The argument in one line.
You can build a self-hosted Telegram-connected Claude Code agent with persistent memory and voice control in under an hour using free and cheap tools like Bun, Supabase, and OpenAI embeddings.
Read if. Skip if.
READ IF YOU ARE…
- A developer or technical founder who wants a free, self-hosted Claude Code bot with Telegram integration and persistent memory for personal or early-stage automation tasks.
- Someone comfortable with CLIs, GitHub repos, and environment variable setup who wants to avoid paid AI agent platforms and understand the architecture hands-on.
- A maker experimenting with multi-agent systems and voice-triggered automation who has 1-3 hours to follow a coded walkthrough and deploy locally or on a VPS.
SKIP IF…
- You need a no-code, drag-and-drop interface or have zero command-line experience — this requires Bun, Supabase config, and GitHub repo navigation.
- You're looking for a production-ready, fully hosted solution with customer support — this is a free course that ends with a paid community upsell for the always-on version.
- You need enterprise-grade security, compliance, or multi-org permission management — the transcript focuses on personal automation and single-user setup.
The full version, fast.
This course shows you how to build a self-hosted, voice-callable Claude Code agent that runs through Telegram with persistent memory, semantic search, and proactive check-ins, using only a Claude subscription, Bun, and Supabase. The build follows a three-tier architecture: Telegram bot as the interface, Claude Code as the brain executing tasks, and Supabase storing conversation history plus OpenAI embeddings for meaning-based recall, with intent detection that auto-tags facts and goals. You clone the public GitHub repo, run the automated setup, harden security by treating the bot token and Telegram user ID like passwords, then make it always-on via LaunchD or PM2. Own your infrastructure instead of renting another wrapper SaaS.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:55
01 · Cold open: live Siri demo
Calls GoBot via Siri, bot accepts a complex GitHub task and deploys slides to Pages in real time.
00:55 – 02:17
02 · Course promise + free vs paid scope
Telegram messaging, memory, voice, proactive check-ins all free. VPS, multi-agent, full calling are paid community.
02:17 – 04:07
03 · What is GoBot: live slides
Slides appear live (made by the bot). 3-tier architecture: Telegram bot, Gateway/Mac Relay, Claude Code. Multi-modal, context-aware, specialized agents.
04:07 – 05:24
04 · Prerequisites: Bun + Claude Code
Bun JS runtime (free). Claude Code via Pro ($20/mo) or Max ($100-$200/mo). Repo: 175 stars, 76 forks.
05:24 – 07:22
05 · Install Bun + authenticate Claude Code
Terminal commands for Bun install. Verify with bun version. Claude Code install docs for Mac/Windows; authenticate with subscription link.
07:22 – 09:23
06 · Step 1: Get the code
Clone github.com/godagoo/claude-telegram-relay or download ZIP. Drag folder into Claude Code, say set this up. Automated setup ~3 min.
09:23 – 11:12
07 · Step 2: Create Telegram bot via BotFather
Search BotFather (blue checkmark), /newbot, pick username. Token = password, never show on video.
11:12 – 12:06
08 · Step 3: Get Telegram User ID
Search userinfobot in Telegram. Security: username + user ID together lets anyone hijack your bot.
12:06 – 15:13
09 · Memory setup: Supabase
Conversation history, semantic search, facts + goals detection, intent detection. Supabase free tier 2 projects. Create org, project, enable Data API.
15:13 – 17:14
10 · Supabase MCP + OpenAI embeddings
Install official Supabase MCP server, provide access token. Store OPENAI_API_KEY in edge function secrets. Embeddings cost pennies.
17:14 – 21:15
11 · Test the bot + first Telegram message
Claude Code populates user profile for context. Message bot on Telegram, typing indicator appears. Bot is live.
21:15 – 23:16
12 · Always On: LaunchD / PM2
Claude Code writes LaunchD (Mac) or PM2 (Windows/Linux) config automatically. Close terminal, bot keeps running.
23:16 – 24:36
13 · Q&A: voice calls, proactive AI, live progress
Voice call demo via Siri. Bot recalls personal context. Live progress + control: real-time step reporting, interrupt mid-task.
24:36 – 26:47
14 · Multi-agent + VPS + community CTA
Board of directors: Orchestrator + Research/Content/Finance/Strategy agents in Telegram groups. Hybrid Mac+VPS routing. Fallback chain. Community CTA.
Lines worth screenshotting.
- A Telegram-connected Claude Code agent commandable by Siri voice messages gives you a 24/7 AI that executes tasks on your computer even when you are physically away from it.
- The free GoBot architecture uses Bun (a fast JavaScript runtime) to run a persistent background process that keeps the Telegram-Claude Code bridge alive even when the computer screen is off.
- Supabase provides the persistent memory layer — the bot stores conversation history, project context, and decisions in a structured database with semantic search.
- VPS deployment (optional) makes the agent truly 24/7 — tasks initiated via Telegram continue executing on the server even after shutting down the local machine.
- The three-tier architecture is Telegram bot (input layer) → Mac relay (routing layer) → Claude Code (execution layer), with smart routing deciding which agent handles each request.
- Triggering Claude Code via Siri ('Call GoBot') converts a voice prompt into a full agentic task execution — Siri becomes a voice interface for a local AI agent, not just a phone call.
- A live demo where Claude Code builds a presentation from a GitHub repo investigation, deploys it to GitHub Pages, and sends back the URL validates the complete pipeline in one take.
- The GitHub repo for GoBot has 76+ forks at time of recording — the open-source structure means the community extends the bot's capabilities faster than a single developer could.
- Multi-agent routing inside GoBot assigns tasks to specialized agents (writing, coding, research) based on the nature of the request rather than sending everything to a general-purpose session.
- Claude Pro ($20/month) is sufficient to run the basic GoBot; Max ($200/month) is what the creator uses for production-level 24/7 agentic work at full capability.
- The automated setup script in the GoBot repository handles Bun installation, Claude Code connection, and Supabase configuration — the installation sequence is: clone, run setup, authenticate.
- The analogy that Claude Code is personal infrastructure with AI — able to execute strategy and take physical actions on the computer — reframes the tool from a code assistant to an operating system.
The free course IS the product.
Autonomee playbook
Goda does not tease the good stuff. She gives it all away, calibrated to stop exactly where you need to pay.
- Lead with working proof, not slides. Siri call in first 3 seconds beats any intro.
- Build the free course to the exact edge of where your paid offer begins. Not less. Not more.
- Reframe your tool for non-developers explicitly. 'Not just for developers' is a line Joe needs for JoeFlow.
- Hybrid local/VPS routing is the right architecture for JoeFlow engine. Steal this model.
- The fallback chain Anthropic to OpenRouter to local is three lines of config. Do it.
- MCP-first setup: let Claude Code configure its own infrastructure. This is the install experience for JoeFlow Batch templates.
- 'Stop renting intelligence. Own it.' Six-word positioning Joe already owns philosophically but does not say this cleanly.
Terms worth knowing.
- Claude Code
- Anthropic's command-line agent that runs locally and can read files, execute commands, and interact with services using a Claude subscription instead of per-call API billing.
- Bun
- A fast JavaScript runtime used here to keep the Telegram bot process running in the background on the local machine.
- Telegram Bot
- An automated account on the Telegram messaging app that can send and receive messages programmatically, controlled by a bot token issued through Telegram.
- BotFather
- The official Telegram account used to create and configure new bots, generate their tokens, and manage their settings.
- Bot Token
- A secret string issued by BotFather that authenticates a program as a specific Telegram bot; anyone who has it can fully control the bot.
- Telegram User ID
- A numeric identifier unique to each Telegram account, used by bots to recognize and authorize specific users.
- Supabase
- An open-source backend platform built on Postgres that provides hosted databases, authentication, storage, and edge functions, often used as a self-hostable alternative to Firebase.
- Edge Functions
- Small serverless functions that run on Supabase's distributed network, used here to handle tasks like generating embeddings close to the user.
- Semantic Search
- A search technique that matches the meaning of text rather than exact keywords by comparing numerical vector representations of phrases.
- Embeddings
- Numeric vector representations of text or images that capture meaning, enabling similarity comparisons and semantic search.
- MCP Server
- A Model Context Protocol server that exposes external tools and data sources to an AI agent so it can call them as native actions.
- Anon Key
- A Supabase public API key safe to expose in client apps, granting limited access governed by row-level security policies.
- Access Token
- A private credential used to authenticate programmatic access to a service account, here Supabase, with the same power as a password.
- OpenAI API Key
- A secret credential that authorizes a program to call OpenAI's models, billed against the account's prepaid balance.
- Intent Detection
- Classifying an incoming message into a category like fact, goal, or question so the system can decide how to store or respond to it.
- Proactive Check-ins
- Scheduled or contextual messages the agent sends on its own to follow up on stored goals or routines rather than waiting to be prompted.
- VPS
- A virtual private server — a rented always-on cloud machine used to host services that need to keep running when the user's local computer is off.
- launchd
- macOS's built-in service manager that starts and supervises background processes so they keep running across reboots.
- PM2
- A Node.js process manager that runs scripts as background services on Windows or Linux, restarting them if they crash.
- Claude Agent SDK
- Anthropic's developer kit for building autonomous Claude-powered agents that run on a server and bill per API call instead of via subscription.
Things they pointed at.
Lines you could clip.
02:47
“Claude Code is not just for developers. It's like your personal infrastructure with AI that is able to connect all the things, both execute on strategy, but also have hands to do things.”
02:10
“Stop renting intelligence. Own it.”
25:55
“This is your AI second brain. This is your knowledge. This is your tools. You don't want to stop working.”
04:41
“I personally been paying $200 to Anthropic forever and that's been my one of the best investments that I ever made.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
analogystory
00:00Hi, Siri. Call GoBot.
00:07Hey, Goda. Hi. I see you've been busy with filming today.
00:11How's the free Claude bot alternative setup video coming along?
00:15We are filming a course today. I have a big task for you. I want you to go to my autonomy organization on GitHub, investigate everything about Go bot features and setup.
00:28You can also look at the WordPress, the course for full setup. And I want you to create a presentation using my autonomy branding, deploy it to GitHub, turn it into pages, and send me a link to the presentation.
00:43Got it. That's a big one. So you want me to investigate the GoBot features and set up across your autonomy key?
00:48Okay. Good luck. Okay.
00:50So I figured how to trigger my CloudBot replica
00:55with the phone, so Siri is finally useful. And now I can watch it step by step executing the task. And this runs twenty four seven now.
01:04So even if I shut down my computer, it keeps going. And after the last video, hundreds if probably not a thousand people asked me how to do it, how to build Clawbot replica inside Clawd code yourself.
01:20So this is going to be the mini course that I promised. I've been working hard on trying to make it as useful as possible. After this video and mini course, you are going to able to communicate to your clot code using your telegram like a messaging app, and you will have memory.
01:36You can use send voice messages, and you will have proactive check ins. If you are interested in the full features like calling and actually going and doing big task and having twenty four seven hosted on VPS. You can check out all those features for the full version.
01:52And if you want to personalize and fully set it up because it's a little bit more complex, you can join our community. Let's dive in now.
02:00I built my new website the business, and it's called Autonomy. I'm very proud of it. And I can also fully edit my website just by making a call, giving it a task, it goes and edits.
02:12That's the whole other video that I'm planning to make. You can navigate to resources. This is the knowledge base and here you go.
02:21Build your own Clobot replica inside ClotCode. So what you're going to build is an AI agents you can call, command, and control.
02:29You can totally get started, and it's free. And I know that people are already building on top of that.
02:35First thing to address that this works on Mac, Windows, or Linux. So it doesn't matter what operating system you have. We are using Cloud Code.
02:42The number one thing for you to understand that Cloud Code is not just for developers. It's like your personal infrastructure with AI that is able to connect all the things, both execute on strategy, but also have hands to do things.
02:57Oh my god. Okay. So, the slides, it just literally popped up.
03:02I'm not cutting this part. Okay? So, we're going to look at the slides together.
03:06So, this is the full version. Okay. What is GoBot?
03:11It's always available multimodal context aware specialized agents. So this is also I'm going to talk about I implemented multi agent. I've like been working with multi agent framework, so I had to do it.
03:23Three tier architecture, Telegram bot, it goes to the so this is the full version. Okay?
03:28Gateway, Mac Relay, I don't want to spend too much time. Smart routing, key features, voice integration, multi user ready.
03:37I don't know. Like, if this is not crazy to you, it is absolutely crazy for me that I can just call, tell it what to do, and it opens on my computer fully done.
03:49And I need to confirm it to push it to Pages so I can publicly share it. Okay? Oh my god.
03:54And I got a link. I literally got the link to the same slides. I need to optimize the slides to be mobile friendly.
04:00Back to the course. Okay? Now what you need?
04:03Two things. We need Bun. It is very fast JavaScript runtime.
04:07So, basically, this allows us to run while your computer so it's basically like running in the background. Like, imagine when you play from Spotify music, no matter what you do, its music is playing. So it's kind of the same principle.
04:19This is completely for free. And then clot code. So for clot code, you can get clot code with clot subscription.
04:27You can even get clot code with clot pro subscription. Monthly, it's around $20. If you want real power, you go with a max plan.
04:36And max plan has two tiers, around $100 and $200. So I personally been paying $200 to Anthropic forever for a moment that they had that option in Cloud Code, and that's been my one of the best investments that I ever made.
04:55So I hope they never increase that price, please. The next thing, the code. So there is a public GitHub repo that I put together, and you can see that 175 people started.
05:06There's 76 people that forked it, so they are making their own versions. You can read exactly about text.
05:12So if you are familiar how to run different commands, just go clone it, and it has automated setup. So now how do you install BUN and ClotCode?
05:21So I already mentioned you need subscription from ClotCode. But for installing BUN, you need to run this command in your terminal. To open terminal on at least Mac, you go into Spotlight and type terminal.
05:35Boom. This is your terminal look. So this is what we are going to use.
05:38This is what developers are very familiar. And to be honest, it's kind of like a full circle because when computer started, everything was in terminal. Now we are back in terminal.
05:47So it's a little bit funny. You would run this command first, and then you need to close your terminal.
05:53So you just copy. Let's open it again. You copy, paste it, and click enter.
05:59Once you have bun installed, you also want to double check if everything went well. So you can go BUN version. So let's do it together.
06:07Again, if I press this, I have 1.35. If everything installed, this is what you will see.
06:14Now installing ClotCode. There are official installation documents. Depending if you're on Mac or Windows, the installation command will look different.
06:22These are the commands. You just copy, paste it in your terminal, and it does the setup. In my case, I already have it installed.
06:29Once you have clot installed, the next thing you are going to do, again, open your terminal and type clot. Let's make it bigger. Just clot.
06:37Just as a name. You press enter. And if you first time set it up, it's going to ask you to authenticate with your subscription.
06:45So, again, you click the link, you select your email that you have a subscription, it generates like a special link, you click on that, and boom, you are inside. It's asking me about accessing certain workspace, so I'm saying, yes, I trust this folder. You are inside clot code, and you have a, like, lovely icon, so it's not as scary.
07:05It's really just like a chat interface. It just looks a bit different, uh, but you type and you press enter, basically.
07:12Now we are going to connect Telegram to our clot code. Step one is get the code. There is two ways to get the code.
07:19You can go to this repo, click on the green button, and download ZIP. And this is if you want to start, but you don't have GitHub account, you can just do that.
07:30However, I highly highly highly recommend if I would not say it's a must to have GitHub account, so you can actually also do version control on your setup. If you have GitHub account and you actually clone this repo, when they push new updates, which I just did two days ago, you can update your whole setup with the new feature that I added or the community added or if there is any bugs.
07:54So having a GitHub account is really really important. If you don't have GitHub account and you just downloaded the file, you go to your download folder and you have this zip file. Unzip it.
08:07And all you can do and this is applies to also images because ClotCode can also analyze images, files, like literally anything you can feed it. You just drag this folder, boom. And then you just say, I want to set this up.
08:22In ClotCode, once you press enter and send it, you actually are seeing what is happening. If you just freshly set up clot code, it might throw a lot of, like, can I do this? Can I do this?
08:33Those permissions that you are seeing, it's not related to the setup that you just got, but it's like a clot code setup. You can change those permissions and loosen up. This is like a gray area, and you have to be really smart about what permissions you give to go on autopilot.
08:48Let's get this setup. This is a Telegram bot powered by clot code. You message it, blah blah blah.
08:54So this is you get how it runs. Bind and clot code are both installed. Phase one estimates three minutes.
09:00Like, I asked my community members. They said, like, it every guide is running in two minutes. It's going to tell you what to do, but also what it needs, and you just paste.
09:08It. Step two, create bot in Telegram. Very simple.
09:11So first of all, you need to download Telegram. Telegram runs both on your phone, on a tablet, on a desktop. Literally, you will be able to access this no matter where you are.
09:20So you go to Telegram. In a search, type bot father.
09:25This is what you're going to see. There's going to be, like, a phishing accounts. Make sure that this one with the blue check mark, open it, click create new bot.
09:34It's asking, like, what's the bot name? So put something like Julie bot. Here here is important.
09:40Treat your username private. Because if you share your username, anybody can find it on Telegram just like we found bot favour. In a sense, we anybody can message you.
09:50I will tell you about security measures, how to prevent that anybody can message you and extract information from your Telegram bot. But even with usernames, we're kind of private.
10:00But we said Julie, let's do Julie one two three four five underscore bot. You create a bot.
10:07Click on your bot, and this is oh my god. I'm so happy that they implemented this.
10:12This is the token that we need. Token is private. Never show it to anyone.
10:17Just copy this, paste it in your automated setup. So I do copy, and here you would say my bot token is next thing that we need is Telegram user ID.
10:29So I'm going to show you how to get it, but before we do that, I want to make it very, very clear. Treat this as your password. I saw so many tutorials on YouTube from people setting up Clotbots, and, basically, they're showing on the video their user ID.
10:46If I have your username and your Telegram user ID, I can spin up an agent in a second and basically message your bot passing in this Telegram user ID and extracting everything that they want. Anyway, go to user info bot.
11:03In the search, type this one. Once again, please click on the green one and not all the fishy ones. So this is the official.
11:13Now I'm not going to click on it because it immediately exposes my Telegram user ID. But, basically, once you click on it and click start, it tells your ID a string of numbers.
11:25Go back to your terminal and you paste the user ID. I'm not going to follow along here with automated setup because I think it's really straightforward. And if you have any issues, you can submit them in a GitHub repo.
11:39But I'm going to explain you what's happening and how to set up the rest. And at this point, you would be able to already message it.
11:48Okay? Now memory. Because if you just send a message and you say, like, hey.
11:52My name is Goda. And just say, hi, Goda. And you're like, what's my name?
11:56There is no memory. Right? We have to set up memory.
11:59That's kind of like a basics that we need. Many different ways how you can do this. I personally use super base.
12:06We are next week having a workshop on comparison between super base and convex. But you want a database that is secure, is hosted in different regions because of data protection.
12:16Once we set up memory, which we are going to do in a second, you're going to have conversation history, also semantic search. Your bot not only pulls the whole messages, but pulls the meaning. So if you are like, hey.
12:28What was this kind of, like, invite that they got last week? Okay?
12:34So this is not something very specific, but there is bits of pieces of information that matches multiple outputs. I also implemented facts and goals. There is intent detection.
12:45If you say something and it's a fact, it's going to detect that this is a fact. Like, for example, my daughter's name is x. It's going to add it to its memory so you don't have to repeat it again and again.
12:56Calls, many people do it differently. You can do task, like, whatever it is. But if you say, like, oh, I want to do this next week, that's a goal detected.
13:04It's going to add. So in proactive check ins, it's going to ask you, like, hey. How is that thing going?
13:09You wanted to do this at that date. Did you finish it? This is, a absolutely, like, a basic basic setup, but I think it already is kind of pretty cool experience.
13:19So where is your memories and data live? Like, we are going to use Superbase for this. So this is the platform, of alternative of Google's Firebase.
13:28Once you create an account, you create your organization. Inside an organization on the free tier, you are able to have two projects. So, actually, I'm going to just delete this project so I can show you.
13:41To create your own, you click to create new project. You need to give it a name, database password. Do create that.
13:49We never know what happens and save it somewhere. Just it's a good practice. Now the region.
13:54Usually, you select the region because this is where your servers live. There are two ways to consider this. If you would ever set it up this for businesses, it's very important where data is stored and all the policies about it, but you're setting this up probably for yourself.
14:10You need to select where you are based because then it's the closest server to use when there's this latency. You always want to be as close to the server that you selected. And you need to enable this because we are going to use data API.
14:22And then you just create new project. That's it. Once you have your project, it's going to be empty, but we are going to let AI to set up the whole database for us.
14:31Edge functions, the tables, SQL is like everything is already preloaded for you. Of course, keep in mind, this is like very basic version. Uh, there is a lot to be desired from this, which you can find in the full version and breakdown, but it's good to start.
14:46It's good to start. You have this functional bot. We are leveraging edge functions.
14:50If you want to learn about that, I put it in writing here. Next thing, in set the automated step, it's going to ask you for your project URL and add on public key.
15:01Go to the top bar where it says connect. In different connections, you can see API keys. From here, you need your project URL and you need your Anon key.
15:11Copy both and give it to your setup. For setting up the whole database, we are going to use Superbase MCP server. In a automated setup, it's going to prompt you to get it officially from Superbase.
15:23That's very important. Use only official one. And install Superbase MCP server.
15:29So there is the full guide how to do it, and automated setup is going to walk you through. Once you have Superbase MCP, it's going to ask you for your account token.
15:40If you want to find access token, you go to account preferences, and here you have preferences and access token. Click generate new token.
15:49Give it a name. For expiry, you can start with thirty days, but, honestly, you can set it to never as well.
15:56But, again, this is like a password to your database. Right? So treat it like one.
16:00Julie, I like, I I I don't know why it's Julie. Generate token. Boom.
16:05So this is what you will see as the access token. Again, copy and give it to automated setup. It's not necessary, but you can set up semantic search.
16:15You're going to have messages. And for messages, we have this embeddings.
16:20So this is basically just a string of numbers and the coordinates of the word meaning for setting up semantic search. It's going to install edge functions, but it's not going to work unless you give it API key. We are going to use OpenAI embeddings.
16:35You can use any other embeddings. To create OpenAI API account, you go to platform.openai.com. Once you create account, if you don't have any, you are going to see something like this.
16:46It looks today like this, maybe tomorrow a bit different. You will have default project. I have multiple different projects.
16:53You can create a dedicated project, Julie. This allows you once you have a project, are more organized and you can track your API spendings. To enable API, you have to set up billing.
17:04And I think if you're new, you're going to get, like, a $5 credit, maybe changed.
17:10But, usually, for new users, they were giving some credits. Once you set up your project, we will need API keys. You don't have any billing enabled, it won't allow you to set up API keys.
17:21Go to settings, billing, and attach a credit card and just add some balance. So I now use OpenAI, I think, only for embeddings for the project, and it's insanely cheap.
17:34Embeddings are super cheap. If you have $510, it's going to last you forever.
17:39Once you added some credit, go back to dashboard, API keys, and, again, create new secret key. You give it a name. We say Julie.
17:47Create secret key. This is what you need to copy. Now important, we're going to place this API key in super base because this is just for super base.
17:55Once you copy and click done, it's going to disappear. You won't be able to copy it again. You will have to create new key.
18:02So copy this, maybe save it in some document. We need to go to edge functions. I have many different edge functions, but for edge functions to have functionality to create embeddings, those need secret.
18:16Add a secret, but here is super super important. You have to give it not some random name, but you have to give it exactly this. OpenAI underscore API underscore key in capital letters.
18:28Paste it like this. And I already have one API key right here saved, so I'm not going to add another one because then this is a variable.
18:38And if there are two, the same variables, that doesn't really work. Regarding cost, super base free tier is insanely general as you get 500 megabytes of storage, and we are storing only text.
18:50In a full setup, we are storing also images. So for example, you can send an image. It detects intent.
18:56Considering that intent, it goes looks at the image. But also, we are storing image and creating embedding. So I don't know.
19:03Like, after a month, you can say, like, hey. I send you this image of camera. What model was that camera?
19:10It's going to be able to tell you because we are storing all, so images become part of our memory. But to start Superbase is free. OpenAI embeddings are pennies.
19:20This is the moment when you are testing your bot. So at this point, you should be done with your setup.
19:26Now it's going to ask you some things about your profile so that basically it's not like dumb and it has a little bit of context information about you. Again, if you're already using Clot code and you have your MD file with information about you, that context It's just an additional thing, and this is for people just getting started.
19:44If you want to start messaging your bot, you click on this username, Julie. Click messages.
19:50To actually start, click big blue start button. Boom.
19:54You are inside. You just say hi, and let's say, for example, here we say thank you. This is awesome.
20:02At the top, you're going to see typing. Well, this is my full setup. I think I implemented typing also in there.
20:10If I haven't, I'm going to add it because it's pretty cool. So you can actually see that something is happening. Right?
20:16Okay. So now this right now, it runs only while your terminal is running.
20:21So we need to make sure that it's running all the time. So this is why we go on always on. And when I say all the time, you have to understand that this runs while your computer runs because it's on your machine.
20:33In a full version, we set up a hybrid model where for when your computer is on, we are using Mac with your subscription, which is fixed cost. If you wanna go and your computer is off or in your backpack, you have VPS, exactly the same function.
20:48We are using Cloud Agent SDK, but more on that in the full setup, and we can teach you, and we can help you set it up inside Commute. Now if you are on Mac, you're going to use LaunchHD. If you're on Windows, it's called PM two.
21:04However, don't need to worry. It doesn't matter on what software you you are. You just basically know to meet it set up.
21:10Just say it, make it always on. It will know what to do. Test it if it works when terminal is closed.
21:16You just basically close it, terminate, gone, and then text your Telegram bot, and it should answer you no matter what.
21:23And if for some reason you want to stop it again, go to Cloud Code and say, hey. Stop my Telegram bot. From this point on, no matter when you open Cloud Code, if you say Telegram bot, it will know what you're talking about.
21:36Gosh. I imagine this video is already pretty long, but top questions that I received from the previous video, want to just address. So voice calls, voice transcription, this is additional, like, phone calls.
21:48Also, that your phone calls have context of what happened. If I call now, hi, Siri. Call Go bot.
21:54So we can see if it actually has any memory. Okay? Hey, good.
21:58Hi. Do you remember what we did last?
22:02Yeah. I do. This week, you've been pretty busy.
22:04You were working on filming content, including getting that free Claude bot alternative setup with a mini course done. You also asked me to create that GoBot presentation and deploy it to GitHub pages, which we got live.
22:15Thank you. I think you'd recovered from the norovirus. Why?
22:20Why is it always spilling a tea? Like, one one thing actually some I I saw a comment that the next task on me to make sure that when it is on a call, that it has like a security measures to identify what is personal information and do not reveal that.
22:37But this is I I imagine this is just a problem that I have. Right? Because I just bring it up on video calls.
22:43I call it call to task. I hope it becomes a thing. At the beginning of video, I included this, like, live progress and control.
22:50So I really wanted that I can see what's happening because I was just giving big task and it goes.
22:57And I don't know what's happening. Right? So now it's reporting you what it's doing, and you can interrupt it and have control.
23:04Proactive AI. So this is where I'm really proud about it. So, of course, you have smart chickens.
23:10So they run on a schedule and they are con contextual in the full version. There is, like, frameworks. It decides, should I even check-in?
23:18Should I call? Should I text? What it's about?
23:21What's important? Then you have morning briefings, for example. If you want to set it up, like, I love this executive briefings.
23:28This full version course is in the community because, honestly, like, if you go, like, oh, why don't you open source everything? Yeah. I can.
23:36Same as any 10 JSON files for workflows. Yeah. You can just give them, but the reality is that 99% of people take those and hoard those things and never set it up because they run into some issue.
23:49I also built multi agent systems. This is kind of MVP, the bigger project that I'm working on, which is autonomous entrepreneurs.
23:57So this is the idea is of board of directors. We have orchestrator. I have researcher, content finance, strategy.
24:03This is how it looks. We are using Telegram groups, and then we have topics. We have specialized agents for each topic.
24:10And for example, strategy, it knows whether it's strategy, content, it knows whether it's content. But the general one is orchestrator.
24:18If I send slash command board, it goes and checks with every single specialized agent what is working, what is the memory, their special skills.
24:27Another thing that each agent can invoke, critic mode. So if they see that they need kind of like an edge, they use critic mode and there is the whole kind of smart prompt behind it. Each agent can also go and consult with another agent.
24:41But more on that in the community. And this production development. So many people ask, like, okay.
24:47Can you set it up this for businesses? But this setup with VPS, You can totally make this production development for the business and make it always on.
24:57On VPS, very important, we are not using clot code.
25:01You are paying per API calls and using agent SDK. So this is different.
25:07And many people ask, like, oh, can I still use subscription on VPS? Theoretically, yes.
25:14But OnTraffic is banning accounts. So I would not recommend it.
25:18What I recommend is, like, a hybrid approach, which I use. Everything while my computer is on is going through the subscription. But if my computer is off, it's going through APIs.
25:29We have, like, smart routing. So if it is just a simple high, it goes to high. If it is, like, something a little bit more com complex on it, if it's really complex, multistep process, it uses OPUS.
25:41So this is for cost management and then, of course, fallback AI chain. So this is also for Mac and also VPS that if on traffic goes down, that we have a fallback model because this is your AI second brain.
25:54This is your knowledge. This is your tools. You don't want to kind of, like, stop working.
25:58For fallbacks, there are different ways to do it, but in my case, I'm using fallback as open router to specific models, and second fallback is basically to the local model.
26:09You can also run local models on VPS if you want to. I'm curious what everyone is going to build on top. I already saw that people are building.
26:18Build something that is yours and your infrastructure. Inside of the community, we have a full course explaining multi agents and how to set up VPS and also how to do it securely, proactive AI, ADS, and we have constant building.
26:34So we released the full version yesterday. I already pushed multiple updates and community members giving their ideas, and we are building on top of that.
26:43So I this is really exciting.
The Hook
The bait, then the rug-pull.
Before Goda says a word of explanation, she lifts her phone and calls her AI. The bot answers, accepts a complex task on camera, and deploys a GitHub Pages presentation before the intro is over. This is not a pre-built demo. It is the product running live, and that is the whole argument for the next 27 minutes.
Named ideas worth stealing.
25:30model
Hybrid Local/VPS Routing
- Local Mac with subscription (fixed cost, computer on)
- VPS with Anthropic API (pay-per-use, computer off)
While computer is on, route through Claude Code subscription. When computer is off, route through API on VPS. Best of both worlds.
Steal forJoeFlow engine architecture: local engine while running, cloud fallback when closed
25:50model
Fallback AI Chain
- Primary: Anthropic Claude
- Fallback 1: OpenRouter (specific models)
- Fallback 2: Local model on VPS
If Anthropic goes down, agent reroutes automatically. Never a single point of failure.
Steal forAny production agent system needing resilience without code changes
24:36model
Board of Directors Multi-Agent
- Orchestrator (general routing)
- Research agent
- Content agent
- Finance agent
- Strategy agent
One orchestrator in a Telegram group with topics. Each topic has a specialized agent. Agents can invoke critic mode and consult each other.
Steal forJoeFlow Batch / Chef orchestrator pattern: one master agent routing to specialized workers
15:13concept
MCP-First Setup Automation
Let Claude Code install and configure its own tooling via MCP server rather than hand-holding every SQL step. AI configures AI.
Steal forJoeFlow install experience: let the engine configure its own dependencies on first run
26:36concept
Free Course as Acquisition Layer
Free GitHub repo + tutorial delivers real value up to exactly the edge where paid offer begins. Creates genuine trust and qualified leads.
Steal forJoeFlow free tier + MCN+ positioning: the free version is the acquisition layer, not a teaser
How they asked for the click.
VERBAL ASK
26:36product
“Inside of the community, we have a full course explaining multi agents and how to set up VPS and also how to do it securely, proactive AI, ADS, and we have constant building.”
Soft and earned. Free course genuinely delivers, paid pitch only arrives after full value is established. Shows Skool community page with course thumbnails.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
