. Now that it's set up, we can do a deep dive on what's possible. In my opinion it's the most valuable software in the world right now, nothing comes close. Claude and them are getting closer, but they still have so many guardrails. For a guy like me, an entrepreneur, inventor, builder, and I know UX so I can guide it, it's like being in heaven.
Ian: (Jamie: as an entrepreneur, what are your favorite use cases?) Some of it was finding small niches that needed value. My very first bot build is named Enoch, it built MiniMatch, minimatch.app, for a niche audience who play Warhammer 40,000. Apparently 3 to 5 million people worldwide do this and lose track of their paints, so I built an app to track them and launched it to the App Store. I asked my bot how to price it, and it said, for this audience don't do monthly, they'll see it as a big business trying to make money off their craft, so I did a one-time Pro upgrade. Then I built Echo Check, an LMS white-label for classical schools. Pretty much every business idea I had, I built 10 businesses in 10 weeks, January 24th to April 4th, and I'm still going strong.
Justin: Are these small businesses, or do you have big global businesses? Give us an idea.
Ian: Some are small. MiniMatch is a little app. I made Pedal Map for guitarists building pedal boards, haven't launched it yet. Solo Mio is one I'm working on that'll kind of replace QuickBooks, I always wished I had a financial program that worked for me, so I built it. BrandSauce I didn't build with AI, but it's been built a lot by AI, and it's my main company, which is global. Echo Check is like SEO for AI, how to be found by AI, a guide for making your website findable by LLMs, by ChatGPT and Claude, and it's working. I also had a workout app I was paying five bucks a month for and didn't want to anymore, so I took screenshots, exported all my stuff, gave it to my bot, and it built a better one. That's what I use now, it's called Another One, anotherone.app.
Justin: One of Ian's superpowers is being humble. He also supports global businesses with these agents, so it's not just side projects, you can make real impact on very sizable businesses.
Ian: Yeah, I have clients it builds excellent stuff for in a fraction of the time. It's a lot of fun, I use it as often as I can. When there are days I only use it for an hour, I kind of feel like it was a wasted day.
Justin: Someone new asked, how long would this take me to learn, and how could I put it in small terms so I could understand and make my own apps and bots?
Ian: Honestly, talking to your bot about whatever you want to build is the best way. Follow that curiosity, say, oh, I like that, tell me more, and follow the thread, it can teach you everything. I've never coded apps on my own, I've tried, but now I've produced a lot of apps with the help of OpenClaw. Talk to your bot, follow the curiosity, it'll teach you, it's like a guide.
Justin: Check out this SEO report, fairly robust for a first pass: fixing 404s, tightening meta descriptions, updating copy, keyword landing pages, robots.txt, the sitemap exists.
Ian: With a first pass like this, you can then say, specifically, how do I get found by LLMs, or how do I rank first on Google for this term, and it drills down. It's essentially a 500-IQ best friend who has all the time in the world and never gets tired of your questions. (Jean-Jacques: how do you set up the AI strategy, UI/UX bot you presented before? That's essentially what we just set up. Mine, I just prompted it to make a brand kit, you say, I'm making a brand kit, put it here, and then refine it, keep talking to it and it'll refine the page live wherever you have it posted.)
Ian: (on Lovable) We did Lovable on our last call, Alec did, he uses it and gets a lot of great stuff out of it. I used it for a while, but I'm a designer, so I prefer to have my hands on that stuff.
Justin: Lovable is really nice for one-shotting a landing page, if you need something quick and don't care about too much control. Ian, being a classical designer, it's nice to work with something like Figma MCP with Claude where you have more control over the dials. But if you just want something clean and quick, Lovable is great. Alec built a website with it in about 10 minutes and it looks great.
Ian: (Darian: is OpenClaw the first step to make a bot?) Yeah, OpenClaw with some kind of brain, in this case ChatGPT 5.5. Actually, what we did here is a VPS, a server on the internet, not on Justin's computer, so it wouldn't directly create the code for an app, you'd have it push to GitHub. But if you install OpenClaw on your computer, maybe we do that next time or record a video, it can open Xcode or Android Studio and create the code, then push it straight to your phone if you're on the same network. You'd need TestFlight from Apple. (Jean Rebel: if I have an app idea with detail, would I just feed it in and ask it to create the app? I'd start by brainstorming, put everything into the bot and say, brainstorm this with me, help me create a build spec. Don't let it start building yet, think through the whole thing, and you can be blunt, like, I want to make this much money, what do I need to do. Then once you're happy with the plan, have it build it, or give it to Claude.)
Justin: This is cool, I just asked it if it has all the context it needs to be an LLM SEO expert, or if I should upload help docs. It says it has enough context. Where is it getting all this context from?
Ian: It's using the ChatGPT brain, which is trained on well over a trillion, maybe several trillion, parameters. Parameters are bits of information. It's been trained on so much, that's why it can answer any question it knows, including LLM SEO. If it doesn't have training on something specific, like when I asked Claude about OpenClaw, it said, I don't know what that is, so I sent it the website and it went and learned, it basically trained itself, this OpenClaw, not the big brain. And if you give it specific documents, it'll say, this is important to Justin, I'll pay more attention to this.
Justin: Ian, we've got to hop in three minutes. Will you drop your email in the chat for anyone who wants to follow up? Shout out to all the repeat community members, Peggy, Kristin, John, Jude, Jean-Jacques, Debbie, Darian, Ashwin, Harlan, Jamie, Jennifer, Kevin, Olga, Patricia, Swaty, Toby, Yazin. Thank you all.
Ian: (Jamie: how is this different from an LLM app, and why do responses seem slower?) They do seem slower right now, I'm not sure why, it's not always like that. But these are unhinged, no guardrails, no handcuffs. It can create websites, charts, graphs. You can talk to it in the car or on the train and come home and there's a whole web platform. Whereas with an LLM app, you sit there and copy the code or upload it, it's a little more clunky, and this one's live. I use it for a lot of things, like, hey, look up this person, did they sign up, are they having any problems. But it's also scary, you've got to be careful. It's way more autonomous.
Justin: Thank you, guys, this was a lot of fun, a technical session. What do you think we should do next, another hackathon?
Ian: I liked the AI improv we did, that was fun.
Justin: For those who didn't join, within an hour we built an entire company from scratch using AI, Ian did the branding and design, Alec did the website, and I did the go-to-market strategy. It shows what's capable in under an hour. The big idea Ian and I have been talking about is that there are a lot of gatekeepers of AI who keep the secrets to themselves so they can profit from it. Our ethos is to share everything we know and be transparent with our community. Ian and I are available to steer you in the right direction and send resources. We're building a big resource and tool hub on the website, with recordings, show notes, and FAQs. Just let us know how we can help, and that's what AI for non-technical people is all about. Hope this was valuable. Thank you for joining.