‘AI Developments at n8n’ - from the Amsterdam Meetup (April 2025)

Nice. So, welcome. Um, I'm uh winging it also a little bit today, but I do have some stuff that I can show you. Um, yeah. So, I'm JP. I uh I'm engineering manager for the AI team. And for the last uh week and a half, two weeks, we've been working on something to get something out uh that you might have seen this morning because it was released this morning. Uh if you know, you know. It's MCP. And MCP is for model context protocol. Uh who of you already knows what MCP is? A little bit. Okay. So, uh, if you have a an AI like Claude here, um, you sometimes want to use that with some tools, right? And, um, if you use Cloud Desktop, you might want to interface with your file system, maybe your GitHub issues, maybe uh, you want to interface with anything. And um, we didn't have that for uh, for NAN yet. And this is what we've built so that you can integrate uh NAN with your cloud desktop. But also the other way around if other people have built their uh MCP servers, maybe you want to integrate that into uh into NADN uh after I show you a little bit of that. Uh it's really hot of the press. I'll show you something else. Uh and I will give you a sneak peek into what we're working on. And it's all AI related because that's my purview, right? Um so this is

uh not what I wanted to show. This is the MCP server trigger. So if you saw just now uh claw desktop and I want to uh maybe attach a couple of tools that I use in N regularly or I want to attach like a subworkflow. I want to expose that to uh a desktop application or like your IDE like cursor or things like that. You want to build an MCP integration and the MCP server trigger is what allows you to do that. It exposes uh your uh MCP your tools on a URL and you can get that like either the test URL or a production URL and it works similarly to like your standard web hooks. uh you can change the path if you want and uh you can add uh authentication so that not everyone that guesses your path can also uh use it and then you can expose all these tools that you want to claim and I set this MCP server I set it up with claude and I can show you that it says now for those you of you in the back it says six MCP tools available. If I click it, I can see that I have this calculator, hacker news, my color tool, vector store, a workflow, and a Wikipedia API all from this NAD server. And that NAN server is the server that uh that you saw uh here, right? I called it NAND. It could you could rename that to something else. Um so yeah these are all these tools and you can actually see that when I say something like uh can you summarize the top uh articles about nan from hacker news then it will uh start to call that um uh that workflow and it asks you permission right so if you use tools in claude it will ask you permission other clients will not ask you permission so for example the client And I will show you next doesn't normally ask you permission and I think also cursor doesn't ask you permission but I can click allow for this chat or allow once usually I and it will contact my uh my workflow and I can actually see that in the executions uh that it just called the hacker news uh tool right and it will give you all the information about uh about nan from hacker news and

um that's basically the MCP server trigger. What I think people will really use this for is not to expose these individual tools, although you can if you have your credentials set up and you just want to expose a group of them uh to your LLM or your other host basically. But I think most people will eventually use it to expose their workflows, right? If you want to integrate uh something you've built, you want to kick off a workflow from uh from your uh LLM, you can do that now with MCP server trigger. So uh that's really powerful. Um this is the server trigger. And you might notice that it doesn't function like a normal trigger because it doesn't kick off the rest of a workflow, but it's uh it's a special kind of uh trigger that we had to uh kind of work out how that works. And we might add more bells and whistles later. So for example, there are a couple of limitations. This doesn't really work in Q mode. If you use Q mode, that will not work. Um uh but we might have also plans to do prompts, resources, and sampling. How we're going to do that yet, I don't know. But um these are things that it's currently lacking. It's just doing tools while the MCP protocol allows for a lot more. Um yeah, so that's uh that's the server side of things. Any questions so far? What is Q mode? Q mode. So you can run N in like a like with multiple instances. Okay. And um so that you can have failover and things like that. It doesn't really work that with that yet because we currently have the limitation of using serverside events. Uh that's the protocol they use. There is a new protocol but we haven't uh really dug into that yet because all the clients still expect either standard IO or server side events but that requires you to keep open a longer lived connection and if you then send a command to nan it might end up in the other instance and it doesn't know how to route that to the correct uh thing but yeah cool um alsocept files. What do you mean with except files? If you want to send an attachment using Gmail tool, right? So, do you mean attach files from like your tool or from like the LLM? So, for example, if you upload a file to Yes. and then say it to send it using the Gmail, right? Okay. So, so the question is, uh, does it also accept attachments from like cloud desktop if you want to send that for example using Gmail? Um, it's a good question. I don't know yet. Uh, yeah. Uh, we haven't tested that, but knowing tools, binary files and stuff like that are kind of hard. So, if it's like images, that's kind of hard with tools. You might have noticed that in in general in N8 using tools and binary files might be a bit tricky to get right. So um probably not yet but uh we'll definitely keep that in mind as a potential use. If you have a really good use case let us know. And so this is the other side of

MCP and so we'll get to your question uh after this bit. So the other side of MCP is a client, right? So if you have an agent, you might want to integrate an MCP server that somebody else built, right? If somebody else built a very nice NAN MCP server, they might want to give you access to that and you might want to expose that in your agent. Or you might have your own MCP server in your NAND instance and you want to interface that as a as an easy way to attach the same tools to different kinds of agents. you might want to you might do that using for example MCP client uh like that. And so the MCP client is basically a tool like you would normally use a tool in an agent uh but you can configure uh like the endpoint uh that you want and you can configure the authentication and then we you have the option to include or exclude uh tools if you want. So right now it say it is in the all uh mode so that it will use all the tools. Uh but you can also say like I want to select only particular tools. So I want to only want to use hackerne or uh the calculator or things like that. And um uh you can also do uh I want to exclude things. Um, if I do all right now and I start chatting, so I can ask it like, "Can you uh explain the colors of the Amsterdam flag? Look it up on Wikipedia. " Then it will call the Wikipedia tool uh and it will uh give us some answers about the Amsterdam flag. And it will call the Wikipedia tool. And you see that here the query is the Amsterdam flag. and then the tool that is called with the description. That might be important if you want to figure out like okay what does that tool actually do and send to my LLM and then you can see the uh the response that it uh that it returns. So and then it gives you a nice little uh thing and that allows you to do all that from the uh from the AI agent. So that's the client that's the other side of this. Any questions before I continue? No, it's not a question. But I think it's very important to say that using MTPs from external sources are very security risks. Yeah. Everybody needs to know what the source of server is and where it's coming from and what you want to do with it. because surface are external sources that can do a lot within your resources. It's a very good remark.

So so the remark basically is that you don't if you don't know who built your MCP and you don't know what's actually happening with that MCP that built. So what happens is it downloads the descriptions of these tools. it lists all these tools and uh it might inject the descriptions of those tools into the context of these LLMs, right? So, if you uh put some prompt injection stuff in these descriptions, if you don't trust that MCP, they can still steal stuff or they can easily call tools that you don't want to be called and things like that. So, so be aware of what the uh what these MCP uh servers do. Yeah, there was another question. also around the security. How do you handle that? Security, you mean the authorization or the security that was just mentioned? Yeah. So, so authorization and stuff like that we handle with, for example, bearer authentication, but the protocol doesn't really specify how to authenticate MCP calls and stuff like that. So, we took what is now common in the community, and that seems to be the bearer token, but there might be more coming. There's a lot of debate around that. that topic. The other side about the uh prompt injection stuff and that sort of things. We don't know yet exactly how we want to do that. But one way of doing it is checksuming the descriptions of these tools and then comparing it against that and asking the uh the creator of the workflow for example like hey is this now changed? Do you want to update your MCP descriptions or not? What they say in the chat will be the same as the structure that will be changing our database. Ah, so to get your question right, you have an MCP that you might want to hook up to a database and you want to know whether that uses the correct input output format. Yes. Uh that's really hard and that's a prompting issue and also a thing that is really hard with a tool that's noneterministic. Right. So Yeah, what you probably need to do if you want to secure that stuff, you want to probably route that through a subworkflow and then get that subworkflow to analyze basically the the format of that uh that answer. Yeah. everything. Uh I I will we can continue that conversation later. Um because I don't know how much time I have. I don't see Bart right now, but uh I will Pizza Man. Ah, he's getting

Man. Pizza man. Ah, he's getting the pizzas. Okay. So I will continue because then we can uh eat pizza uh quicker. Um this is a really cool little snippet that we did uh and what it was released on Monday. Um this is the think tool. I don't know if you guys have seen uh this little white paper by uh Anthropic. They said like okay you have your uh reasoning models like uh 3. 7 sonnet that has a thinking mode. Uh you have your 01 and your 03 models from uh from OpenAI and things like that deepseek. Um but what if you want to do these kind of reasoning steps without maybe a um a reasoning model and therefore they decided to do the think tool and the think tool is really basic. It's just a tool that has a description and that's it. it doesn't do anything, but the description says something like, "Hey, use this tool uh if you want to think about something. " And that gives the uh the agent a little scratch pad to think about. And um that's really cool. So you can uh you can ask it, for example, a little um uh like a riddle and then you can see that it's actually calling that think tool uh to reason about that riddle and to uh to work through that problem to come up with an answer. And I can show you like here uh um the question was uh something about like Jack is looking at Anne, Anne is looking at George. George Jack is married, George is not and we don't know if Anne is married. Is a married person looking at an unmarried person. And then it starts to analyze okay this is what we know blah blah and it will call that think tool and it will give you a response uh basically that that's echoing that um that input. So, so basically this doesn't do anything except giving it a scratch pad and a little bit of a description to basically do a chain of thought reasoning uh but then with a tool and you can use this with any kind of uh model even like your uh things if they allow you to do tool calling basically. So that's a little snippet that's out there. You can pop it in and see if it improves your uh your workflows and your agent reasoning. Cool. Uh last bit I want to

show you is a really really early sneak peek on what we're working on and this is uh the AI workflow builder and um this is not out anywhere yet but uh this is uh something that uh that we can uh we can show you. Um let me pull up like a little prompt. So the thing I'm going to ask it to build is a workflow that says every morning at 9:00 a. m. get the current weather for Amsterdam. Generate a nice image based on the forecast. send it to JP at nitn. io and if it's going to rain that day, send a Slack message to user JP reminding uh him to get an umbrella and I'm going to ask that to generate that workflow for me. Uh and it's showing you the steps that it's taking. So it's generating a couple of like uh steps for like uh these are the the things that workflow needs to do. Then it's about selecting the right nodes that it want wants to pick out. So now it has selected the workflow nodes. It's composing that and generating a workflow now to do uh just that. And um let's hope it works. Uh sometimes it uh it's still gets stuck. So because it's this is all very new and beta. Um but yeah, so it had has composed the workflows now the nodes now and it's generating the final workflow. So, this is all going to end up in a workflow. And I still have to click a button uh because it's generating a JSON file. I click cleanup. And you can see here that it has a schedule trigger. Uh it's set to 9:00 a. m. Uh then it will get the Amsterdam weather. Uh still have to put in the credentials. Of course, it doesn't know yet how to do that. Um but yeah, that's uh that's what it does. It even uh created a an if uh node. It's creating like a little bit of documentation uh around this workflow. And uh there you have it. This is a very quick sneak peek at what we're building at the moment. I was already interested in is how I can train my own workflows to make the answers to the questions that I give inside of a chat as you already did. You have given the information you want to know the weather. So you get the weather. So yeah. So what this does basically it has access to a database of all the nodes uh that are currently available in it. It queries that from the internal database in NAN what are the tools available and that and get those get added to the uh to the context if they look like they are nodes that like belong to this uh to this workflow. If I have tools that are not known by the uh space. How can I is there a possibility that you can No, this is all working internally right now. So, so uh yeah, this is why it's not built in an N workflow yet, but might be in the future. Yeah. How are we doing? One more one last question. Okay. Very enthusiastic. Yeah. Go ahead. Sorry. Can you repeat it? Yeah. So, not yet. We're working on that as a as an improvement after this. So, so right now it's always generating a single workflow, but after this it will probably uh No, it will definitely be possible to like iterate on your workflow and ask it to add things to your current workflow and things like that, but that will come later. Yeah. Very cool. That's it.