Mastering AI tools in Cisco Catalyst Center

Oh, right, thank you for joining everyone, good morning, good afternoon, wherever you are. It's, Wednesday 10th of June and.00:00We're about to start the expert insight series webinar on mastering AI tools in catalyst center. Actually, we will see more stuff than in Catalyst Center. We have quite interesting sections for you.00:20To use AI within and outside of Catalyst Center.00:40My name is Alex. I will be the moderator of the session. Today I have the honor to have with me Andres Franco, Lorenzo Dasilva and Alex Connov presenters of the different sections, as well as.00:45Who with me and the rest of course will try to answer any questions you may have on.01:06On the topic. Quick note, what you will hear are best practices, suggestions, recommendations that we have for you. If you want to apply any of those, it's up to you to take the decision, strategic decisions, business needs.01:12It's your responsibility. We are happy to have a call with you to go further, to go deeper, but what you hear and what you see is recommendations and best practices from us.01:31Today's session, mastering AI catalyst center. We have another one next week and this is in Cisco catalyst center expert insights series. Next week we have Cisco workflows and ways to enhance.01:48Cisco Catalyst Center automation. This will be the last session of decision and from September onwards we will be again with you with the revamped series.02:08The logistics, you are muted. If you have any questions, please use the Q and A panel. We won't be able to answer questions in chat, so if you have questions, send them in Q and A and we will try our best to respond. If there are any unanswered questions at the end, we will.02:24Will collect them and respond by email so you will get your answer at some point. Any technical difficulties use the chat and we will try to take care of.02:42This is the agenda for today, but before we start, we have, as you see, we have AI and catalyst center. We have integrations and the APIs, how to use scripts with APIs and the MCP servers and catalyst center. But before we start a quick note on.02:56On Cisco IQ. I don't know if you have heard. We now have the new platform for asset management assessments and more. And the question is, are you on board already? Have you onboarded to Cisco IQ? If not, this is a quick a highlight.03:15But if you like, if not, we are happy to help you.03:35Please send us an email and we will contact you to see what is needed for you to onboard to Cisco IQ. Very simple. You access the site, you create an account and with your contracts, you can start using Cisco IQ, but.03:38This is not our topic today. Our topic is AI and we start with our 1st presenter, Andres for a Cisco IQ, feel free to send emails and contact us and we will take it further. Andres.03:57Welcome. The ball is yours. Yeah, thanks Alex for for that great introductions as always. Hi everyone, I'm Andres Franco. I'm really looking forward to walking you through how AI is transforming the landscape of the Cisco Catalyst Center and how this.04:16AI capabilities can help streamline your network operations. So during this 1st section I will provide an overview of some of the known AI capabilities while also introducing the AI assistant for the catalyst center.04:34Okay, so let's start with the AI and Cisco Catalyst Center and by reviewing the evolution of the network operations. So basically, as you can see, we have come a long way from the manual CLI and SNMP error in.04:50The it is to the GUIs and the network management systems in the 2000s. Some of you may even remember the Cisco network six assistant or maybe the security device manager for routers and firewalls. Then after that.05:10We did move into the software defined networking and programmability. So that's where we did introduce apic EM which evolved to the DNA Center.05:30And more recently to the to the catalyst center, and we could also start talking or programmatically to to our network devices thanks to things like APIs, and using.05:42Tools like Python scripts and ansible playbooks. Additionally in the 2020s we started seeing artificial intelligence for IT operations with AI ops and the assurance and what we call the the predictive analytics, so being.05:57Able to not only.06:17Know what is currently happening in the network but also the trends and insights to predict somehow what will happen later. And finally, we reach the point today, right, where we are moving from simple automation into the realm of agent.06:19AI, which can act, it can use reasoning. It can really break down complex problems into steps and all of these, you know, using interaction through natural language.06:39And finally getting closer to the concept of the self healing network. Alright, so let's also quickly review some of the AI, let's say main concepts. So we have artificial intelligence has the.06:53Umbrella concept that is everything and is related to all the tasks that are performed by machines or computers that would normally require human intelligence. Then we have the machine learning, which is about pattern recognition or learning.07:13Others from data.07:33Then has a subset of machine learning we're also having deep learning which was key for the current state of the AI and is basically using the multi layer neural networks to be able to learn more complex patterns and finally, what we.07:35Called the generative AI, which is AI that produces content and is also the technology that allows us to use the natural language to interact with these conversational assistance. Let's also recall a little bit about the catalyst center.07:55Right has the controller for the campus network and the component that is able to drive network insights, automation, security, and also integrations. So we have different personas in catalyst center, AI OPs, NetOps, secops, and devops.08:14For the scope of our discussion today.08:35The most relevant personas are AI ops basically because let's say AI alone is not able to produce any actionable network insights, it needs to be able to, to collect the network telemetry data.08:38In addition to using the analytic engine and once we have that information then AI OPs will be able to provide AI driven visibility, insights and the troubleshooting for our network.08:56And the persona is more related to the internet APIs the software development kits and the integrations and all of this is allowing us to execute a tasks a lot faster through the catalyst center.09:11So just a quick recap on that. Then it is also very important when we talk about AI in Cisco networks to understand or recognize that this is not.09:28A new experiment for us, basically since 2018 and within the scope of the catalyst center, we have been in a continuous journey to integrating AI capabilities as you can see in this timeline.09:43We started with the foundational AI network analytics, then we include analytics, AI RRM, site analytics and this days now we're having also event analytics and the AI system.10:01So our customers that are using AI enabled features are able to benefit by identifying the issues before their users do. They are also able to speed up the time to resolution and they are also able to increase the network security.10:19And performance. Of course, as with everything we are having some prerequisites that are related to the catalyst center AI ops. I'm leaving this mainly here for your reference, but we need to have our catalyst center up and running our.10:36Or upgraded to.10:56One of the recommended releases, so at the moment we're still having 2379 and 23710 has the recommended releases, but we can also move to 23711 if you want to benefit from the functional stability improvements and also.10:58Is required if you want to start exploring the AI system. So the prerequisites, as you can see are pretty much the same prerequisites that we have for.11:17Assurance and automation is like defining your network hierarchy, adding the devices to the entry, assigning devices to site and so on. And maybe the most important ones for our discussion today is to make sure that we have configured.11:27AI network analytics for AI driven issues and insights and that we have updated the machine reasoning knowledge base. So once you have completed all of this you will be ready for AI OPs in Catalyst Center.11:41Alright, so let's start exploring the different components that we're having regarding AI in catalyst center. I will explore or revisit some of the features, not all of them. Let's start with the machine reasoning engine. So you can basically think of.11:59MRI has the expert inbox capability that we are having by collecting the knowledge from our Cisco experts, security researchers or even our partners, and we're putting all of that together into a knowledge base, a knowledge package.12:19That then runs on the Cisco catalyst center directly, right? So this is also combined with the data that we are collecting from our network devices from from our network infrastructure and thanks to that combination.12:39MRI is able to automatically diagnose and provide the root cause for complex issues like spanning tree loops, power supply failures, interface flapping, and high CPU.12:56And this is one method that allows us to move from ticket to root cause troubleshooting and resolution. So that's the the concept of the MRE. Then we're also having the network tools.13:11Which you can find if you're not going to tools and network reasoner. So basically this allows us to troubleshoot various issues in our network very quickly. In this dashboard you can find the different workflows and those workflows will allow you to see the affected devices for a particular issue.13:27Let's say in the last 24 h, let's say high utilization or as I was mentioning before power supply failures, and you can also see the network performance impact that.13:47Might derive from running any of these workflows in your in your network. One of the tools that I use, well validated tools that I use with customers a lot is the one that is related to assurance telemetry analysis whenever they are having reporting problems with telemetry assurance telemetry data, right? So.14:01Please feel free to to explore these these capabilities if you haven't done that yet.14:20For this to work, a quick note is you must install the make sure that you have the machine reasoning package installed and up to date, ok? Then moving on with the Cisco AI network analytics, and this is more of a solution overview.14:29So again, we have our Cisco data center has the controller for the campus network and we are receiving telemetry data from our network infrastructure, meaning our routers, switches, wireless controllers, and access points, and then we are using a security channel to forward that to the Cisco cloud where we are having our AI network analytics engine that is used.14:44Seen.15:04Things like machine learning, machine reasoning, visual analytics, and also decades of Cisco networking expertise to then be able to discover some patterns in our network, and build this dynamic baseline. And once we have the dynamic baseline, then we are able to define the.15:05Normal behavior for our specific network and detect the anomalies.15:25Based on deviations from from that baseline. We can also get some insights practice exploration of what what is happening in our network, the the trends, let's say that we are having in our network, we can perform a comparative analysis by comparing our network.15:30To the network of similar peers or even by comparing KPIs internally within our own network.15:49For different allocations, buildings. And finally we have the predictive analysis which can help us anticipate and prevent failures, right? So that's that's the overview. If we start to revisit some of the, of these capabilities, starting with.15:56Baselining, then you may recall that this have been explored in.16:17Sessions before, right? That normally when we are not having AI network analytics, the alerts are, let's say, calculated based on going either above or below custom thresholds above a certain maximum and minimum, so each time we go.16:22Above or below the the threshold and alert will be generated and well.16:42The consequence of this is that we may end up having alert overload, right? We may have too many alerts and we will have a hard time on focusing on the, the area problems that are important, right? So with the AI network analytics and the AI driven dynamic baselining.16:48So we define the baseline which is this blue band that you see here at the bottom, and we only trigger alerts when there is a deviation from the baseline, right? So that's the the important concept. This help us to only obtain their relevant anomalies.17:07And help us to to prevent this alert overload that we were having with the, with the custom thresholds, and in the end, you can focus on the issues that really matter for your specific environment. Moving to another capability, we have AIRM which allows to up.17:25The radio frequency environment without any manual work.17:46Or manual configuration, so it basically allows us to practically optimize and fine tune your wireless configuration. So for this to happen, what I, what the feature is doing is analyzing up to 14 days of historical historical telemetry, so we are basing the analysis on things like e.g..17:50Bigger usage time or it can also be persistent interference and then based on that we can make decisions.18:10That are based on long term trends instead of making decisions that are more based on momentary spikes. We are also having this other capability that is related to the ARM, which is the temporal awareness or well you know it has the VC hour feature that allows the.18:19ARM to identify when your network is the most active so we can suppress any disruptive changes from happening during these specific time intervals. What you get in the AI.18:40Our end dashboard is the summary with the frequency performance and the RF coverage and you also get some actionable insights and some additional performance metrics. Okay, so this is really one of the features that allow you to to have a continuous experience.18:58Optimization in an automated fashion.19:18Then we also have network insights as part of the AI network analytics. So this is again another long term trend analysis across access points and buildings that allows us to mitigate network risk in our network based on different KPI.19:22That are related to capacity and throughput, so one example could be as you can see right here on the screen, for a period of four weeks, we can see the increase in the number of clients for a particular access points particular access point or we could also determine.19:43E.g., what is the trend regarding their radio throughput deviation for a particular access point? So this is how you get to see trends happening in your network with long term data, right? Very, very useful to, to know what is happening.20:02We have some other features that I'm not showing here like the the network heat map that also provides different information regarding KPIs for a period of 30 days but continue with the period comparison, this is another feature that allows us to basically.20:24Benchmark our networks?20:42Performance against similar networks and you can do this again by comparing different APIs that are related to radio frequency and throughput. So one example would be RSSI, but this could also be radio resets. It could also be radio throughput, and then you can select to compare.20:44For the all days of the week or for a particular date of the week and have this view of how your network is performing when compared to, to the similar networks, which gives you this additional visibility.21:04There is also possible to perform a comparison within the your own network between different buildings. We are having that capability as well as part of the AI analytics. Okay, so.21:19Just to continue our discussion here, our, our review of the AI capabilities, we have AI driven issue identification. This is what the issues that we are able to to detect as part of that deviation from the dynamic baselining that I was commenting before, right? And there.21:34Root cause analysis. So what we get with this feature is to see who is impacted by the issue.21:54What is the issue about in this case we're talking about failures, right? Then you can also determine things like when and where did the issue happen. So you have the timeframe and the particular location, the building associated to to the issue. Then you have, of course, how.22:02How the issue was triggered or or detected. In this case the KPI that was for which the deviation from the baseline was seen.22:22So here is failures in the, in the onboarding, right? What we're what we're seeing. And then of course you have a couple of additional KPIs that will tell you the why. So these additional KPIs are the most is the correlation.22:33Of other let's say events happening in the network and that would provide us with the most probable cause of the issue that we are evaluating. So this is one perfect example of how AI analytics provide a lot of.22:50Value when it comes to, to visibility and and troubleshooting of our network. But not everything is about issues. We also have events so candy center already had this capability of let's say collect.23:06In events happening in the network either because of configuration changes or.23:26Changes in the state of the network, e.g., interface going down, right? But it was very hard for us as user administrators to perform the the correlation of all those events. That's why event analytics was introduced for both the wire and the wireless network, and it allows us to basically.23:33Expose and correlate trends across multiple event types. So some of those event types are syslog.23:53Reachability and also client and and radio events. So this is making our life easier when having to to correlate high number of events that are detected by the by the catalyst center in our network.24:00If you haven't explored this feature, I do encourage you to give a try. Then another one that I'd like to to revisit here is site analytics. So site analytics allows us to basically.24:17Highlight the sites that are having the worst performance in our network based on different KPIs. So the KPIs are very specific ones like onboarding attempts, onboarding duration, connection and speed, roaming attempts, and so on. The important thing to know here is that you.24:35For each of these KPI you can track.24:54From the site level all the way down to the floor level, even you can reach, you know, see at the access point level what's what's happening, right? And see how the access point or what issues the access point is, is seen and regarding.24:57Problems that are contributing with the poor performance of a particular site.25:17All right. So now that we have, you know, have this quick overview of the AI analytics and also the machine reasoning engine, I just wanted to present you quickly with a quick comparison and this comparison is not to highlight that one capability is better than the other is more to.25:25You.25:45Communicate that these two capabilities work together, they do combine to provide the best value for, when it comes to insights and troubleshooting of our network, right? So when it comes to the data handling in the case of the AI.25:46For analytics. So that data is cloud based with the identified network event data.26:06That is forward to the cloud and it's the process in order for us to obtain the the specific details and in the case of the MRE or the machine reasoning engine, this data handling is based on the knowledge base with the human expertise that we were discussing, right? And then.26:13The primary function that's where we're having the only, the other.26:33Main difference, the AI analytics is more to detect anomalies and obtain trends and insights of what is happening in our network while the machine reasoning engine is more oriented to automating the troubleshooting workflows and receiving the.26:37Remediation steps, ok? Alright, so we have reached maybe what might be the the most interesting part of this 1st section and it has to do with the AI assistant for the Cisco catalyst center. So basically the AI assistant.26:55Is nothing more than this conversational AI interface that you can use to monitor your network, troubleshoot issues, and also search the documentation, right? And all of these you can use with natural language which will make you feel that you're having a conversation with your own network. So.27:15So, we're having is very important to you to know that we're having a common architecture for both cloud and on prem. That's the reason why you also have the AI assistant in the let's say Meraki dashboard, right? And this is available in the case of the catalyst center.27:35In the releases 23711. Initially it was controlled availability, so you needed to to fill a form in order to gain access. These days you can gain access through Cisco DNA center. So basically after you upgrade to the 237.27:5211th release and if you release if you registered to the catalyst, Cisco Catalyst cloud, then you should be able to to enable the AS system, and it's also available natively.28:10In the cathy center 322, ok? So another thing to to highlight here is of course using this this kind of assistant system we're moving into igantic operations little by little so we can obtain information.28:25Very, very quickly in an in an easy way, right? So we are doing this.28:46Faster. There is of course one important thing to know about the supported intent categories, and this is a high level, so you can somehow determine the kind of questions that you can ask to the AI assistance of the domains that we are supporting in this initial phase is.28:51Are basically assurance, switching routing wireless, and we also have monitoring for client devices and the compliance and also the software management with the swim feature, ok?29:10Has with everything, we're having some additional prerequisites for the Cisco catalyst center AS system. The 1st one we have already covered, which is that you need to run a minimum supported release to 3711 or 32X, right? For P one, my understanding is.29:26That it will also be available for 316 when it becomes GA through cisco DNA sense, but maybe we'll wait until that release because GA to to have that information confirmed. And then the other main.29:44Prerequisite is to guarantee that our catalyst center instance is having cloud connectivity. So we do that by navigating to the system settings and finding the Cisco Catalyst cloud section and making sure that we're registering our instance to the Cisco Catalyst cloud.30:02Okay, so this is very very important. Also, if you, let's say if you perform software upgrade to catalyst center to 3711, you may come across this better banner that will allow you to see that you have been provided with early access.30:17To the center AI assistant functionality.30:37And it will also provide you with the information or the link where you can find information about the Cisco AI system offer disclosure. This is still the in the Cisco trust portal. And this is where you can see, you know, if you are concerned about data privacy and what data is collected, this is the the document that.30:41I will provide you with, with those details.31:01Perfect. I hope I'm not going too fast and that you are, you know, engaged and watching our session very actively, maybe asking some questions. At least the time to, to present the demo for the AS system and I will start with the.31:06Initial verification of this tool in catalyst center. So as you can see in this particular network environment, which is more of a lab environment running 23711, then if I use the global search, I cannot get to the catalyst cloud.31:23I just to show you that I have registered my catalyst center instance to the Cisco Catalyst Cloud portal. If you have not done this, then here you will see the, the option at the bottom to, to do the registration, right? Once we have validated that we should be.31:40In a period, I think there is some some kind of period that might range between 24 h and maximum 48 h, then you should see the assistant icon showing in the top bar of the of the catalyst center.31:58This icon is visible from any page within Catalyst Center, so it doesn't matter to which page you navigate, you should be able to spot it. Then we can navigate to system and the software management and check the install applications.32:15And there you should be able to perform a quick search to find that the AI assistant has been deployed, as you can see here on the screen, you can check the the version of the AI system as well. And finally for this verification phase, you can navigate to the system three six.32:30And check the services that you're having in your instance and then you can search for the AI agent. So the AI agent service should show there.32:50Up and running, and that's how you can validate that the assistant is is actually running on your particular implementation, ok? Let's now move to the AI assistant capabilities, right?33:01So in this part I will be opening the AS system, so 1st thing is that we are having by default the docket view of the system, but you can change that to a floating view or even a full screen view.33:16Then we're also having some suggested prompts that you can access directly from the AS system so they can guide you if you don't know what kind of questions you can ask, and then you are basically ready to start asking questions, right? So we're having, as I mentioned, two main categories of questions that you can ask.33:32Ask you can ask questions are related to documentation, documentation questions, and you can ask questions are related to intent intent related questions within the domains that we did present in one of the previous slides, right? So I will start showing you a quick example here on the documentation search.33:52On how to add devices to the catalyst center. Of course, there is some processing time and normally it's very fast, right? And you can see here that I got my response on how to add devices by either using the plug and play.34:11Or maybe the discovery tool and also the manual method. Then you can also move into the intent questions, so I'm asking here for top issues in a particular site. Notice that I didn't get the response in the 1st attempt and the reason was that I was not.34:26It's very specific with the site.34:46Name. So this is very important because in the background, what is happening is that the the AS system is calling the calling center APIs, right? In order to to filter the information. So it's the same concept of garbage in garbage out. You need to make sure that your prompt.34:51Is very specific if you want to get the best response possible. So here I just corrected my question by providing the specific site name as you can see, and then that did allow me to actually get the top issues for, for that particular site, right? So that should show.35:08In a minute, let me just maybe just jump into that. So I'm having a couple of issues. One is related to fabric border note not having internet and this is because as I mentioned, we are not supporting SDA fabric as intent, but it is included in the assurance.35:28Part so whatever assurance is reporting, we will be able to see it. And then we're having some other issues related to to the wireless network. If you are not familiar with the particular issue, you can query the AS system to provide you or to explain that particular issue.35:48So I'm asking here to you about more details about what poor RF in the 5 ghz bands on a floor means and then the AS system will basically be able to provide the with the details of that issue, it will also.36:06Provide me with the impact and their recommended actions, right? Then you can also be as specific as you want. I mean here asking for a particular gigabyte interface on a particular switch. So I'm providing again the specific name of the switch.36:25And here the center was able to determine what was the root cause for that interface to go down. Actually, this was a quick test that I was performing in the lab environment. I did manually shut down the interface and as you can see here, I was provided with the different event types, the timestamp.36:43And the data center was able to to spot that the interface was changed to administratively down by the admin, right? So this root cause was was accurate in this particular example.36:59What else can we do here? You can ask for a particular detail instead of using, let's say the the devices name, the IP address of the device, so I'm asking here for gigabyte 102 details for that same switch but using the the management IP address of that particular device and I'm able to get.37:17Information about the duplex, the speed and the MAC address that is seen on that particular interface.37:36Then I can even ask for the configuration of the interface, so it will basically be running the command runner and will be presenting the the configuration of the particular interface, and here I'm just checking for compliance, right?37:43Since I'm not having too many devices in my network, my question was kind of road, so tell me the compliance status of my network devices and he showed me which devices are compliant and which devices are non compliant. In your case if you're having, you know.37:59Medium to a large number, but I would recommend is to be specific and ask these details based on a particular site so you can get the best result possible. Then we also have this capability of opening a new chat.38:16Right, if you want to to keep like separate threats and the heat here I'm just checking for golden images in my center implementation, so I was able to get the image name and version for the different device families and finally just asking here a quick question about their recommend.38:31Software images for the switches in my network. So you can ask questions that are.38:51Maybe combining two things, right? You are combining here the doc search capability or searching the documentation with the specific of your network, right? So of course depending on how you ask the question here, I was not specific and I'm getting iOS XE 70.38:56Fifteen.five as the recommended list for for my switches. This is what the Cisco TAC will recommend for catalyst nine case switches for most of the catalyst nine case switches. If you go to the Cisco compatibility matrix, you may see a different version as the as the recommended one. So yeah, you need to have.39:16Those kinds of, of considerations into into account.39:36Oh, sorry, sorry for that. Let me just, I think I just keep one important part here. So here you can have those multiple threads and you can delete a particular thread or rename a thread, that's that's possible. If you click in the, in the hamburger icon in the AS system.39:40And just to conclude the demo, I just wanted to show you that at least in this particular version for 23711, we can switch from the default mode of the ASI system to a deep reasoning mode, which is in beta at the moment.39:57And allows us to query for more complex and time intensive reasoning questions. So in my particular example or example that I'm presenting here, I'm just asking the AS systems to summarize all the critical issues and events in the last week. So I'm even providing the time frame which is.40:17Good practice if you want to get better results. And of course this will take a little bit more time than the default mode, but once it's processed, you will get a lot of details so you get to see kind of a report with the critical issues and events for the last seven days executing.40:36Summary is highlighted there.40:56The active issues showing the status, the locations, the affected devices and their timestamps is also presenting the site health overview, right? With the device health scores, a breakdown of the client health is also presented with even some root cause in some, some cases.40:59And the health for health for the different devices or device family that we are having in the, in the inventory. And finally, of course, the recommendation on how to mitigate this issues, so notice here that we have not reached yet the point in which the AS system.41:19And this is very important to understand, right? We'll execute or take actions to to mitigate the issues. We are currently at the point in which we are presented with the with the data and with the recommendations to to mitigate the problems but it's not actively let's say taking an action to to fix the problem, so.41:38This is still in the face in which we are performing or we're having just focus on visibility and and troubleshooting let's say. Of course, if you want you can give some feedback, either positive feedback or if their response that you are getting is not what you're expecting, you can also provide that feedback.41:58By selecting here from these options, maybe the output that you got was incorrect or was not helpful. And then what we are asking you as a customer is to provide with the additional description of what you saw your expected output and the catalyst center version so our our teams can work in.42:18Improving this tool.42:38Even even more, ok? So that's basically what I wanted to present as part of the, of the demo. I hope you have found it interesting. Just to wrap up this session, I also wanted to present, ok, what is the the future, right? And the future is identic ops.42:41So we can think of as you giving AI a goal and then the AI will execute the steps to achieve that particular goal. So Cisco is building agent ops for the enterprise, so.43:00This in general, right? Not only the context of the catalyst center and is doing that by providing these different tools, the Cisco AI system, as you can see here, the Cisco cloud control which was commented a lot during the Cisco life. If you did join or.43:18Or watch the some of the sessions and also the AI canvas. At the moment current center is not having like native integration with Cisco cloud control, but please stay tuned for future announcements on this, ok? And with this I will finally pass the ball to my colleague Lorenzo to continue.43:38With the presentation. Thank you Andresh for presenting this 1st part of our session covering the AI capabilities that are native to Catalyst Center. So before we go to the last session, which is going.43:58Going to be about MCPs and specifically about an MCP server in Catalyst Center. Let's talk a bit about APIs because I think it's important.44:18To talk about APIs before we talk further about the MCP server and actually the AI assistant itself in the background is also using these APIs and while it's thinking sometimes it's also interesting because you can see the type of queries it's making that resemble a different API.44:33I'm not going to be covering extensively extensively this this in this section. Just to be aware that in catalyst center you have more than 1000 API endpoints, so not only can you cover most of the features that will be available through.44:54Was a graphical user interface, but you can also extend this capabilities and you'll NO longer be limited to the functionalities that you have in the graphical user interface. And as you can see here in the left, you have the different.45:13You have the different API categories and also this is a screenshot of catalyst center where you can also check the different APIs and even test those APIs. If you don't have a lot of experience with programmability, you don't have to go directly to your.45:32Terminal or to your preferred testing platform like postman. You can test it directly within Catalyst Center, which I think it's very nice because you can also see what are the required parameters for your different API calls. So let's start 1st.45:52With that, but before, if you want to get started with this also of course for basic queries you can leverage the, the Cisco devnet sandboxes. There's actually one always on sandbox you can leverage.46:12Which is the this one the the catalyst center always on sandbox and you don't have to, you have to log in with our Cisco credentials. You don't have to wait for it to, for it to to initialize it.46:33Still instant, of course, you'll be limited by the capabilities but it's a great way to get started. If you want to, if you don't want to test this in your own environment, I think it's a great way to, to get started. And, the way you would.46:52Access the APIs in in catalyst center is in if you go to platform and then developer toolkit and APIs. I have a video here but I think I can show you directly in the demo here. This is the devnet sandbox lab, you can test it as well. So.47:13We're talking here about version 23710. And if you go, if you would go here to the menu, to the hamburger bar as we call it and then to.47:33Platform and then to developer toolkit, you can see here all of the available APIs divided into the different sections and you can also test the APIs here and you can also download this file. This is going to download your JSON.47:44File with all of the available API endpoints. And if you want to test something, you can hear so let's let's say, let's try to retrieve our network devices which is a need.48:04The API you can call. You have the description here, you have the actual URL endpoint. Also, if you want to test this outside catalyst center, you have the different query parameters. You can also click here and it will redirect you to developer.cisco.com, which is.48:24Your, let's say source of truth if you want to check more in depth API reference and specifications, you can download the open API document pretty much like you can download it directly in in Catalyst Center, and then you can see here also the different, all of the different.48:44Api's available. And if you want to test it directly to Catalyst Center, you can go to any of them. Some of them remember.49:04Will require you to use the device UUID, so you have to fetch the device UUID 1st with that specific get API and then go to to that one. In this one the retrieve network devices you can query all of your devices at the same time. So if I unselect all.49:13Of the parameters and I run this API call, you're going to see all of the devices that I have in my inventory, and then you can see actually the.49:32The the the call the URL here and you can also use this left plane to filter for the query parameters so if you want to only see let's say devices that have that are in a managed state, you can go here to the managed to the state.49:47And just filter if you want to see, and to filter by the host name, you can also do that by the family role. So let's say, ok, I only want to see access devices you can filter it. So.50:07I think this is a good way to get started if you don't have a lot of experience with, with APIs specifically with catalyst center APIs. I think this is a great way to get started. So remember main menu platform, developer toolkit and then you'll be able to see.50:24All of the different APIs and you can also test them.50:44If you want to bring this outside catalyst center, of course you can test it directly in your terminal. Just remember that if you are testing outside Catalyst Center, you will need an authentication token. This is, this recommends you can run.50:50I'm hard coding the credentials here. I'm not using an file because I'm just using the sandbox lab, so I mean the credentials are publicly well known, so that's, that's NO problem. So, you can and the insecurities to this.51:10Comment here is just to ignore and bypass the.51:30Basically the self signed certificate that the lab might use. Just to to tell you, and I will also show you this in another application that if you want to do any API call you'll have to authenticate 1st. So you have to use this post request to get your authentication token.51:34You can also check this here in the the documentation you have to use this API. You'll receive an authentication token and then you actually use the token to perform the API call, and I have here a demo to show you.51:54So I'm just using, I mean in this case.52:14Well, like I was saying, I'm hard coding the credentials of the lab. You shouldn't do this, but for testing purposes, it's fine if you're using the, the sandboxes. So as I was saying, you're going to use this post API. You're going to receive this authentication token and then you can.52:17I'm exporting this authentication token as a variable just to facilitate the the call in the end. And then as you can see here with this.52:37And with this request, I'm calling the same API that the retrieve network devices API directly in my terminal using curl if you have it installed. But as you will see, the output will be very, very informative, right? So it's not very consumable like.52:48Say human readable and as you will see, it's the exactly the same information as I was showing you in the catalyst center graphical user interface, but not very let's say user readable of course you can use some there are some ways.53:08In which you can present it in a nicer way but as you can see here, it's not. So you have other ways in which you can consume the APIs. There's a very known tool which is called postman. And there's.53:28Official Cisco postman collections. So basically you will download this collection and you will import it to postman. In this case i'm using Bruno, which is a similar application, but you can also use the postman collection in Bruno. So you you will go to developer.cisco.com, you will download.53:46The postman collection and you would import it in.54:04You would import it into as and as you can see here, this collection and the view would be similar to what you would see in the the catalyst center user interface and then you will export this collection and then you would import it here and actually.54:09I I can show you. So I have imported it here, and you will see something here like.54:29Similar to what you see, you have the different categories. Like I was saying, if you're testing this outside of the catalyst center, you have to authenticate, so you'll have to use the post API 1st to to to get the token. So in this case you can use.54:38An environment to to to set your your URL in this case that not has a sandbox and then you have to authenticate using username and password. You'll get the token which is going to be saved. It's going to be saved. Then.54:58And then you can actually use this this token to retrieve your your network devices as you will see here and you cannot test all sorts of of APIs here. Okay, so this is another way and of course that.55:17Format is a bit better postman also and Bruno also can generate code snippets if you want to test this thing curl and so on and so forth.55:38So coming back to our presentation, this is all nice. I mean is still very useful to test your API endpoints, but I think since we are in DIA area, most of you are using.55:50Llmiss from AI like GPT from at Opus etc. So why don't you leverage these models to create something with your APIs and to.56:10Start to leverage your APIs, even if you don't have a lot of experience for testing purposes and for read APIs. I think this is a fun exercise, so you just go to an AI tool of course, it might not work at the 1st attempt, you might need to tweak.56:27Your your prompt but you can achieve nice things. And the good thing about big LL.56:47Is that they are usually aware of the Catalyst Center APIs or the catalyst center SDK, and I have two examples here and in this case like I have prompt here, let's say shortened, e.g., but just to show the intents so you can just go.56:56To cloud set CPT.57:16Whatever if you have an internal tool that uses these models and you can just tell it. In this case you can see here I hope it's not too blurred so using the catalyst center Python SDK an SDK is just a wrapper, let's say a library that will wrap from the.57:17Api's and then in your code instead of calling the APIs directly, you can use functions in Python that will map to this APIs just to maintain the code.57:37Not only more reasonable but all also if the APIs will be updated and the endpoints change, this will be reflected as well in your functions. So as you can see here, I'm just saying, ok, can you create me a Python script that 1st connects to catalyst center, instant using the.57:52Fetches my network devices and allows me to run one or more device show commands in one or more devices, and actually I was trying this yesterday and I was trying just with a single prompt, and as you can see the prompt is not that big and it works. Of course, I was trying with.58:12Which is a bit more capable I was trying as a testing and also as a fun exercise with other models at the 1st attempt, they were giving me some issues. But even if you use a smaller model, even a local model, if you go to the AI IP at.58:32To the API specifications and you are able to extract the correct endpoints, you can also say the AI, ok, just use this specific endpoints. So in this case and let me show you.58:51This video I was I mean it's speeding up because opposites it was taking its time. The whole, the whole process was taking around 3 min something but I shortened the video for, for demo purposes, and as you can see, it could.59:06Created the the file. In this case I told it to hard code the credentials.59:25For testing purposes just to have to use a single file, but as you can see, it created this file. I copied the file and as you can see the credentials are hard coded here. You should use an M file, but for testing purposes, it tells you that you have to install the.59:31The catalyst center SDK, so I have copied the script, I have pasted it I have installed them the dependencies on the libraries and I have run I run the command and as you can see.59:53With a single prompt, with a single script, it's sorry for this. It's connecting to the catalyst center. It's retrieving the different network devices. Of course you don't, you won't see a lot of devices because this is the.01:00:08The sandbox lab, you can see here the different devices and it allows me to select one or more. I select one or more and I just show you run a show command e.g. show CDP neighbors, and as you can see, it was able to run some show.01:00:24Comments.01:00:44Of course this is nothing very complex, but it's just to show you the power and what you can achieve with the help of AI even if you don't have a lot of a lot of knowledge in programmability, and I know most of you come from a networking background probably. So.01:00:48Sometimes you tend to run away from programmability, but if you know more or less what you want to achieve, if you investigate a bit, with the help of AI and the large language models, you'll be able to create some interesting things.01:01:06This other example and the last example I have before handing it off to Alexe to cover the MCP part, which of course will also let's say leverage these APIs. Another thing you can.01:01:22You can do as, as you are well aware in the catalyst center homepage you have the these dashboards which you can which you can configure. You can create your favorite views, but.01:01:38You might want to receive a report, let's say every day, every two days, every three days. You can schedule it and you will receive a report, let's say an HTML file telling you the state of your catalyst center and you can actually choose what you will see in this report. And you could also, let's say, schedule it to.01:01:56For to send it a report over email and then it would send you over Microsoft Teams, Webex, whatever. It would send you.01:02:16Let's say a summary of the executive report. So what I did is, I was telling I was using cloud Opus as well, but of course if you use a less capable model, you'll have to iterate more or be more specific with your queries and.01:02:25Tell ok for the network health to retrieve the network health, use this specific endpoint, ok? But generally I was just asking can you please create me a python script that will generate the report a network health report. This was the output, and actually we can, I think we can.01:02:45Run it quickly in 2 min.01:03:06I have the the the script here which is using the, the catalyst center SDK. In this case I'm using an environmental file for safety purposes. But as you can see here.01:03:10I was able to, to create the script quickly, it's run the report and it saved the, the report and as you can see here, it's fetching the actually information that you see here in the catalyst centers you can see here the health of.01:03:27My network is network devices is 56 %, it's fetching the devices, it tells me.01:03:47Okay, you have seven out of nine reachable devices, two unreachable devices. You can see the client health. If you have critical issues, device reachability, compliance, whatever.01:03:56What I usually say when talking about this topic if you have APIs only forget which center but for every other system with an LLM when with the limit, they imagine your imagination is the limit.01:04:12And as long as you have APIs for that and API endpoints for what you want to achieve, the sky is really the limit. You have the flexibility you want. If you want the report to show you something else to be delivered to any other destination.01:04:28You'll you are free to do it, right? So I think and just to wrap up before handing over to Alexe.01:04:48Catalyst Center APIs, I think they provide a lot of value because it provides you with a single point of entry into your, into your infrastructure that's managed by catalyst center of course and with a single, not a single API but with a single type of APIs, let's say you can query your info.01:04:56Infrastructure, ok? And you don't have to authenticate with individual devices and so on and so forth. So I think it provides immense value. If you haven't tried APIs before because you are scared of programmability, you can just.01:05:16Leverage AI to help you in the way. So thank you for attending this part of the presentation. I hope you stay because now it's I wouldn't say the most interesting part of the the presentation, but I would say very, very interesting.01:05:32So Alex say over to you.01:05:52Hey, thanks Lorenzo, thanks a lot. So let me share my screen I hope you can see it now.01:05:57Yes Alexi. Go on, ok great thanks. So let's talk about the MCP servers and how we can use it with Cisco Catalyst Center. So we'll talk about MCP MCP protocol, what it is and how we can use it, yeah?01:06:16Yeah, so model context protocol or MCP it's some traffic developed. It's now open source and community driven. It was released well just in.01:06:342024 in November so it's quite relatively new. Yeah, it's an open standard that enables AI agents to use natural language, to discover and call tools resources and prompts, but it's more about tools which made this.01:06:53Protocol that popular. Okay, as a network engineer, she is a human, I'm going to ask a question and I ask an AI agent which has LLM on back of it.01:07:14And the AI agent is just the application, yeah, which will communicate with me. This agent could be in a very complex agent or could be basic one, it can use local or remote LLM, but choice of LM is quite important here and small models or which are not.01:07:29Trained with the specific knowledge domain, let's say for networking or catalyst center might give low quality answers so EI agent here is just responsible for orchestration, controls of the workflow, managed state called tools, handles execute.01:07:49Etc.01:08:10And then, yeah, AI agent calls the tool through MCP client. Again, you see tools resources and problems, but we're not going to touch resources and prompt because mostly it's just kind of the tools what made the MCP popular, yeah. And a lot of MCP servers don't even have resources.01:08:12Prompts implemented.01:08:32Yeah, the agent calls the tools through MCP client and LLM kind of understands the reason selects possible actions and summarizes results. Let's talk in the next slide about more the concept of the MCP. Again, it's more or less the same call flow, but we can.01:08:35See that from MCP perspective, we can call AI agent LM and MCP client just MCP host.01:08:54And, the host provides me the user interface yeah and the agent provides me all the orchestration as I mentioned within the MCP host and allows me to access the tools, yeah, this kind of the the value of it has understanding how to communicate with the MCP server through the MCP.01:09:02Client and MCP server is not part of this MCP host, yeah. And then the last language model is the one that makes reasoning and selects which tool to use.01:09:22And the MCP is a way to access all different kind of platforms with API and get information from those platforms, so it's just kind of, you know, just a gateway or from APIs to tools and to connect to.01:09:32To LLM, yeah. So it makes it very easy because when, e.g., I'm looking from under architecture perspective, for the MCP host, I don't need to know what is the API, you know, documentation for specific API. So it's all done, let's say on the mc.01:09:48MCP server side, just add a new MCP server to your MCP host without changing anything in the MC MCP host itself, yeah, in your kind of application or it can be like cloud quote or VS code or cloud desktop or whatever, yeah. And, yeah, talking about this protocol.01:10:08I think we have like two protocols, we have ctio. It's.01:10:28Kind of code is running locally just under the input output protocols so your MCP host or MCP client is able to communicate with MCP server. Normally use it just.01:10:33Locally your MCP server, just the code which is running there and they're using this protocol or can be remote so streamable HTTP supported, let's say you have MCP servers somewhere deployed and you can connect to it and you can use it remotely so.01:10:50Yeah, what's next? Yeah, you, it supports multiple MCP servers so MCP hosts can communicate with many of them. Yeah, you can connect different MCP servers and.01:11:09Yeah, e.g., here's here we have MCP client for catalyst center MCP client for Meraki and they will connect to different MCP servers. Again, they will get all the tools while resources have been prompt, but as I mentioned it's mostly about tools what are interested kind of interesting for us.01:11:27Yeah, you need a separate MCP client here, but again, the nice thing it's kind of in most of those applications like low desktop or, in, let's say VS codes. It's just few lines which you have to add in.01:11:49Into configuration to be able to connect you don't care about this MCP client stuff unless it's really your application you write from scratch.01:12:07Yeah, let's talk about the workload here a bit and not the workload but what more happening in the MCP application under the hoods. So all the communication between the agent and MCP client and large language model is happening inside this application.01:12:16So, this we call it ok from MCP perspective it's MCP host and yeah again the agent in our case is just orchestration. Well, it gets input from user communicates with LLM and MCP client, so it's kind of.01:12:36Keep this communication orchestration, yeah. Then again the large language model could be on prem or could be in the cloud as well and I'm going to just get the message from the human, from perspective, send it to from send it to large language model, large language model checks the.01:12:55Message and kind of understands what you are asking. Maybe there is NO tool.01:13:14Needed to be called for that like I don't know, you ask what is catalyst center e.g. and then you just get an answer from immediately without kind of using any tools or maybe decide I'm going to ask about kind of, you know, just what is my catalyst center topology and then.01:13:21Okay, thanks LLM knows about the tools and it can call those tools, yeah. And then.01:13:41Yeah, again, this is what LLM is doing or kind of this reasoning and it's important to understand that this LLM decides which tool to call, yeah. And then the agent sends the tool request to the MCP client and then client calls the MCP server, yeah. And the MCP server returns the information and the last language model.01:13:51We decide, yeah, I have kind of the final answer or NO, I kind of don't have the final answer yet. I need maybe to use another tool because maybe you go, you know, just in chain asking something like 1st thing you have to get the topology and then maybe you ask for about specific device which has to discover and to get all this information.01:14:11So yeah, it's kind of very important what's LLM does for us in this case yeah and again if you have a large language model that doesn't have knowledge domain, maybe building an MCP server for networking is not going to be very successfully in this case, so it's very important.01:14:33And what kind of you are using yeah and again MCP client is focused on just communication to the MCP server. It knows how to format the message, so usually it's not very complex, sort of clients already have kind of declined part there, as I mentioned, like e.g. my demo will see.01:14:53Be using VS code, so I need just to configure few lines which point to a local or kind of remote MCP server whatever MCP server you're going to use and we'll start the client part for you. Yeah, I don't really bother about it.01:15:13Yeah, again, let's go through the common workflow from user experience. So again.01:15:33When MCP host starts, it will register to all the servers and finds, all the tools and the schemas e.g. here we have two servers, Meraki and Catalyst Center, so we'll know at that moment we have these tools for Catalyst Center and these are the schemas, so again what kind of parameters we can use.01:15:45To call those tools, what would be required for them, yeah, to know what tool to call and how to call it with which parameter, yeah. So then as a human, I'm just going to ask the question and then.01:16:06The language model will receive that question yeah but the context such as it's not just the questionnaire, you receive a context says such as user request again available in CP tools and tool descriptions so everything what was kind of discovered in previous steps.01:16:25So all kind of tools and the descriptions they will be propagated to the.01:16:45Yeah, to the LLM, yeah. Then again, maybe previous tools tools results and system instruction, conversation history again and previous tool results and conversation history probably just started it.01:16:50So for kind of already existing chat and it will reason and we'll have a plan here for you also says ok I'm going to call three tools e.g. and in this sequence and then it will instruct to call the tools from the MCP server to get the.01:17:05Information and then the MCP servers execute tools and return results to the MCP host through the MCP client.01:17:25And then, yeah, at this time is kind of the answer goes back saying that DLM make a decision to have the answer based on the question what was received from the user or not and if that just the final answer that is going to be provided.01:17:38To the users through natural language. If it's not the final answer, then again it will go back and well I need maybe to think again. So what kind of information is missing to answer this question? What can I do? And then maybe I need a different reasoning one and I need to get decent information because I don't have.01:17:57Everything that they need and then again can call few more tools e.g. and the quality of the model kind of important here to make all these decisions, yeah. And, yeah, one thing to highlight that LLM is part of MCP.01:18:17Host application, yeah, host stack, but not the MCP server, yeah. So MCP server is kind of LLM is not connected there, it can be pretty **** device, it just, you know, convert let's say in our case.01:18:39We have catalyst center APIs from one side, but then it will just to convert to this MCP protocol so LLM can use APIs. Again, the main purpose of this MCP servers MCP concept is just.01:18:54We need to provide some extra context extra data to the LLM yeah to good to make good kind of reasoning for us yeah ok why MCP server not direct API calls? Again, we were talking about the API calls in the previous slide.01:19:13Right.01:19:33Lawrence was covering it, but yeah, again, you can see a lot of bullet points on this presentation on this slide, but I think what's very important is what do I do different between when I use APIs and when I use the MCP server, yeah. So why should I use MCP server?01:19:36So again in API case, I have a kind of a network workflow that I want.01:19:54Get a report for all the clients or e.g. this location and I need to make an API call 1st to catalyst center to get a token. Then I need to get maybe the site hierarchy through the APIs, then I need to find the sites and maybe the.01:20:00Pass information and go to get the clients and make another API call and maybe multiple API calls because I, I need details and yeah, I have to go to get details for different clients. And then.01:20:20The thing that I have to create the loops, I have to pass the data in my code and so on and so on. So it can be complex and time consuming, yeah, to implement it. And it's very specific for your task you run, yeah. Once you create kind of this workflow.01:20:35It's not always easy to use it for other tasks and you decide to do something else yeah but you have to rewrite your code and so on, yeah. But if I use MCP servers, I don't create the workflow here, because I use tools that already exist and then the last language model has the reasoning to understand I need to call.01:20:52So these four APIs, e.g. or well for tools here to pass the data and I do the, I don't do any kind of parsing so it's kind of on.01:21:12This MCP and the LLM part, is doing this. So yeah, NO regex or whatever, yeah. So versatility of MCP servers it can it's really nice, it's fantastic, yeah, I would say because.01:21:24Let's say with 20 tools I can resolve many different types of use cases versus traditional way with APIs, and the the thing is that the LLM can kind of, you know, just combined or calling different order or those APIs to achieve different results. Yeah, you don't need to write really the workflows. Yeah, it's kind of LLM making.01:21:41Decision for you. And in case of APIs I have to create a purpose built tool every time that I need something.01:22:01So, yeah, again, MCP tools options, so what are the DMCP server tools? At the end of the day, they're just an API obstruction, like in our case or they could be operations with files they can go to search online something.01:22:12Kind of that's a default tool for a lot of agents, you know and the MCP servers could be just an API call or maybe more advanced tools like a sequence of APIs that they want in a specific order or even more advanced tools like AI flows agent driven, multi step or flows.01:22:29Combining multiple existing MCP tools and tools could be also auto generated as well and we are going to talk about auto generated tools in few slides and again, if you think about APIs, we have kind of, you know, just for like a thousand I think endpoints here and they.01:22:49I don't think that we should auto generate tools for all of the APIs. It's a bit kind of naive to create a tool for all the APIs and use only so like that you.01:23:09Yeah, I mean it can be a limit of some limitations of LLM itself, like context window, because as we mentioned in the workflow, all these tools they have to be propagated to the LLM, yeah. And it will consume your context window e.g. and.01:23:21I think it will be far less deterministic what what we get, yeah. And but other generation is kind of important because it makes tooling a lot easier, yeah? And it's not the only way to do it but.01:23:43This is maybe a kind of a good approach here. Yeah, we're going to talk about this a bit later. So, yeah, let's go to a quick demo of what we can do here. Again, this is just a visual studio codes. Again, you can use.01:23:59Anything else where you can use maybe you have cloud code or cloud desktop or open quote or whatever, yeah, for.01:24:19For OpenAI or your own application, yeah? And, so, yeah, basically we want to ask some questions again here in VS code and in other applications you can choose different models again model.01:24:27Is quite important then yeah the reasoning.01:24:47Also, yeah, you can put different, then, these are the tools you can see that, e.g., here I have 46 tools selected and these are for catalyst center specific tools. And then, yeah, let's say I want to ask something. Again, nothing was kind of trained or whatever, just connected.01:24:53The MCP server to my catalyst center and that's it.01:25:12And, let's say I ask for some something basic like, ok, what is my site apology? And then again you can see here ok knows about the the the tools and it decides, ok, to use this specific tool because it has kind of description of it, what it can do and it knows that, ok, I can get site apology using this.01:25:17Specific tool, yeah, which at the end we'll just we'll.01:25:37Will be an API call from MCP server perspective. And again, quite quickly I get a response again, it's kind of a small lab setup but then it provides you even kind of summary kind of key location it decided on itself or what other key location and so on, yeah. So, but you can maybe ask for better format and.01:25:41So on. Then again, you, e.g., ok, what's next? I can ask what kind of, what kind of devices I have with IP addresses and software and this is what maybe I want to see. Again depends on your.01:26:01Probably if you have too many devices then you have maybe to be more specific what you want to ask about specific site and so on. But at the end it you don't even care about the pagination and other stuff. So yeah.01:26:18And another tool was involved like get network devices, it will correspond to specific API, again, it's just one to one, it's quite simple here. Yeah, you get all the information, you get all the devices again maybe it's not the best state of this lab, but yeah, some devices are up, some are not.01:26:32But you can get all the information about these devices, what I ask for, then I will just maybe go a bit forward and.01:26:54Let's say I was just asking about, I want to see interfaces and state of of those interfaces on on my specific device I selected. So it's a switch. So again.01:27:04Just one, tool was selected here. Again, you can see that it selected specific device ID, learned it already in the previous call when I was getting the summary of all of my devices and then it just provides the information for me.01:27:20Again, it's pretty straightforward here, not kind of too many tools were called and then e.g. for the last question here, I want to, to get it in a bit different way yeah just ask about.01:27:36Running the comment show AP interface brief e.g. comment or it can be whatever kind of show comment you want to use and then yeah again LLM doing all the reasoning on behalf of me and here e.g. you can see that there will be a number of.01:27:54Tools involved yeah let's finish it and yeah we can talk about it, about the workflow, so.01:28:13As you can see here again I didn't do anything here. I don't need to write any workflows, any loops, any, any controls so what it does when I ask ok can you run this specific comment, it has a specific tool which is run read only comments. So again, we don't run.01:28:32The comments directly from the device, but at the end, yeah, maybe you can have another MCP server with I don't know radkit or whatever where you can execute it directly on the on the device once you discover it through from catalyst center and you know e.g. the its management IP address. But here the workflow is like.01:28:51Yeah, you run the comment then again you get just the task status back in this tool and then the LLM itself is kind of, you know, just clever enough that ok yeah in this comment I don't get the reply of.01:29:11Or or the output of the, of this comment, I just get the task ID and then, ok.01:29:31For task ID, I have another let's say two which is get task status, yeah. And then it checks the status to check that the task was completed and then once it's completed, ok, it also knows, ok, I need to run another let's say two, yeah, I have to use this.01:29:39All get task detail when it kind of run this get task detail and in this get task detail again it doesn't get the output of the comments directly, it just gets kind of the link to the file which can download and then it was another tool so it's kind of four two where it can download the file with the output of of.01:29:59The comment, yeah.01:30:19And, yeah, once you get it again, it's analyzing, reasoning and it kind of provide you and again you can run different comments e.g. in this case, yeah? And yeah, that's kind of a good thing because otherwise you have to write all this workflow if you're just using APIs and it will be.01:30:20Maybe specific to I don't know for this workflow, but here it's kind of for you and yeah you can see what was kind of inside, what kind of output you get from MCP server and just propagated to.01:30:40Do LL is doing all this kind of reasoning and provides the final answer like yeah you have just two interfaces up, and so on. So yeah, that's it. That was kind of the demo. Again, the nice thing that you can quite easy to.01:30:57To use it, yeah. It's not that much needed again. You just need kind of LLM subscription and so on and kind of those code desktop or this code or whatever, yeah. To do that.01:31:14So, again, let's maybe quickly check simple MCP server again it's not catalyst center related, but just to give you a concept again, it's in the python, but Python is not the only language you have SDK from, let's see from Andropic for different languages, here it's even using this fast.01:31:28MCP ok it's.01:31:49It's part of MCP but it now kind of goes a bit in parallel. I don't go don't want to go into detail but I just want to highlight that you have different ways to implement and to run this to run your own let's say MCP server, yeah. And if you check this one.01:31:51Again, we have just some lines of code for MCP servers to start from specific library, then we just have two tools and basically create a server and we are using the top two functions here just by python tank.01:32:10Functions and we will wrap it and for some and for to multiply any kind of LLM will learn with how it it will learn.01:32:291st of all description is very important here in this function. It describes the two to LLM and also in our case MCP library automatically described the parameters to the LLM, yeah, so it kind of know that it can use float and not just e.g. string as an.01:32:38Input and so on yeah and this is just an idea, yeah, for you. So again, it's kind of this library itself, it also takes care about converting it from, let's say Python way or it will be another language to the kind of MCP format, yeah, here.01:32:58So let's go to the next slide, but yeah, again, we are going, this is the link and open source code. So what about catalyst, center MCP server, maybe if you want to play, ok for now we were talking with catalyst centers and Cisco Cisco BU.01:33:17And Cisco strategy is to publish official MCP server, so we are just kind of waiting for the official one but this one is just an open source and it's published by.01:33:37I don't maybe knowing the guy his principal technical marketing engineer, the official one should be available soon, yeah, according to our discussions with BU, but the nice thing again about the MCP servers that it's easy to move from one server to another one.01:33:50It's kind of plug and play it's like a USB or whatever you just connect to your laptop, yeah. Again, this MCP server is not part of MCP host itself, so if you have kind of this MCP host I don't know cloud desktop we can call it or you have open AI kind of subscription with.01:34:07You use their application, you can add MCP servers there or open code and you can start with this one or even you can start, you know, just using some other MCP servers and then just to go to the official one, it's quite easy to, to go there.01:34:27But, let's talk about, about this specific implementation. Again, I don't want to go too much into the details but maybe to talk about some kind of challenges we have and in case of the server, we can do the same isn't our math example but.01:34:47Yeah, if we kind of, we can write the code functions corresponding to API endpoints and we have to describe the parameters of the API endpoints as per open API specification and to do the description and so on and so on or maybe you can use there is e.g.01:35:06There is a nice way where you can.01:35:25Kind of download the open API file and you can convert to to MCP tools. The thing is too many tools is not also that good as I mentioned for the, for the LLM and and for the context window, yeah. If you have too many tools, it can.01:35:28Consume your context window, but.01:35:48Again this kind of it starts to get very hard to maintain or especially if I want to do, e.g., let's say you start with 2030 tools, it's already difficult here you have to write all this in the code and you have to add something else. So another option is actually to define the tools as a tool definition.01:35:51In this way as you can see on the screens, you have this catalyst config yaml file. So just in the YAML and then the MCP code, you see the MCP codes, it's kind of, understands how to import this yaml file and create the tools for me. And the only thing that I changed is the YAML file here. I don't go in.01:36:11Did I change anything in the corner. This is how this.01:36:31This MCP server which I shared in the previous slide with you, if you can go and download and use it is implemented. So, again, if you want to play with it and e.g. you want to add the tools, you can easily do it. E.g., if I have, I don't know, ten tools, then if I want to add another tool I just add to the yaml file.01:36:35And nothing changing on the server from that perspective, so NO code has to be changed and so on, yeah. So it's results a major challenge that we have.01:36:55Building tools much easier than actually do it by hand. And again, you can also ask to create a tool for your for you just let's say you need a specific API to be converted to the tool, you say, hey, just, can you.01:37:05Do it for me kind of created in this YAML format. E.g., this is YAML format for you. You can give this file as an example and then you can provide some open API specification which can be also e.g. in Jason, you can download it as Lorenzo mentioned from Catalyst Center or from Cisco.01:37:22As well, yeah.01:37:42And it will create it kind of a new tool for you. Then, yeah, just to talk a little bit what you have in this YAML file, what's kind of needed there and of course you need the API endpoint as you can see, I need the method get post, put delete.01:37:44But again, I would highly recommend to start on your tools with get because with the.01:38:04If you provide all these functions with postpone delete you can do some well let's say LLM can do some call APIs and delete something, probably you don't want to have this result. And then the parameters that they need in the location of those parameters again we have kind of these parameters in this specific.01:38:10Case, you have this get issues details, e.g., and you have this.01:38:30ID parameter and then it's kind of what type is, what is the description of these parameters or then we'll know again it is kind of important to get this good description then if it's required or not required, e.g., and then location location just where it is in terms of the request to in the API request, so it can be like in the pathway.01:38:38Like we can see it in our case, you see the ID just goes in this assurance issue and then just ID you put it there.01:38:57Or it can be in the body for some other APIs or in kind of custom headers, yeah. But basically specified in the API specification. So again, how you can edit e.g. for VS quote and it's more or less the same for other like open quotes you have like specific file like it.01:39:05Jason where MCP JSON e.g., it's called in my case and you can add the servers, as you can see we have just two servers here. They're using cdio, and.01:39:26You just have a name and you just kind of point to the Python interpreter, like here it's in virtual environment for you so basically it's like to run the Python script here you say the python, which one you are using and then the script which you want to run.01:39:38Which is your.01:39:58Your MCP server, yeah, and that's it from perspective of those like visual studio quote or open court or cloud desktop, how you can go and how you can add the MCP server, so you don't care about MCP client stuff and that's it then once it's enabled, it's running, you can kind of.01:40:00Can go and ask the questions like I showed in the demo, yeah? So again here just to local MCP servers and yeah, usingdio. Again, they can be kind of remote e.g.. So yeah, that's it from my side and.01:40:21Going back to Alex, thanks for you guys, for your attention.01:40:40Thank you, Alexi, and thank you all, Andres, Lorenzo, Albano, but most importantly thank you the attendees for spending the last hour and 45 min with us. I hope it was an interesting, very interesting.01:40:44Session.01:41:04If you want to see more webinars on old technologies, you can visit CS.co/EIS that's Q4. For catalyst center recordings, webinars information, you can go to cS.CO/EMEA that's EI.01:41:05And if you want to contact us, please send your email to annex CSS.EMEA@cisco.com. Once you hung up once you drop you will get presented with a post webinar survey. It's very important. Please fill it.01:41:26Been spend a couple of minutes, maybe less.01:41:45But please send it so we get your feedback and we try to improve in our next webinars. Last thing before we go, if you haven't onboarded to Cisco IQ let us know in post webinar survey, let us.01:41:48Know through an email and we will contact you to do it together.01:42:08That was the expertise site series webinar on mastering AI tools in Cisco Catalyst Center from Andres Franco, from Lorenzo Pino Gazilva, Alexi Biscuunoff and of course myself, Alex get this.01:42:12Thank you so much. Have a great day.01:42:33