Transcription
We send a message, and he went to work. That's it, I'm creating an epic personal assistant in Ctrl+C, no Ctrl+V, I'm not doing anything at all. That is, if before we needed to spend hours on setup, here it is ready in 5 minutes. N8N MCP server, which, I hope, will allow us to make everyone's life a little easier, contains more than 2,300 ready-made workflows. This dashboard is generated right in Claude, it shows data better than the CIA. [music] Guys, hello everyone. Today I will show you how I created an N8N MCP server, which turned creating workflows in N8N from a hassle into pure pleasure. If you have ever tried to set up a workflow in N8N, you know that feeling, you spent several hours, or maybe days, broke your head, and in the end, it still doesn't work as it should. The main problem is that even no-code platforms require you to be a semi-programmer. N8N is an excellent platform, but creating workflows, thinking through all the logic can turn into a real hassle and sometimes stump even an experienced user. That's why I created the N8N MCP server, which, I hope, will make everyone's life a little easier. Let's briefly go over what MCP is in general, once again. It's essentially like USB-C standard in mobile phones, meaning it's a unified standard for connecting any tools to any AI agents. It allows you to enrich the context of your dialogue with them with real and up-to-date data, and also allows you to perform specific actions in external tools. That is, in the context of N8N, we have all the N8N documentation, ready-made workflows as the dialogue context. Well, and performing actions in the context, ah, N8N is, accordingly, creating workflows, debugging them, adding some new nodes to the workflow, viewing statistics, and so on. To be honest, I've been thinking for a long time that creating workflows in N8N needs to be automated somehow. Change or die. And after a large number of tests, having tried several solutions for a really working tool that could program workflows, I, unfortunately, did not find one. I tried both ready-made N8N MCP servers and simply tried to get Claude without an MCP server to create ready-made workflows. But as a result, almost always, it turned out to be, well, some kind of mess with non-existent nodes or which, in principle, cannot be inserted into N8N and see what it has done there at all. But I realized that I had to take the solution to this problem into my own hands. Having diligently tackled this task, I made literally several versions of this MCP server, I managed to create six different variations of the knowledge base, having written literally kilometers of prompts, after almost a month and a half of my creativity, this project was born. As a result, I got an MCP server that knows everything about N8N, almost every node, every parameter, contains more than 2,300 ready-made workflows, and can also work directly with your N8N instance. That is, it allows products like Claude, for example, to upload ready-made workflows to you directly via API, so you don't even need to do Ctrl+C, Ctrl+V. It can suggest that something is wrong with your existing workflows, add the necessary tools there, and even generate custom dashboards on the fly based on your workflows with execution statistics and data about leads that have passed through this workflow. In general, literally everything that comes to your mind can be displayed on the fly. And the most important thing is that it's all based on your description, essentially in Russian. I created this tool for both automation beginners, who it will help in creating workflows from scratch, and for professionals, who it will help to speed up their work literally by times and will allow them to free up their thoughts for working on really important things, such as prompts, instead of blowing their minds and heads, you know, about what should go where, which node to choose, look at the documentation, and which variable to substitute where. In general, okay, enough talking. Let's go see how it works in practice. Let's start, probably, with the classic. This is with an agent that works in Telegram and acts as a personal assistant. How did we do it before? We go, select the necessary nodes, configure the parameters, go in, read the documentation for this node. And often, yes, people write to me that they've been racking their brains for a week about what should go where, help me, please. And if something breaks, well, brother, good luck with debugging. How about the message when you activate a non-working workflow? Yes. Please resolve outstanding issues before you activate it. Here you go, learn English, that's why all this is still happening. Here we are, let's check. And now look at the magic of my N8N MCP. You just go into Claude and tell the agent: "We need to create an AI agent, a personal assistant, that will work in Telegram." Look, maybe we have something from ready-made workflows already. We send a message and wait a bit. It simply loads the instructions I've already worked out for you, which will help it do everything correctly, it accesses the knowledge base, and starts looking for ready-made examples, then it offers them to you, all you have to do is confirm your choice and check everything in the ready-made workflow, connect the necessary API keys. Then it will even write a brief instruction on how to do it. Great. We have several excellent ready-made solutions. The most suitable one is this one. Let's look at its architecture. Brother, look what we have. This is pure gold. A ready-made assistant with voice input that can. Ta-da-da, ta-da-da. Here's the diagram it drew. Do you want to create this workflow right now? You tell it: "Yes, let's get it into our N8N." We send a message and it went to work. That's it, it's just wow. It's getting additional information from the knowledge base here and continues to work. Super, I now have all the necessary information. I'm creating an epic personal assistant. And here it's already going to make us a ready-made JSON. That is, if before we needed to spend hours on setup, here it's ready in 5 minutes. While I was talking, it's already done. Bam, ready. And the assistant has been successfully created. That is, we don't do any Ctrl+C, no Ctrl+V, nothing at all. Here. Telegram assistant, voice and text. First of all, it immediately writes us articles, but step-by-step setup, what to do, and a description of common problems and solutions. It even put a disclaimer here, my friend. I might mess up with the code, so if something doesn't work, don't panic. Common errors. If nothing helps, write to me. We'll figure it out together. Yes, it sometimes messes up with node placement. That is, we just need to tidy up so that everything is clear and logical for us, and check everything. Fill in our placeholders. We just need to open each node that is highlighted in red and check everything. That is, here we will now connect Google Calendar. It put Gmail here for us. And send a response to Telegram. Well, and let's check how it works now. Execution. And an error popped up, most likely because I have a Russian server, and GPT-4 mini won't work for us. Well, we'll update to OpenRouter now. If something doesn't work out for us, we go back to the chat and tell it: "Ah, listen, change the OpenAI agent node to OpenRouter." Yes, of course, we can change it manually, it will probably be more logical and faster, but let's try, let it change it itself. That is, it gets information about this node from the knowledge base and will replace it now. The only thing is, for the update, yes, it needs to get the workflow, rewrite it completely, so, well, some minor things are easier to change manually, it will be simply faster. But in order to check, we'll find out now how it will change everything for us. Done. Replaced the engine with an open world of models. Well, as it called it. Let's see what it did for us. Let's update now. OpenRouter. Well, yes, node. OpenAI chat models. Everything is okay. Let's choose a model. Let's say, I don't know, like, let's put something like this. Let it be ours. Let's save. So, in Google Calendar, we need to go through it again, fill in all these parameters. Well, that's why I say that some minor edits, especially those that, you know, are certainly easier to do manually. And so. Well, and we start communicating with our agent. What will we write to it? We'll write to it: "Ah, hello. Who are you? Go to work. Everything worked correctly. I got a response here. I am your helpful assistant who speaks Russian. I can work with your Gmail, Google Calendar, answer questions. Well, and so on. Let's update something else. What should we do? Ah, ah, let's do this. Ah, listen, add a tool to our agent, ah, Supabase, so that it gets, ah, from the users table, ah, a list of all rows and returns it when the user requests it. We send it. Let it think here for now, and we'll check. It rewrote the prompt for us. Well, yes, it's as simple as possible, of course, but it didn't even take it from a ready-made example that was there, but immediately, in principle, it made it in Russian, even here. Well, with variables, it did mess up here, of course. That is, it put the date incorrectly, as far as I understand. What is the date today? Well, yes, the time variable needs to be corrected. It did it a little incorrectly. Or possibly just update it. So, let's see what's there for Supabase. It's still updating the workflow. That is, it searched for information about it. Add OpenRouter credentials there and so on. Well, we'll wait then, when it finishes. Yes, there is one more important point, that the size of the dialogue, ah, still has a certain limit, and it's better to have one or two, well, three operations, because, ah, in our case, yes, it just didn't have enough dialogue size, and we'll have to start a new one. That is, I, in principle, just took the message from that chat, copied it, well, and put the workflow ID into it. Let it update everything for us. Magnificent. I have successfully added the Supabase Tool. And here is our Supabase users DB. So, well, we need to connect so that our parameters are loaded. Here is the users table, return all. Let's test it. Output the email of all users from the Supabase users table. It went to Supabase. Now it should give the data. Well, here is the list of our emails from the users table. That is, we literally didn't change a single variable, didn't add a single node manually. The only thing we did was communicate with the agent in the chat. What, it was possible like that? And what usually takes a day of setup for someone, a week, maybe even here, it took, well, how much, well, 5 minutes, 10. If you convert this into money, yes, I have different viewers from Russia, not from Russia, so we'll count in dollars so that everyone understands how much a middle developer costs on average. That's like, well, $50 an hour, let's say 8 hours, that's $400. And with my MCP prompt for 30 seconds, wait 10 minutes. For everything else, there is the Mir card. Let's take something a little more complicated. Let's look for a content factory. Ah, Renat, a special greeting to you. Ah, listen, look, what ready-made workflows for a content factory do we have in our knowledge base, so that there are, ah, many nodes, services, and tools. In general, so that you, for example, give, ah, content, and it is automatically published everywhere. Look, I gave such workflows as vaguely as possible. Let's see what it will find now and what it will do. Excellent, comrade. Let's see what ready-made workflows we have. I suspect there should be some gems. First, it received instructions, ah, ready-made workflows with multi-service auto-publishing. By the way, about searching, I'll tell you briefly. That is, we have vector search here, and Full-text search, in general, it was a separate story to make it search correctly. Brilliant. Let's search for more powerful ones. So, bingo. Now we have a golden collection. Let's look at the most powerful workflow. Ah, there is a social video generator that posts directly to Instagram, Facebook, Meta, recognized in the territory of the Russian Federation. You know yourselves. This is power. Listen, brother, I just showed you an absolutely killer collection. This is the top content factory from the base. The king of content factories. Block Automation Platform. We need to publish something. Auto-clip for YouTube. Social Media Agent. Let's try a deadly move. We have as many as 38 nodes here, so that it has enough memory. Let's get the first one for us. Without changes, only. And let's make sure you can handle it. Ah, don't make sticky nodes, leave the ones that are there. Uh, and in the example, when there are many nodes, yes, the model might simply not have enough, and the output window, and it, well, simply cannot create such a large workflow. By default, it is now set so that it makes sticky nodes for you, so that it is clear to everyone what goes where, what data is needed, and so on. And if you don't need it, or it's a large workflow, then it's best to tell it separately not to do it, and not to specify the details in sticky nodes. Well, and accordingly, it went to build a workflow for us. We'll also wait for it to finish. A content factory probably sounds like a project that takes a week, or even two, maybe for an experienced automator, and for a beginner, I don't know, a month, two. Well, let's see what it creates for us now. I'll try again, brother. These large workflows, yes, I show them as they are. In principle, it sometimes has problems, because, well, here, you know, there are many factors on which it depends, even, in fact, on how much limit we have left within our subscription limit for token output. This is a separate topic for discussion, how neural networks contrive, but they have no other way, because in fact, these subscription prices are all free. What to do in such cases when it cannot create a ready-made JSON? That is, some large workflow. You can, of course, ask it to break it down into pieces. This is, well, this is, you know, a pain. It's easier to do this: that is, here it outputs by ID, and we can simply from the output, ah, in principle, in the old-fashioned way, you can say, manually take just this, ah, parameter called workflow JSON. That is, we take all this, highlight it. You see, right? Here's the workflow that it just went crazy writing so much. Up to here, Ctrl+C and just paste it here. Oh, you even see how much it pasted, even the MacBook got tired. Such a thing, of course, yes, it's difficult for it to create here. There's just too much information here. Well, as you can see, even from such situations, you can get out of it, well, you can say, in the old-fashioned way. And it turns out that our base for the content factory is at least ready. Of course, there's probably something to change here, check everything, replace it with more necessary services by hand or also through a dialogue with it. Ah, yes, there might be errors here, you'll have to rack your brains a little, but how much faster is it than doing it all from scratch? What else can be done? It's clear, you can ask it about each node, it will tell you about them, what, where, what parameters, how to do it correctly, and so on. But let me show you a feature. At least, I haven't seen anything like it anywhere else. This is that you can create a dashboard, a control panel, based on your working workflow. Yes, many people have asked me, both in private messages and in comments, I think they wrote, teach us to create dashboards for workflows, like to see what requests, who wrote them. How many came? Statistics. In general, this is how we can generate this dashboard right in Claude. Now, for example, we'll take some workflow. Ah, are we going to search manually? We'll ask it now. Ah, listen, look, I have, uh, a workflow, uh, that's something like Telegram Proxy Bot, what's it called. And we need to make a dashboard for it, so that it displays data on execution statistics in real time. That is, they shouldn't be made up out of thin air, but real ones that are there now. Let's say, the last 50 executions, take them. And let it display the last message, who wrote to the bot with their contact information. And briefly, what the message was about. This is the kind of dashboard we need. Excellent, comrade. Switching to analyst mode, dashboard destroyer. Now we'll find your mysterious Telegram Proxy Bot and make it a dashboard that will show data better than the CIA. Why not? Why not? In general, it started loading instructions here, so as not to stretch the video, I'll edit it so that when everything is ready. So, in general, the system is active. Last update, executions 50, success rate 100%, average response time. Active users for today. Ta-ta-ta. Well, here, you know, messages are already coming in, guys, so I'll blur this. Well, in general, it works and loads data. At the same time, you can customize this workflow, in principle, however you want. You know, add a graph by hours, display some statistics on leads, yes, if it's some kind of agency bot. In general, there are many possibilities here, but yes, it sometimes starts to make up data, so you need to double-check it and give it a kick if it starts to generate unrealistic data. That is, you just tell it to use real data, get it via API, and, in principle, it will do everything. It's clear, yes, that this is still more of a compromise option than a separately developed dashboard, but it's great for quickly looking at something. What did we get in fact? That is, if you are a professional, then you speed up your work literally by times. What used to take you a day or two of setup, now takes an hour. If you are a beginner, then you don't need to rack your brain too much, yes? You'll have to work, you don't need to pay tens and hundreds of thousands of rubles to automatizers. But I repeat, yes, of course, you'll have to figure it out, but the savings in time are huge, and frankly, in effort. The main costs now are only, well, for tokens, or for a subscription. Ah, naturally, I do not recommend working via API, although such a possibility exists. And on average, creating one workflow, like the basic agent I showed, costs from 70 cents. This is if nothing had to be corrected, up to 5-7 dollars depending, yes, and well, on the complexity. Understood. During development, I managed to significantly reduce this expense, because the first versions cost up to 25, even 30 dollars per workflow. I recommend it's easier to get a Claude subscription, of course. Here, I think the choice is obvious. And Sonnet fourth Opus works best of all models. The others, unfortunately, even with the most detailed instructions, are not yet capable of doing this. For example, one of the prompts, and we have at least five of them here, for creating a workflow takes, well, roughly speaking, 800 lines. Yes, that's tough. And even with such instructions, others are not capable of doing this even somewhat qualitatively. Purely technically, my MCP server can be connected to any tool that supports MCP, that is, Claude, Cursor, Codeium. By the way, a very cool thing. I wrote a post in Telegram, you can go and read it if you haven't. In general, to be honest, this project was born solely thanks to the support of subscribers of my private channel. Their support and patience, even at some moments, allowed me to implement such a thing that will be useful to literally everyone who is interested in automation, from beginners and enthusiasts to professional developers. So thank you very much to them for their support and again for their patience. This MCP server is available for free for subscribers of my private Telegram channel. There you will also find a detailed video instruction on how to connect everything. Connecting all this is very simple, literally in one line, where you need to insert your domain and key. And I also explain in detail there how it all works under the hood, what are the difficulties, how best to use it. In addition, you will find all the materials from my videos there. Instructions for installing N8N, instructions for installing Supabase, BaseRow, access to Claude, and there is also a chat where I regularly answer your questions. If you liked the video, then give it a like, not for the algorithms, but for me. And also, please write in the comments what, in your opinion, could be improved in this project. Thank you very much for your attention. Bye.