Transcription
Hello everyone on the AI Dialog channel. A few days ago, Google introduced a new tool, Gemini CLI – it's a code agent, uh, open-source, uh, powerful, and absolutely free. In this video, we will look at the installation process, explore its capabilities, uh-huh, who, and most importantly, why to use this agent. And most importantly, we will consider three practical tasks, real tasks, and test its usage. We will fix a bug, implement a new feature, and create a project completely from scratch. By the way, if you are interested in creating products using such agents, then follow the link in the description. We have prepared a practical, hands-on intensive for you. Let's start our review with the press release. The press release is dated June 25th, where Google introduced the free, open-source Gemini CLI tool. So, what is it? Google says that a command-line tool for developers is not just a tool, it's practically a home. A home. There. And with the help of a terminal or command line, developers can solve all their work tasks as effectively as possible. And that is precisely why they are presenting their Gemini CLI tool to you, an agent that embodies the full power of their flagship model in the terminal. And, of course, this tool is excellent for solving tasks related to programming, writing code, and everything associated with it. But, as it were, they are also ready to present much, much more. And they say that they have integrated this agent into, you know, the Gemini Code Assist environment, yes, Google also, it turns out, has an integrated development environment. Well, also based on VS Code, of course. So, what's so good about it? And what does free mean? To use Gemini CLI for free, you just need to log in using your account. Accordingly, for regular users and developers, uh-huh, the opportunity is provided to use the flagship Gemini 2.5 Pro model with a huge context window. A million tokens. It will be 100% enough for you. And what does free usage mean? They provide unimaginable, actually, limitations. Ones that hardly any of you will be able to hit. It's 60 requests per minute and 1,000 requests per day absolutely free. There. But if you are a super-efficient developer and these limits are not enough for you, then, of course, you can use your own keys, and then usage beyond that will be charged. They, of course, note that this is still a preview, there, uh-huh, like a preview version, it has just come out, uh-huh, it provides good code understanding, file usage. How could it not? And command execution. Problem-solving. Well, it's as if a tool cannot be called a code agent if it doesn't know how to work with files and execute some commands. These are practically hygiene functions. And a fundamental update to command-line interaction. There. And something else. They provide the ability to use Google Search, they provide rich, in our opinion, capabilities for expansion using the MCP protocol and provided extensions. That is, open source means that any developer can figure out how everything works and write some extension function if they lack something. If you don't know what the MCP protocol is, don't worry. We had a video on this topic on our channel, you can watch it separately. Also, this tool provides rich customization options for prompts and instructions using configuration files and various, uh, specific files for contexts, some limitations, and peculiarities of your project specifically for such agents. And, of course, this tool, as we see here, yes, and probably like most videos you will be able to watch, is an interactive mode of interaction, but they also say that this tool can be used non-interactively, yes, and such non-interactive usage is precisely the sphere of all sorts of automation, CI/CD, or any other tasks where calling this tool with some prompt can be embedded into automation scripts. Open-source, extensible. Written, by the way, in TypeScript for 95% for expansion, it provides the ability to use the MCP protocol, configure prompts in a MD file, and use various configuration parameters. We've talked about the code assistant. We won't be looking at it today. We will only be looking at the console utility. And let's start immediately with the installation process. They suggest we go to the GitHub repository, uh-huh, where, uh, everything is written about what needs to be done. In principle, uh-huh, two installation options are provided. You can use the command to start without installation, but this will probably be suitable for a one-time acquaintance. If you don't want to install this utility on your system, you can just run it and experiment. We want to use it regularly, so we go to the terminal and run the global installation command for this utility. This will allow us to use the Gemini command in the terminal later. This same command is also used for updating this utility. When launched, uh-huh, we see a welcome message, a description of capabilities and usage methods. But, uh-huh, initially, after you install it, you will need to go through the initialization process. This process involves setting up the theme and choosing the authorization method. I have, in principle, already done this, but, uh-huh, it's worth mentioning authorization separately, yes, that at any moment you can re-authorize. Uh-huh, there are several options, which I mentioned. And we are interested in the free option when using a Google account. Here it should be noted that a large number of users with different accounts have had problems with authorization. Authorization, I believe, didn't work with business accounts, some newly created accounts, so, uh, well, and generally various errors occurred, which can be viewed and solutions found for your case. See above on GitHub. We choose to authorize with Google. We select our account. We grant access, we see this window. Okay, we can, we can continue working. Uh-huh, interactive mode. Uh-huh, how much context we have used, what model is being used, how the agent itself prompts us, yes, you can use the help command and see what it is capable of. In fact, there are three groups of commands for the agent. The first, which starts with a slash, and slash is no exception, are commands for managing, uh, the agent itself, yes? We will analyze them all in detail later. Commands that start with an at sign, as in this example, yes, @src, my files, TypeScript, are commands that allow you to work with context, an important component of a code agent's work. And commands that start with an exclamation mark are commands that allow you to work directly with the terminal. And this is also a significant part of a code agent's work. Let's go through them. We will consider all these commands for working with context and executing various utilities with examples. Let's go through the agent management commands. As I said, at the beginning, uh, well, here, for example, 16 commands are presented. In the previous run, there were 15. Uh-huh, literally, I don't know, half an hour ago. Uh-huh, and let's look at the hygiene commands that are not related to the process. Uh-huh, the first thing it suggests doing after installing, uh, Gemini, is to set up the theme. Well, you can, accordingly, choose to change it later. Uh-huh, here are different themes. I like the default dark one. Uh-huh, we choose next, information about which version is currently installed, in what mode you are, how you are authorized. Not an unimportant story for a preview version is the ability to report, uh-huh, some errors, of which, in principle, there are plenty. There, and at the current moment, there are enough. Let's go through, uh-huh, through all the main commands. About shows information about the version. Authorization. You can change the authorization method. Bug. Send error information. Chat is dialogue management. Uh-huh, we will look at it separately. Clear. Clears the screen, dialogue, history. We clear it. Uh-huh, compress. In principle, we see the context window here. Yes, 100%. If during interaction the context fills up gradually, then it can be packed, compressed to itself. A very useful function when you work for a long time. Documentation opens in the browser. Uh-huh, we go, read the documentation on GitHub. You can set it up. Since all work is done in the terminal, for some tasks, for example, changing some files, clarifying, comparing, you can use an external editor. You can configure it, accordingly, choose one. I have Cursor selected, you can choose any other or leave the built-in one. Uh-huh, editor. Help. We looked at it. MCP shows a list of available MCP servers. I don't have any MCP servers installed yet. Accordingly, in such a case, it is suggested to read information about how to configure them. They are all configured through JSON files. Memory is a command that allows you to manage memory. Memory is a very, very important component. It is essentially the basis for such a systematic, correct interaction of a code agent with your project. Privacy information display, exit. Statistics, uh-huh, statistics on token usage, time, we haven't used it yet. Tools display information about tools with descriptions. Read folder, read file, search content, search files, edit, write file, get content by address, document content. A useful function for web search. Read multiple files, execute command, memory management, Google Search. We will definitely look at its usage. So, all commands have been reviewed. So, let's move on to testing. We will be testing on, uh-huh, a web application. Let's start the server. This is the landing page of our AI Driven development intensive. So, and we will work with a bug that is located in the reviews block. Yes, we can see that here, the navigation arrow is not very visible, poorly positioned, there is no handle. Let's ask it to improve it. What do we need for this? We need, of course, to go to the application folder and run Gemini. Let's start by asking it to analyze this project. So, we ask it to look at the structure, language, tools. Accordingly, everything it does, it reports. It read the package.json file, understood what it represents, what framework, what programming language is used, what the application structure is. I hope it also looked at the README file. Well, and probably, it would be good for it to determine the package manager. Well, let's say, it's already better, yes? Let's save the project information. Let's say there is such a section. You can, in principle, always allow editing files, yes, or, by the way, here's an option when you need to apply this change using an external editor. Well, let's, within this session, allow it to edit files all the time. So, it says it updated. Well, let's check if it updated. Yes, the pointer appeared, but it's not aligned vertically. Let's take a screenshot and say, please, multimodal model, update the styles. So, debugging, yes, it thought, so it presented us with a plan for such a micro-change. In principle, it thinks for half a minute. Once more, the plan. Debug. Well, let's apply it. I can't wait to see how this fix ends. So, it updated. Okay. It looks plausible, yes? Made it more prominent and aligned to, uh-huh, to the center. Let's ask it to do something more global, probably. Uh-huh, what do we want? Let's implement a new feature, yes? Uh-huh, let's try to tell it in Russian. Why not? Uh-huh, please add a new section, uh-huh, which will present quotes from leaders of large companies regarding, uh-huh, the requirements for using AI for employees, about the future of employees, about the future use of agents. Look on the internet for companies like Zapier, Amazon, Microsoft, uh-huh, and others who have publicly spoken on this matter. And from these quotes, please create a section. Uh-huh, so, let's add the file about the structure to its context, which we saved, and let it try. So, accordingly, here is an example of using Google Search. We asked it to search for statements. Uh-huh, it went to search for statements. So, it went, it found. So, plan. Let's look at the plan. Create a new component, place it in the components folder, which will be called "Future". It should be similar to the reviews. Uh-huh, so, take the same card. And this section needs to be added to the main page layout. In principle, it sounds logical. Uh-huh, here are the quotes. Uh-huh, it even found logos and will put these logos in the pictures. And I'm eagerly awaiting the result. So. Uh-huh, and by the way, uh-huh, not only Google Search, uh-huh, Google, uh-huh, provides the ability to use several of its multimedia tools through the MCP protocol: image generation, video generation, audio generation. So, in the same way, you can integrate them into your workflows, if, of course, it's possible in your case. Yes, we have encountered all possible examples of using the preview version, yes, and slow responses, and, you know, two errors here during generation. Something is not responding, but it seems like it handled it. So, our context has slightly reduced, in principle. Well, let's, let's see. So, the server is running. Let's see where it added this section. By the way, I didn't check. Uh-huh, and after the reviews, it added this section. What can we say about the reviews? Oh, so. Well, in principle, not bad. It made a title, a subtitle, and presented several quotes from Zapier, yes, reading basic math. As it were, leaders of large companies say that AI is now a basic skill for all employees, and they need to master it. Well, it seems like it handled it, yes? Let's see what we can, for example, do with this dialogue, yes? We can save this dialogue and say that it is so. And let's see. And, accordingly, in the future, we will be able to return to this dialogue. Yes, in principle, we haven't filled up the context too much, yes, but if it were filled up, we could use the compress command. The last, final task of maximum difficulty level, yes, is an existing project, in principle, as it were, not bad, yes? And what will we ask? We will ask the following. I made a small template here. We will ask to create a selling page for developers from this press release, yes. In principle, a landing page that will sell, yes, convince developers to try Gemini CLI, a full-fledged landing page. Let's try. Uh-huh, what did we say here? We gave design requirements, structure requirements, key information. And what didn't we give? We didn't give the address of this landing page, oh, the official press release, but it would probably find it, but just in case, let's add it so it doesn't get confused, and let's see what kind of landing page it can create for us, given the speed, of course, with which it generated something before. And plus, considering that we see, apparently, high load here, we will have to wait some time. Will there be any interactivity? So, and let's, of course, open it. In principle, the option with the terminal, yes? We use the at sign. Oh, not the at sign, but the exclamation mark. And let's open, let's look at the page. Wow. Wow. And your agent in Russian is free, smart. Capabilities, capabilities, yes, the color palette, of course, leaves much to be desired. Uh-huh, developers say, well, a typical landing page, in a rather strict color style, but, but it handled it, it's not so bad. Uh-huh, start for free. And where did it send us? GitHub, in principle, it found it correctly. No, incorrectly, yes. Well, shall we consider that it also handled this task? Not perfectly. Of course, it could be much better. So, let's summarize briefly, what specifically I liked. Yes, indeed, the basic set of functions is implemented. A sufficient number of commands, cool search integration features, the ability to manage context usage and its expansion, a preview version, errors occur, slow speed, sometimes some functions don't work as you expect. Something is possibly not as convenient to use. Yes, there are a sufficient number of issues and errors open on GitHub, and requests for improvements. It must be said that working with the terminal leaves much to be desired. The endless cycle of fixing, self-correcting errors also has room for improvement. But this is, as it were, a good claim to compete with the main players, Code and OpenAI Codex. And, of course, to get effective, high-quality solutions, as you could see, you need to know how to interact with this agent, correctly manage commands, correctly manage context, correctly build iterative development cycles with feedback. And this is precisely what we will teach you in our intensive. We have gathered and crystallized all our practical experience in using various agents and their use in a repeatable, systematic way, absolutely without code in the Cursor AI tool, more mature, more convenient. And in this intensive, we will teach you to create products, no matter who you are. And whether you are an analyst, a manager, a product manager, or someone who used to write, and now doesn't write, as it were, an entrepreneur, a freelancer, a startup founder, everyone at the intensive will be given a systematic approach, we will go through the entire lifecycle from idea to deployment into the industrial circuit with very high-quality accompanying documents and structure, we will teach and show how to make the process of using agents repeatable, systematic, which yields a high-quality, good result. Uh-huh, and in 4 hours of the intensive, you will get, in principle, a ready-made assistant who, who will solve your task, will learn to create such assistants. Uh-huh, the start is this Friday. Everyone is invited. Uh-huh, if you liked the video and found it useful, then be sure to subscribe, give it a like, leave comments. Thank you all, bye everyone.