Video: Understanding MCP: A Practical Guide for Enterprise AI | Duration: 2700s | Summary: Understanding MCP: A Practical Guide for Enterprise AI | Chapters: Introduction to Dataiku (5.44s), Agent Building Introduction (127.71s), MCP History Overview (216.155s), MCP Protocol Explained (370.85s), MCP Server Types (523.275s), MCP Implementation Tips (753.725s), Implementing MCP Strategy (1060.975s), Security Analysis Process (1382.69s), Jira Ticket Creation (1623.555s), Agent Tool Configuration (1664.39s), MCP Server Reliability (1793.605s), Role-Based Access Controls (1854.095s), MCP vs A2A (1934.165s), Managing MCP Functions (2042.395s), RAG vs MCP (2138.45s), Agent Testing Strategies (2213.945s), Connecting LLM Models (2461.61s)
Transcript for "Understanding MCP: A Practical Guide for Enterprise AI": Alright. Hello, everyone. Thank you so much for joining. Really excited for today's session. We're gonna take you through a practical guide on MCP, specifically within the enterprise. Vincent, if you wanna go to the next slide. Yeah. So I'll just introduce myself. For those of you who who haven't seen me on these webinars before, my name is Chris. I lead our team of GenAI field specialists. I'm joined by Vincent. He's on on my team, and he's gonna take us through a majority of the session. If we can go to the next slide, I'll just run you through at a very high level for those of you who don't know what Dataiku is and what we do. So Dataiku is a universal AI platform, and we have four key pillars that that hold true to everything that we do within within our product. These have remained the same since 2013. That's something that we're really proud of. And so the four pillars are, firstly, your people. We believe at our core that everything that we do within within enterprise analytics and data science and especially agent building, of course, well is all around collaboration. So making sure that everybody can be involved throughout the process is is really, really important. We connect into your data. If you just stay on the other side. Yeah. So we connect into your data no matter where that data lives. So most organizations that I work with have various different data sources, of course, that we need to connect into. Governing this across the entire analytics pipeline. So from data preparation through to machine learning all the way through to, obviously, deployment, making sure that we govern at every stage of that process. And then finally, your technology. So we we're a tech agnostic platform. We believe in tech optionality, and especially within the world that we're living in today where, you know, every week it feels like a new technology is is beating the other, that's really important to make sure that you're you're maintaining that that competitive edge. Next slide, please. Cool. So before I hand off to Vincent here, just describing the actual product itself. So as I mentioned, Dataiku is a universal AI platform built around collaboration. So whether you're a subject matter expert all the way through to data scientists, I have the pleasure to work with a huge variety of of people in different skill sets and across your entire analytics pipeline from data preparation through to model development, traditional model development, of course, into building out agents. And I think specifically in in the in the world of agents, building agents is actually surprisingly easy. Right? Fundamentally, it's an LLM, a prompt, and some tools. I think the hard part is context engineering. And so Dataiku, because of our foundational aspect, we've always allowed organizations to be able to quickly move from raw data all the way through to the right context, and that's paying off when when the organizations that we're working with are now actually developing agents. Because it means that instead of having subject matter experts just building prompts, they can actually work on the right assets that those agents need to to consume to get the right answers for for the people. Vincent, over to you. Thank you, Chris. Hello, everyone. For this session, we're gonna cover main principle about MCP. We're gonna see how it became a main standard currently in the androgenic world and how we can use it practically in in data coup, and we'll give you some tips and practical guides to to use them efficiently. So feel free to ask any questions if you have some, and we'll finish this session by a demo showing you how you can do that in data through using agent hub. So let's start by the beginning, a brief history of MCP. It has been a bit more than one year that MCP has been launched by by Entropic, and it's been a crazy year for MCP. So it has been launched 2024 with the SDK for Python and TypeScript, And they were very in early adoption by Cursor in the 2025. And at the middle of the year, main actor or, like, Google, Microsoft, AWS, OpenAI committed to MCP collaboration and started to integrate that into their agency offer. In October, there were an official MCP registry launches to to move further in the MCP expand the MCP expand. Sorry. And 2025, MCP has been handed over by Entropic to Linux Foundation. So now it's it's monitored and managed by by this foundation until we go further. If we summarize MCP in a couple of metrics, we have around 10,000 published MCPs. This is those are metrics coming from Oncopaedic websites, and it's two months fresh. We have 300 remote MCP. We would see later what is remote MCP. And we have around no. I think we have 100 millions mostly SDK downloads. So the the market is really there. These are demand. These are need. I mean, we will see exactly what what is MCP and why why is it so popular. So what's MCP? MCP is a new protocol that has been launched by Entropic, as I said, and it allows AI, so LLMs, to connect seamlessly to any data source. You can see every everybody call that USB c for connecting AI models in the real world. So, basically, in a if I summarize it, MCP is run on on the MCP host, which can be data true, cloud code, Versus code, or even your laptop. And in this MCP host, you have your MCP clients. The MCP clients, it just a bridge between the MCP server that contains all the functions and all the the the knowledge to call the the to call the the tools. And the bridge between the two are done with JSON RPC two. Basically, this is a a JSON protocol. So everything goes through this MCP server. You get access to a list of functions that you can trigger. So for example, for Salesforce, you would have some function to call your Salesforce database, and you can get access to many many tools that that can be locally or remotely. And the cool thing with MCP is that you don't have to manually define all the connection and explain how it works. You just have to use this fill this MCP ID card. I would summarize it that way. Let's explain in brief words how the MCP works as is description, is input, and output schema. And this is why MCP can be handled very easily by LLMs, knowing exactly what he needs to give and what he needs to get from an MCP. And just to show you why it's it's very important and why it's solving so much problem, before MCP, we had the all the LLMs had to create their own connection to each API. So basically, if if you use Gemini, if you use Mistral AI, if you use any model from GPT or Boost, you had to create those connections. So it can create a lot of work for your dev team. It's inefficient. Over the time, you need to create maintenance, and it's a lot of work. So with MCP, you just have this block, which is a main entry door, and most of the time, it's maintained by the by the cloud services provider. And you don't have you don't have to to to to review that and maintain that across the time. So it's really plug and play, and that's why all people like it. And it's very, very efficient way to get access to fragmented data that are across all your company in in your agent environment. Just a short example to show you why why it's not good to why why MCP is better than than creating API. When you create for example, if you create a Google Drive API connection with your LLMs, you have to create, of course, adaptation for your engineering team depending on the number for LLM you have. You have some edge cases and error that that are not handled properly because sometime, these new version of API and your connection is not up to date, so you need to have a a huge maintenance over the time. So it's it's not efficient on on the long run. So now we have seen why MCP is so popular. Let's see why what are the different kind of CCP you can get access to and what are the pro and cons. We have the local MCP server. Local MCP server is something that you can host and run on your computer, your server, or your cloud instance of Dataiku or any cloud provider. It's something that you can build by yourself, and you can create private and custom MCP. You can find some on the community, on the open source community. And there are plenty of them. There are 10,000, but probably more now we are speaking. So so the main thing to understand here is that you need to deal the the very good thing with that is that you you have plenty of options, and you have a MCP for everything. You can look for MCP for cyber security assessment, or marketing connection with different kind of tools that are very not known. But this is a good thing. And, also, you keep the control over the data infrastructure. So you make sure when you build your own MCP that you give the right access and you give the right data to this MCP. You are sure about how it works and how the data is processed. On the other hand, you have the security on your on your hands because when you are using local MCP from the the the community, you need to be sure that there is no prompt injection, there is no malicious malicious intent in this MCP, so you need to review it. And and this is something that you need to keep in come to con consideration if you if you see that an MCP doesn't have a a lot of preview, for example. Another thing is that a look at MCP, when it's set up, for example, on Dataiku or any cloud provider, you can get access to external tools, but you hack hard code the API key into the local MCP. So there is no dynamic authentication. The thing is, if you create an agent using a local MCP, you will always have the same kind of write and rules, and you cannot when you add new user using these agents, the agent won't be able to endorse your right and roles and your permissions dynamically. So this is the concept of this. On the other hand, the remote MCP, it is almost a total opposite. We have less lesser number of MCP, But those MCP server are hosted on the cloud provider or the service provider. So, basically, the MCP server that we saw before is not on your computer. It's located on the on the server of the cloud provider and data provider that you are connecting to. So it's a good thing because, as I said, you have a dynamic auto So you you don't have to have app code API key, but you you can have the whole protocol, which is better for for your agent. And and also, you have a delegation of infrastructure management. So you don't need to take care of how it will work, if it's some boxes, if it's if it's in a container. Everything is processed externally. So you it's a good thing for for scaling and and not thinking about the infrastructure aspect. The on the other hand, as you need to go outside of your of your server and your environment, there is a higher latency. And and as I said, there is less options. So so yeah. You need to find a trade off between your initiative, but, basically, remote MCP server will be the main option in the coming years and the the best choice. But at the moment, there is not too much options, so local MCP can make can make sense. So let's see that in practice in data queue. How does it work and how do you leverage MCP? First, your IT admin need to define who creates local MCP. So it it is for for both MCP, they will need there will be different kind of price. So he decide who can create local MCP. And for the remote one, he will probably define the connection with the API too. So for example, let's imagine that in your company, you are using Jira, Slack, and Salesforce. Your IT admin would set up a connection, make sure that it works, and use the auto notifications so that every new user will have to just use this connection, existing connection, and don't and don't have to do anything about it. And and then it it give access to some people or group of people to use it. So then user, it's really easy. You just do what you what you want to do in data tool in your project. So you just create a new tool. I will show you later how it works, but you just need to create a new tool, MCP server, and you select existing connection, or you create your own local MCP server just to form. I wanted saw to to show you couple of tips that we realized with Greece and want to to to really make the most of most of this MCP server. So first, to improve your security, you can do many things. And the most important things is that you can leverage from transaction guardrail in Datacom. So when you're using an MCP, there are data that are probably getting outside of your company, and you probably want to make sure that the data that is going outside of your company doesn't contain some PRI data, like, numb phone numbers, I don't know, names and so on. So you can set up this kind of prompt injection to make sure that it's detected and it's blocked. The second thing is that if you use MCP server on critical tools, you can probably want to review what the agent will do with these tools, especially if you are going to update in your CRM some some data. You probably want to make sure that the agent is is doing the right thing, putting the right the right naming and and using it properly. So you can put a human approval before before the agent executive section. Second, can reduce the agent latency. So it's pretty important because when you are using a lot of MCP, sometime it can gets longer, and that's normal. So you have a couple of tricks here to make sure that when you trigger your MCP, the container where where the instance is running is already already start up and is running is and they're running during all the discussion, so there is no startup delay time. You make sure also that the time out of the call are very reduced based on your on your knowledge of of the of the tools. So you don't spend, like, for example, one minute waiting for the tool transfer. This is a kind of trick that really reduce the latency and make sure that the user has a good experience. Third things is the the cost and context. So when when you are using MCP, your agents have access to plenty of tools. So for example, if I take the GitHub GitHub MCP, I think they are 25 tools and which which represents probably $30.30 k tokens. So basically, if you don't pay attention, you connect you connect this MCP to your agents, you get 30 k tokens directly upfront to your agents, which can increase the cost, of course, and degrade the the the performance because you you over overwhelming the agents with not necessary job descriptions. Really important to say the right and really even tools for for your actions. And finally, the leveraging multi agent architecture is very key. Sometimes, you have plenty of MCP, plenty of tools, and you want to give to your agents those 10 tools accessible. But at some point, if you give too much tools, it gets it gets overwhelmed too. So it's good to divide your agent's architecture in sub agents and saying, okay, this agent is really focused on this task and we'll give you the appropriate tools. And this one will be focused on this one, on this task, and I will give him the appropriate tools. And then you have the agent architecture monitoring all the sub agents and making sure that they are all doing the right things. So, yeah, this is our main tools, improving security, reducing latency, managing the cost, and improving the the accuracy with guilty agent. And so, after the session, probably you will want or you ever try to make a a first step in the MCP world. So what we suggest is that you start with a high risk, high high impacts, low risk use case, and you deploy maybe two or three MCP server with a limited user group. And really measure the time saved and the accuracy and the user satisfaction. This is something very tricky. We can do that in in agent hub, analyzing how people interact, how many interaction these with the agents, and how what are the ratings, and and really see if there is an impact or not. And then if this an impact, you can move forward as much more servers and and probably implement a governance with data quality mesh, making sure that people with the right title, the right roles have access to the right MCP tools, and also train user on the on the on the approved use case. And finally, if you move forward, you can really establish corporate guidelines for MCP tools and and really make a regular monitoring on agents to see how it's moving forward, how how the MCP implementation at scale is is is taking in your company. So now I've talked about all those sorry. So we'll see our aspect of MCP. Let me share my screen to show you how it works in in agents of the I will take you I will take a simple example for for this demo. So we'll take the person of Josh. Josh is a SOC analyst at ABC Bank. So SOC is someone that is really monitoring potential threats attack coming on on on your on your website or on your on your platform. And his role is really to analyze the the alerts, making sure that the security risk are cover the appropriately and blocking access if there is some attempt that is not good and potentially dangerous for the company. So this is the kind of person that is really technical and sometimes that can be lost in many logs. So the goal of this agent is really giving him an easy access to to data, to logs, to get a summary of what's going on, on square technical and business question, and really help him to prioritize his next work, the low low complexity tickets. And if something that needs his team to work on, he can log that into Jira. So we we give him an elastic search MCP server. We will leverage as to the Jira MCP server, and we will use a local MCP server that is a a threat analysis. So, basically, it's that analyze the IP from any coming user, and we can define if on the main platform that we need to know the the risk, we we can say that this IP is very dangerous or this one is not dangerous. So it's it's an agent very useful to really define if some specific logs and specific connection, there is some risk. And if so, let's put some information in Jira, and let's analyze what the enterprise policies say about this potential risk. So let's move to agent hub then. So for those who don't know agent hub, agent hub is the platform of Datacru to make sure agents available for all your Datacru user in one single place. You can create some conversation. You can create your own agent, and you can have access also to enterprise agents. So here, we are gonna use a cyber monitoring agent. And if I select it, I get some some recommended questions. So here, I just let's say that Josh or me, we heard that these are some events that's happening since the beginning of the year. Some really weird event that failed, so some connection that failed. And then, as Josh, want to see this by account. So you want to have an aggregated view, and you want to really filter out all the connection that are coming outside of Europe because the company is mainly operating in Europe. So so the the connection outside of Europe would be very suspicious. So I'm sending this request. Because it takes a couple of seconds to answer, I will directly go to this tab to show you how it works because I already had this discussion. So I ask a question, and then the agent directly answer me with this nice table showing me by account name the number of event that failed. So I can see directly with the connection to to Elasticsearch that we have an account name here that have a lot of failed events counts outside of Europe. So it's interesting. I can check what the agent has done here with the details. So, basically, we can see that the agent has been syncing and then called the ES remote MCP to retrieve the information from from my index. Okay? So I see that there's a problem with this specific account. I want to see the details of the events for the account, specifically, skins. I give him this instruction to give me the details. Here, I see all the details. I can see that they are not too much. They are six, and we can see that most of them have a status code bonded to one. But the most interesting thing is that I can see that the the level of the IP network is sometime cloud provider, which makes sense. But sometimes it's no no ice IP, sometime known for exit node. It's it's it's a bit weird. So those seem to be suspicious, and I will ask the agents to make an analysis of the IP that I'm not part of cloud provider network. So the agent will take those IP and trigger them into my and the threat analysis tool to see what's what's going on and if there's a special risk. And we can see that the risk level is very low. These are I'm not sure. Apparently, all the IP are okay, basically, based on the platform. But still, it's coming outside of Europe and there's something for the same clients. So these are some details, and we can add some some view about it. And what which is enter what what is interesting here is that I managed to pull all those data, but I have also I saw the possibility to ask my agent to analyze the result bay compared to the policy of the company. So here, really, my agent is using Rag in comparing the results to the to the to the policy. And I we can see that he's he has some opinion. So, basically, he would suggest to block this one immediately and isolate this IP here, and he can explain his thinking by showing evidence from from the from from the sorry. From from the knowledge base. We can see, for example, here that the sources are here explained, So we we can exactly see the recommended actions, the name of the file that has been used to answer this question. So we can see that you really used some specific file in the policy, and we can understand his thinking process thinking. We can also see what he did. So basically, he called the elastic search. He used the tool that we set it, and we use a security policy tool, also the Rag, to understand what should be done. And the last the last suggestion is to open a Jira ticket to to track it for and maybe for to have a concrete action for for later. So I tell him, okay, please open a Jira ticket and to discuss it for next week with the team and assign it to me. So the agents execute his work, and we can see that here, we I set up an approval. So I didn't want him to open Jira ticket directly. So we can see here that these this request for validation. So he suggested me this title and this description. So it's a bit hard to read, but we'd see later before before I create the the the ticket in Jira. So I approve it. And then it gave me the results, saying, okay. Jira created. It gave me the link, and it gave me a summary of what it did. So as Josh did, it's just been a couple of minutes to really understand what's happened and managed to leverage two interesting tools. And we can click here, for example, to see that we have access directly to our to our Jira tickets with the with the name, with me assigned as an assignee, and with the description of what happened in the logs. Okay. So that's about it for for the demo and to understand how it works. Just to show you a little bit, of how it works in the back end. So it's it's not that complex. You have a couple of things here to to create. You create a rag directly with with all your documents. So it's it's very it's very easy. You just have to put it in a folder and create on the on the lie on the shelf some some embedded embedded documents. And and then you have you define your agent. So my agent is here. We can see that these prompt, and we can see that we have these different tools here that I have set up. So the four tools I have explained before. And if you want to create, for example, a new MCP tool, I go here to agent tool. I create new agent. I look for MCP. Let's say MCP here. I create. And then DataQ ask you if you want to create a new MCP. So you can create a local MCP server. You select the environment, and you create your local MCP server with with the the JSON that is provided by the by the by the creator, or you select your remote MCP. For example, if I click on Jira, I create it, and then it's ready. So you can see that you can releverage it, load you have access to all the tools. You define what tools you want to use. I was explaining before, I don't want to give all the tools, so I prefer to keep only the main ones, the one that are very important for for me. So fetching the data, looking for Jira issue. I think those are the three main one, and I can test I can test it. I can test my my tool directly and and then save it and leave it to my agent. So, yeah, I think that's it. I don't know if some of you have some questions. I would be happy to answer. Yeah. There's quite a few questions. Some really good ones. So I think one of the first ones that we got is and it's a good one. What if MCP server goes down in the case of a single instance? So I'm happy to to actually take this one. So the MCP servers that that you've seen, especially the local ones that you might be might be hosting inside of Dataiku, if you want to have various different tools that your teams might be using and package them up into MCP servers, We have the ability to deploy that on containers. Right? So in a Kubernetes environment, so you've got containerized execution, and therefore, it it scales. And there was a second question on on active to active in AKS, I think. Yeah. We're tech agnostic. Right? So if you want to deploy that in your preferred containerized execution engine, then you can do that, and we support active to active. So, yeah, that's that's answering those. So let's see here. The the Vincent, maybe the this one's for you, but I'm I'm happy to take it as well. So can you use role based access controls, or does everybody get the same output for for the prompt? So, yeah, it's a good question. It depends on the kind of MCP you are using. So if you if you are using a local MCP, as I said, the API key is outcoded. So, basically, if you are calling any kind of tool, you will call it with the same right hand roles, and you you will have the same output no matter who you are. If you are using a remote MCP, let's say that I have set up a remote MCP and you are your new user that's called the agent, the agent will tell you, I don't know you, so you need to authenticate. And this is why it's more dynamic. You authenticate with the OAuth protocol, and then the agent knows the kind of the the the MCP, the the propagates all the rights that you have on the tool. And so it's based on on your rights on the tools. So then it is you have access to one folder out of 10. The agent could be able to to to collect that data about this unique folder. Don't I know if it's answer your question. Let's Yeah. And then there was a question on compare MCP to a to a maturity acceptance pros and cons. I'm happy to take down one as well. I think that the the two protocols are inherently different. They're they're trying to solve essentially two different problems. A two a, as the name implies, right, agent to agent is is trying to solve the challenges of of agents working together. It does take some of the core philosophies from MCP, namely, for example, the the agent card, whereas MCP has the tool card, right, the tool tool ID essentially, to to quickly share why you might select certain agent. So whilst they do share some similarities, they have very different purposes. Right? If you want to connect an agent to a set of tools, then you would use MCP. If you want an agent to be part of a multi agent system that requires cross communication between various different agents, then a two a might be the right right protocol to use that. A two a does bring a lot of complexity with it, so it's it's something to bear in mind. A lot of the use cases that that that I've seen actually make it to production, that would actually be too complex, specifically when you're looking at kind of evaluating these these genetic systems, which is another question which we'll get to here in a second, bringing in a two a can introduce, yeah, a lot of complexity when it comes to evaluation. And so I think, specifically, right now, it's an amazing protocol, and I think we're going to see it grow. But whilst we're focusing on achievable use cases that can actually get tangible ROI, I don't see it see it making it in yet, but I do expect it to be in the future. Whereas MCP you. is real. We've seen real use cases using MCP, making it towards towards production. So, yeah, that's that's very, very useful. We got a question here, Vincent. I think you might have shown it, but maybe if you wanna cover it again. So on the point of manage cost in context, what is a good way to restrict the set of MCP functions available to to the agents, specifically listing tools? Would you do it in config file? How would you do it? You can do it directly in dataXUI. And, yeah, you you don't have to to define it in a in another place. The the thing is, sometimes it can be a bit tricky because let let's say, for example, when I was building this demo, you realize that some function are necessary to call some other functions. So you you need to do do that to add it's in in a interactive way. You but at some point, you need to define what is the most critical function for my agent to run, and and then you just send it in into the UI data group. Yep. Yeah. So you can just turn them on or off. It's it's really nice. One of the questions is, can we add external tools that aren't in Dataiku? Yeah. That's a that's a really good question, and that's fundamentally the the point of of MCPs. I've I've used that that quite a lot. And I think in the demo, this has showed some tools that that wouldn't be natively available, and we're using MCP MCP to do that. Let's see. Ah, there's a good question here on on MCP versus Rag and when you might use them. So it's an interesting one. Vincent, do you want to do you wanna take that one? Yeah. I can start and complete if you think I'm missing some. point, but so RAG is is the best way to analyze semantical information in a vector database. So document, PDF, and any kind of information that is, I would say, knowledge. MCP is is not only accessing data, is also making some action. So I would say that if you want your agents to have some knowledge about something and when I say knowledge, it's not querying some database. It's just understanding a presentation, understanding a PDF, understanding some visual aspect, graph and plots and so on, use Rag. If you want your agent to be able to perform task on third party services and make some query, collect some data, push some data, and so quickly add actions. I would recommend to use MCP, definitely. Yeah. There was another question here around around evaluations and testing strategy. So I'm I'm happy to take that one. So super asked, how do you know if the data is providing or if the data that is providing is accurate, and what is the testing strategy? So when it comes to kind of agent ops, right, and testing and and evaluations, it's it's absolutely critical. And depending on your use case, you might want to over overinvest time, especially if you don't have humans in the loop or humans on the loop, then it's really, really important that you you run a robust testing strategy. So the the way I see testing strategies is there's there's two things, two techniques essentially we can apply, two categories of techniques. The first one is having ground truth available. Right? So that is that is by far and away the most the most accurate way to to test the performance of of an agent. You would have ground truth on the tool trajectory. So a question gets asked. You know you you know what tool trajectory you you would want the agent to do. And a good way to kind of think about this is what would a human do in the in the same situation. If I ask you a question on, hey. Fetch this data from from Jira. What are the things that you would be doing? Right? Okay. I go to Jira. I do I perform a search to find the right account. I then look at that specific segment, and then and then, essentially, you're plotting down the different tools that you expect the agent to do. So you have a ground truth trajectory, and then you have a ground truth answer. Right? So you know exactly what the agent should be responding, and then you can compare that. And DataIQ supports that by providing the agent evaluation function. It's one of my favorite parts of of the platform. It allows you to really quickly start implementing those tests and get performance analysis. So we can essentially say the agent is performing exactly as intended, you know, across all of these different tests at an accuracy of 95%. Right? And then you need to analyze that that remaining 5%. What is the reason that it's not hitting that? And then oftentimes, that's due to due to context engineering, like I said, or prompt optimization that you might employ. The second category is, of course, if you don't have ground truth and yeah. That's usually the case actually because it's it's quite hard to get ground truth. Right? It's a lot of work, a lot of effort to put in, but I think it's worth it. You should definitely put in the effort. Then when you're thinking about your agentic use cases, that should be one of the first things that you think of. It's like how is do I have ground truth for this use case? Is it easily to get the ground truth? If not, you might be going down the wrong wrong route. Right? Because if you can't prove that your agent is going to perform well in production, then it might be a nice POC, but it's very hard to get signed up for it. So that's definitely a thing you should be thinking about. But there is a second technique that we can use, we and can employ this also for continuous monitoring. And that's, of course, having an agent review it. Right? So LLM is a judge essentially where you're essentially tasking a LLM to review the trajectory that an agent took and the answer, and essentially then getting getting certain metrics off of that. It it can be very effective, especially if you use different types of LLMs, and this is what I often recommend. It's can you often see LLMs and, specifically, the companies developing these LLMs as, brains. Right? They're they're trained on certain sets of data. The Gemini models are trained on very different sets of data than the OpenAI ones. Although there's some overlap, Google will always have some proprietary data that that they've gathered from from doing the the search engine work that they've been doing for the last ten years. And so you can use that to essentially start, yeah, start testing your agents by using these different LLMs. I I always recommend do that. You can also have council of models. Right? So you can have multiple different LLMs reviewing a subset of of the tasks that the LLM has done. And, again, all of that we make available in Dataiku. It's a really interesting topic, one of my favorite topics to to discuss. So so great question there. Okay. Let's see here. Oh, there's a there's a question on how do we connect the actual LLMs. Vincent, do you maybe wanna discuss the LLM mesh? Yeah. So in in Datacru, you have the ability maybe I can share my screen. So let me screen yeah. I can do this. So if you go to administration here, then you have the ability to define connection, and then you can create new connection. And so, for example, you can define a new connection in any bigger cloud LLM provider. So if you want to define a new connection, for example, with OpenAI, you create OpenAI here. And then you define a name. You create your open you OpenAI, the key, for example, and then you define the the LLM that I'm not sure what to access to. So this connection will be available for a specific number of people or everyone if you let it to every analyst. And yeah. And that's it. You you just have to have an API API key. Sorry. And and it's set up for for everyone or for a small number of group. And you have many many many option. You have this side, and you have girl, I think. Yeah. I think it's well, okay. You have you have many of them here. So it's a new session and new connection. Yeah. And there was a question on can we connect to local models? And, yeah, we can. And we we've actually done quite a bit with with some of our more heavily regulated customers. We can connect to to your locally hosted LLMs, and we've got a great partnership with NVIDIA as well. So, yeah, we can. I think that was most of the questions here. Where are we on Dataiku Academy training on this topic and the tools built into Dataiku? I expect it to be coming very, very soon if it's not already out yet. So, yeah, expect all of these new features to be part of the the academy courses, so don't worry about that. And then, Daniel, you asked a good question. Can you show how you perform deployments of agents in different environments, prod test? So everything that you saw today, including the consumption, was within development. Dataiku has a really robust strategy for deploying agents into production environments. That's a whole topic in and of itself. We probably need to spend an hour, if not more, on that. So, yeah, it's it's a good one. I'll note down. And I think as another topic, I will note down as well as maybe running a webinar on on agent evaluations, kind of what the different techniques are for it. Frank, I I saw you ask some questions as well. Those are good ones, but probably take a bit more time to to dive into than what we have today. And then Ram just asked, is it a cloud based SaaS, or do we have all functions local in the enterprise? The answer is both. Again, we're very much agnostic to how we deploy Dataiku as well. So if you want to deploy Dataiku, we have Dataiku Cloud, which is software as a service, or you can deploy that on premise or in your cloud stacks, your AWS, your GCP, your Azure. That is there's no problem. And the same for MCPs. Right? So you can deploy your MCP servers locally. You can connect to cloud based ones. It's it's really up to you. Cool. I think that was most of the questions. Thank you, everybody. Those were some really, really good ones. Thank you, Vincent, for for for preparing the presentation. Thank, yeah, we hope to see you in the next one. Thanks very much. you. Thank you, Bye. bye.