Transcription
I use Cloud more than any other artificial intelligence. But, honestly, even I dismissed Cloud CD for a long time. I thought: "Well, this is purely for programmers. Why would I complicate my life with a console? And how wrong I was?" As soon as I tried it in action, I realized this tool is too powerful to ignore. If you want artificial intelligence to actually build systems and do work, not just generate text for copy-pasting, you need cloud code. Today I will show you how to start working with it, even if you are far from programming. And most importantly, we will create a team of AI agents from scratch that will work autonomously. Let's go. You might ask: "Why do I need Cloud Code if there's regular Cloud in the browser?" Look, in the browser, you usually use Cloud Project. That's cool, but your chats are isolated. You have to act as a data transporter, copying text from one window and pasting it into another, keeping track of the context. In this chain, you waste a lot of time on routine. Cloud Code removes this barrier. It works directly in your file system on your computer. Locally, it sees all project files simultaneously, can run tests itself, find errors, and fix them. Here, Cloud becomes not just a consultant, but your full partner. You set a task, and it selects the tools, manages agents and MCP servers, bringing the work to a real result. But who exactly does it manage? Let's figure it out. So, the anatomy of an agent. For artificial intelligence to work like a pro, it needs three components. All this is configured in a simple Markdown file. Look at what our agent consists of. First is the role. You clearly define who it is, what experience it has, and what it is responsible for. Second, Workflow – this is a step-by-step algorithm of actions. What to do first, what next, and which skills to connect. Third, MCP tools. These are its hands, access to email, Google Sheets, or your database. And today I will show you how to create your own MCP server and connect it to Cloud. For example, we will assemble a team of four agents for a marketing agency, and Cloud CD itself will become their team lead. It will distribute tasks and check results. An important nuance: when you design your agents and their colleagues, give them narrow, non-overlapping tasks. The clearer the role, the fewer errors and disputes between agents. How to access this power? There are three main points, but I recommend one. First, the local terminal CLI. This is the foundation and the most complete experience. Anthropic has made a convenient installer, so even a beginner can handle it. I've put a link to the instructions in my Telegram channel. This is where you will have full control over files and even a visual token counter. Second, the extension for VS Code. It's ideal if you're already coding something; you see the code in one window and Cloud Code in the adjacent one. Look, I enter a couple of commands, and it builds the site structure on the fly without unnecessary clutter and extra permissions. Third, Cloud Desktop. This is a familiar application, convenient for working with MCP tools. And this is what we will use for some tests. Today, it's a hybrid. Computations happen in the Anthropic cloud, but the model has access to your local files. An important point: I work on MacOS, but on Windows, everything is configured similarly. The main thing to remember is that in the browser, you will always be limited to a single tab, while in the terminal or desktop version, you are the master of the entire system. Let's move on to practice. We create a folder for our project. Let's call it Marketing Team. This will be our Workspace, where Cloud will have full access to the files. First rule: immediately put a file with a description of your business or project into the folder. Without this context, the neural network will hallucinate and work blindly. Now, we open this folder in Cloud. I'm using the desktop version. I simply give the application access to the directory. That's it, the connection is established, but don't rush to pay agents immediately. First, let's lay the foundation. We will create a cloud.md file. For the system, this is like the company's charter. CD will read it and understand who it is, what the regulations are here, and how to structure the work. I simply type in the chat: "Create a folder structure for a marketing agency according to the Cloud MD file." And that's it. One command, and it creates all the directories and files itself. Full autonomy. The result. I open all the code. If you still think this is only for programmers, you are mistaken. Today, this is the gold standard for everyone: marketers, scriptwriters, analysts. It's a free all-in-one tool that makes your work 10 times more convenient. Don't ignore it. We hire our first employee, a content strategist. Now we will lay the groundwork, but I warn you, the real meat will start a little later, when we connect MCP servers to it. The agent will stop being just a chatbot. We will give it access to search through Perplexity and generate visuals through the Nanobanna Pro model. This is the foundation for 2026. The agent must be able to analyze and create content end-to-end. Wait for that moment, and I'll show you how it all works. In the meantime, let's go to the terminal. Why specifically? Because only here do we have full control over creating autonomous agent roles. The feature of Cloud Code is slash commands. We press slash and see a list of shortcuts. We need the main one, `agent`. This is an updated command configurator that appeared in recent versions. We select "Create New Agent" and assign it the Project level so that our Starteg knows only about our current marketing agency. We describe the role: "You are a content strategist. Your task is deep research and creation of content structure." Life hack: so you don't struggle, I've already written the ideal system prompt for this agent and put it in my Telegram channel. The link is in the description. Grab it, copy it, and use it. Cloud Sam generates instructions. We select the model. For the strategist, I set Cloud Sonnet 4.5. I'll tell you about model costs and how not to blow your entire budget at the end of the video. Be sure to watch until this point. It will save you a lot of tokens. The agent is created. Now it can be called in the chat via the usual at symbol. And let's also check what's under the hood of the agent. We open Visual Studio Code. Look, in the project folder, a new Markdown file with instructions has appeared. Everything is transparent, no magic, just a clearly written role. But an agent without tools is just a chatterbox. Therefore, we give it a task: "Create working templates." And this is where you will see the power of Cloud Code. It doesn't just write chat text; it physically creates files in your folders. Company plan, brief, strategy. Everything is automatically organized into the Templates folder. No more copy-pasting. Honestly, I'm re-calculating these templates now as I edit the video and realize this is the level of a strong marketing agency. I know people who work without briefs and documents at all, but usually, they remain small and proud because chaos doesn't scale, and here artificial intelligence builds a system with you. Yes, of course, you always need to proofread the generation. Trust, but verify. But personally, I'm 100% satisfied with this result. Take it and work. Agent testing. One. Testing. The agent's TG via the at symbol and we write the task. I'm enabling planning mode. I warn you right away, it's a bit more expensive. Here, a thinking model works, and it spends tokens on generating the plan, on thinking, and then only on implementation. But for complex tasks, this is a must-have. Fewer errors, quality is many times higher. Let's launch. It Googles itself, pulls up our template, and provides an ideal document. Conclusion: the right mode plus templates equals a result that pays for every token spent. Where do I get access to neural networks today? Right now, getting Cloud Pro or keys is quite a task. So that you don't fall for scammers and get banned in a couple of days, I've created a private group on Telegram. There, I share contacts of sellers from whom I buy subscriptions and keys myself. Plus, we figure out together how to bypass blocks and set everything up correctly. Link in the description. If your Perplexity crashed or you have problems with Cloud Code, write to me in a private message on Telegram, and I'll help you figure it out. Moving on, we create the second agent, a presentation specialist. But this time, we'll upgrade it and add Cloud Agent Skills. What are Skills? Skills are ready-made instructions, a kind of extension package that allows Cloud to perform specific tasks. For this agent, we will need official document handling skills from Anthropic. In Cloud Code itself, they are disabled by default. You can, of course, go to GitHub and download them manually, but the easiest and most reliable way is to do everything through the built-in plugin manager in the CLI. Let's do this. We type the command `plugin`. We select the "marketplace" option. We enter the package name `anthropic/skills`. We need to install `document skills` and preferably `example skills`. Pay attention to the screen. Do you see an error? The system says that I already have skills installed. I deliberately left this moment in. If you have the same, don't worry. I just go back to the menu and install only what's missing. For example, the `examples` package. By the way, there are already plenty of ready-made skills for any task online. And if the one you need is missing, you can always create your own. I've already made a separate detailed video about Skills. The link is on the screen now and in the description. If you want to understand in detail what this is, be sure to watch it. Then, we can use a prompt to confirm twice that these skills have been successfully made available within Cloud Code. We return to Cloud Desktop. We launch configuration mode and create our second agent, a presentation specialist. Its task is to take raw data and turn it into ready-made PowerPoint slides. An important nuance. I immediately set the brand's color palette in the settings, so the agent won't just slap slides together but will immediately format them in our corporate style. Is the agent ready? Let's test it. I have a file with performance indicators in my folder. All you need to do is tag the agent and ask it to create a report based on this file. By the way, in the terminal, this is done very simply. You just specify the path to the file, and the ID picks it up itself. This is much faster than uploading something manually to the web version. Look, it immediately activates the document skills we installed. A couple of minutes, and everything is ready. It provided a six-slide presentation. By the way, I didn't specify the number of slides. It decided how many were needed itself. But from experience, if there's a lot of data, it easily creates up to 20 slides. The result is quite workable. The colors are ours, the structure is there, charts, diagrams are in place. Yes, the graphics are basic, but for a business report, it's what's needed. Minimal edits, and you can present it to management. The coolest part is that you don't need to upload files to the chat at all. Just put any reports, tables, or documents into the project folder. Cloud will scan them itself and assemble a presentation for you. And then the most interesting part begins. We will not just insert ready-made pictures but generate unique images through Nanobanna and automatically embed them into the presentation. There are a couple of tricky nuances here that many stumble on, so be sure to watch the video to the end and don't skip. If you miss this part, the connection of all agents, like mine, won't work. Let's move on. Excellent. Our third agent is Data Analytics. And here we move to a new level because we add MCP Model Context Protocol tools. Don't be intimidated by this abbreviation. Officially, it's a protocol that allows Cloud to connect to external tools and get live data. And in simpler terms, it's a USB Type-C in the world of modern neural networks. A universal connector to connect anything to anything. This is where our entire team becomes truly powerful. You can set this up on your computer; it's nothing complicated, but there are two conditions to start. First, make sure you have Node.js installed. Nothing will work without it. And second, it's highly desirable to install Python version 3.11 right away, even if it's not needed specifically now. In the world of artificial intelligence, you can't do without Python. Do it once so you don't get distracted later. Look, usually, you work like this. First, you go to the browser, open Perplexity, type in a query, wait, then copy the answer and attach it to Cloud. You are just a data transporter between tabs here. It's slow and kills all focus. But the Perplexity API works very simply. It's not image, video, or audio generation. It's just data that arrives instantly. So, why be a mailman when you can give Cloud direct access to knowledge? We will create our custom MCP server. Utopic already has many ready-made MCP servers. There are really many of them. And connecting them is literally a one-button operation with authorization. Sometimes it's simple, but I want to show you the whole process from the inside using Perplexity with the Sonar model as an example. And later in the video, we will connect Google's Nanobanna Pro. The process is similar, but there are nuances. Off-camera, I've already bought an API key for Perplexity on Platin Market. I'll leave a link to a trusted seller in my private Telegram group. Everything is there, so go ahead. This is not the only way, but it's the easiest and fastest for testing. I'll show an alternative later. First, I send a prompt to Cloud. I ask it to generate server code. I specify the key and connection logic. Here's an important point. The Perplexity key I bought has its own API interface that works according to the OpenAI standard. This is currently the most popular format. It's used for much more than just ChatGPT. Now, installation. Cloud Code's built-in terminal can't import servers yet, so we open your computer's regular terminal. We navigate to the project folder, in my case, `cd Docs Marketing Team` or your path. We enter the command `cloud mcp add from cloudDesktop`. This command picks up all MCP servers and our Perplexity server and gives agents eyes on the internet. Now they are not limited to their standard Cloud search but can use Perplexity Sonar in real-time. Note: Some other MCP servers may require additional authorization, but this is simple. We return to Cloud Desktop, go to Settings, Developer, and see one MCP server. This means everything is installed. And now, watch carefully. I bought a key to the simplest Sonar model, and in the prompt, out of habit, I wrote "Sonar Reasoning." Naturally, the server gave an error. This is classic. Always check which model you have access to. I simply corrected the model name in the config, and everything worked. This example clearly shows how Cloud helps fix bugs, and it does it itself. So, everything is configured easily. Just substitute the data, and the system understands it. In parallel, I have my personal account open where I monitor token usage. In terms of money, I bought a key for $1, and it was enough for a huge number of requests. The average price of a complex request for the basic Sonar is about 1 ruble. All prompts and settings are already posted on my Telegram channel, link in the description. Take them, grab them, and use them. Now that the server is ready, we repeat the agent creation process as we did before. We configure the role of a data analyst. Its task is to take raw data, for example, an Excel sales report, and turn it into understandable conclusions. We could connect it to Google Sheets or databases. Remember, you can connect anything to anything, but for simplicity, we'll give it access to local files and, of course, specify the same brand color palette for it. Let's test it. We ask the agent to analyze the effectiveness of our content campaign and build a dashboard. Look, it calls the Perplexity MCP server to find current data. It analyzes our file and instantly builds an interactive HTML dashboard. Everything is in our colors. There's the essence, metrics, charts, and trend lines. This is what Cloud is truly good at. Strategic storytelling. This report can be immediately exported to PDF and sent to the team. Raising the stakes. We are creating the fourth agent, an SMM specialist, but just giving it a prompt is for beginners. We will write custom skills for it. This will make it not just another bot but a carrier of our brand's DNA. Our goal is to create the "Branded Visual Style" skill. I'll preface this by saying that even the latest Cloud CD Opus 4.6 or Sonet 4.5 are still not leaders in graphics generation. They write code and text perfectly, so let's get SVG files. Here's the detailed prompt I prepared. Let's create this skill. Look, I'm not just asking to "make it beautiful"; I'm uploading our brand references and giving strict technical requirements. Pay attention to what is written in the instructions. No raster. I explicitly forbid generating PNG or JPEG. I need clean, valid XML SVG code. In terms of style, I set the format to "Clean plus Hand-drawn." The size is strictly 180x1080 so that everything goes directly to social media. I ask Cloud to read the official Canvas Design Skills and, based on them, write our custom skill. It launches the Skills Creator. Look, it analyzes the database, understands my SVG requirements, and generates a new Markdown file for the skill. And the last, but critically important, step. It's not enough to just create a Markdown file with the skill. We need to officially introduce the agent to it. So, I go into the settings. I do this through the convenient Cloud Desktop interface and add this new skill to the configuration. That's it, the system is assembled. The agent sees the tool and is ready to generate content. We tag the agent, "Agent Social Media Specialist," and give it a specific task. Look, I'm asking it to prepare drafts for five Telegram posts on real business topics. Best website tools, a set of software for business with a budget up to $200, automation of invoices and payments. And pay attention to my conversation at the end of the prompt: "Try to do it with quality, so it's not embarrassing." Let's see how it understands this task. The agent starts working, generates texts, calls the "Branded Social Visual" skill, and saves everything in the "Telegram Posts" folder. Let's open the result. The texts are excellent. And what about the pictures? They fully match the brand. Our colors. There are Light and Dark TM versions. This is pure vector that can be edited in Figma. But let's be honest, despite my request to "not be embarrassing," the design still looks very bad, except for the first image, of course. And here's a life hack. We run these files through Nanobanna Pro. I do this within GMI. There, you get 100 Nanobanna Pro generations per day on the Pro plan and almost unlimited regular Nanobanna. We just upload these files in PNG format and their prompts and ask: "Make it beautiful." And here's the result that's actually not embarrassing. The images turned out to be high-quality, meaningful. It looks like a person understands 3D modeling, even though we just fed the file to the neural network. This can be safely published. The connection works well, but there's one nuance. At the beginning of the video, I said that CLC Code is a unified command center. But in fact, we are now jumping between tabs, copying prompts, going to the browser, downloading files. This breaks the entire flow of our workflow. And the beauty of MCP Model Context Protocol is that we are not limited to just text or code. You can connect absolutely anything with an API here. Want video generation? Please, connect CL or Runway. Need music for a video? Connect Sunna. Analyze databases, external services. Cloud can manage all of this from one window. But let's take it one step at a time. Right now, we need to automate graphics so that the agent doesn't just write text but provides a complete content package, organized into folders without your intervention. We will implement this through the AI service "Polza AI." Why this? It's a purely pragmatic solution. Many direct APIs are currently unavailable or require foreign cards. Polza AI provides access to top models, including Canada Banana Pro, and allows payment in rubles from a single balance. For our task, the Cloud + generation connection, this is more than enough. MCP connection, Polza AI, and generation. Let's start the setup. First, we get the API key from Polza AI and insert it into our Config. I didn't write the MCP server code from scratch and asked Perplexity to find the latest Polza AI documentation. I fed it to Cloud and said: "Write me an integration." It understood everything, provided the code, and we connected it. We launch Cloud so that it sees the new tool. Polza AI for Nanobanna is asynchronous. This means we send a request, and the result doesn't arrive instantly. At first, Cloud thought it was frozen and simply timed out. We quickly fixed this by adding a status check to Cloud so that the agent doesn't fall asleep and understands that generation is in progress. You just need to wait. Now it works stably. Now, about quality. For the pictures not to be random, we need a unified style. Remember, we generated it in GMI. I simply ask GMI to formulate this into a clear system prompt. We take this prompt and pass it to our agent in Cloud. The task is simple: "Generate a cover for post number one, using this style, save it in the folder for Telegram posts." We launch the Nanoban Pro model. The process is not instant. The neural network needs time to render, but we just wait. By the way, about money. This comes out to about 15 rubles per generation. I consider it pennies for this level of image. It also generates Russian text perfectly. And here's the result. The file is already on our disk without watermarks, in full resolution, in the exact style we specified. No browsers. Cloud went itself, ordered itself, downloaded itself. Polza has a lot of other models. You can go and familiarize yourself. All documentation is on their website. Study and connect what you need. I've left links to trusted sellers in my private Telegram group. We also discuss settings there, and the prompts and scripts themselves are in my public Telegram channel. Links to everything are in the description under the video. The most exciting part of the video is coming. The agents are ready. The tools are set up, but for them to work as a single mechanism, we need a conductor. Remember the Cloud MD file we created at the very beginning? Now we will turn it into the brain of the entire operation. Key step. I ask Cloud to scan all created agents and update this file. Look what's happening. It writes instructions on who to call and when. Now it knows the triggers. Need analytics? Call agent number three. Need visuals? Agent number four. Need a presentation? Agent number two. It even added a "Multient Workflow" section. This means that agents can pass tasks to each other in a chain without my participation. This is simply amazing. Let's test the system on a large task. The goal is to create a complete marketing campaign for the coffee coworking space "Bein and Feature" from scratch to a ready-made presentation for the client. The entire process will go through four stages, and each subsequent step will use the results of the previous one. Step one: Market research. Cloud starts with web search and uses MCP Perplexity. Step two: Create a Word strategy. Now Cloud takes everything it learned and writes a full strategic document for our pages. The main idea: a coffee shop where they don't kick you out for using a laptop. Tariffs: day - 12 euros, week - 45 euros, month - 149 euros, plus three ready-made texts for Instagram. And here's the first miracle: the document looks professional. Brand colors, beautiful tables, correct fonts. Nothing needs to be done; you can show it to the client immediately. Step three: Draw pictures. AI generation. The next stage is visuals. We generate three pictures: "Capsule of Silence," "Morning Coffee," and "Community." There was a difficulty here. The service works asynchronously. You need to wait. I sent the request. Waited 10-15 seconds. Checked. It took about a minute and a half for three pictures. But the result was worth it. All pictures are in a unified style, in the correct brand colors. Dark blue and beige, without watermarks. Step four: Assemble the PowerPoint presentation. The final stage is packaging everything that artificial intelligence has created. In the end, we have eight slides: cover, description, idea, sample posts, prices, and plan. All pictures are in place, everything is beautifully created and added. What did we get in the end? All documents tell one story. The research says: "We need silence." The strategy answers: here are the tariffs. The pictures show what it looks like. The presentation packages it: no gaps, no manual work between stages. Yes, there were difficulties, waiting for generation, downloading libraries, meticulous checking of indents. But the main thing is that the entire process went automatically; I didn't interfere. In conclusion, friends. The topic of agents is now gaining maximum popularity. The technology has matured. Today we connected databases, documents, search, image generation, but the possibilities are much greater. You can connect Google Drive, Slack, GitHub, databases, the same PostgreSQL, yes, anything. Write in the comments what tools you use in your work and what you would like to connect to Cloud. I will definitely analyze the most interesting ideas in a live format, with all the mistakes and nuances, in new videos. Conclusion. And before the finale, let's talk about the painful part, limits and models. I'll be blunt, Cloud is currently the most honest neural network. There's a shared pool of limits for all models. My strategy for using it in 2026: Haiku. Simple and cheap. For ranking or simple tasks, it performs excellently. Opus 4.6. The king of analytics and finance, powerful, but it devours limits like crazy. Only use it on the Max tariff. Sonnet 4.5 is my choice. It's the golden mean. It's more profitable for me to ask Sonnet 10 questions and clarify details than to burn through the entire daily limit on a couple of Opus questions. But I'll note that Opus is truly powerful and doesn't hallucinate. And the main rule: don't flood the chat with gigabytes of text. Context is not infinite. For large data, you need RAG systems. This is critically important, and I will make a separate video about it soon. It's strange that I haven't done it before. Now about money. The Pro subscription costs $20. I buy it with my own money. And I'll be honest, it's worth the money. Look around. Sam Altman is running around companies. Investors are tired of fairy tales about AGI. The hype [music] is subsiding. ChatGPT has a slowed-down model. The paid version. And Perplexity is a storm there. First, they gave away annual subscriptions, and now they themselves ban users and demand card linking – constant swings. Against this backdrop, Cloud is an island of stability. But why am I on Pro? Because of honesty. Cloud has no hidden tricks. When the limits run out, they don't quietly switch you to a dumb model like ChatGPT or Perplexity do, which drives me crazy. It honestly says: "That's it, the limit is exhausted, pay for API or wait for daily limits." I spend daily limits, but I haven't reached weekly ones yet, so decide for yourself. By the way, regarding payment, don't waste time on virtual cards; Cloud doesn't accept them. There are no promo codes or free trials either. I activated access through a reliable seller in exactly 5 minutes. I left his contact in the private group, link in the description. There, I also help with questions about Perplexity and bans. If you need help, write to me in a private message on Telegram. I try to respond to everyone. There's a lot more interesting content ahead. Thank you for watching the video to the end. Be sure to subscribe to the channel so you don't miss the next episodes. They are already on their way. If you need breakdowns of specific topics, write in the comments. I read everything. We're working. Maybe not as fast as we'd like, but with quality. Important announcement. Very soon, I will launch projects for training neural networks. I've accumulated a lot of material, so I'm preparing something useful. This was Ivan, the channel "Artificial Intelligence in Practice." See you in new videos. Goodbye, everyone.