Hi everyone, my name is Martin mcbrede, and I'm a customer success specialist. And in this video, I'm gonna walk you through how we can use Cisco ACI together with the model context protocol or known as MCP.00:06To make ACI easier to operate using natural language, the core idea is pretty simple. Instead of expecting to remember a number of different APEC endpoints, payload formats, and object relationships, we can expose those operations as structured tools that an AI client can call on.00:20That gives us a more intuitive interface for requiring troubleshooting, and even provisioning a new fabric. For ACI in particular, that's valuable because the policy model is powerful, but it can also be fairly complex to work with directly.00:39We'll 1st start off with why ACI benefits from MCP, and then we'll go into automating ACI with natural language, and then we'll go into a demo to see how it works. So to understand why MCP is useful here, it helps to remember how ACI works.00:56ACI is built around a rich policy model that's managed by the APIC, and most meaningful changes involve multiple related objects such as tenants, vrfs, bridge domains, EPGs, and contracts.01:15Even a simple request like deploying a small application stack can require several API calls and they have to be in the right sequence. So the challenge isn't that ACI lacks automation, it's that the automation surface can be very broad and detailed.01:30So MCP helps by putting a structured abstraction layer in front of that complexity. So at a high level, the flow starts with an end user entering a request using natural language. The LLM running in the AI client interprets that request and then decides which ACI tool.01:47Protocol.02:07The AI client, which can be something like cursor or VS code, then sends that tool call to the local MCP server. The server uses a shared ACI client to authenticate to the APIC, perform the get or post request, and then return that data in the form of structured JSON.02:08Then the LLM takes the results and turns it back into something readable for the end user. So the LLM handles the intent and presentation while the MCP server handles the actual system interactions. So here's an example, using a simple request, create a tenant named.02:28The model doesn't generate a raw REST API payload on its own. Instead, it selects the tool that it can use such as ACI tenant create.02:49That tool takes a defined input, builds the apic payload, authenticates, and then posts it to the correct endpoint. The APIC returns a success response, and then the server passes that instructured form. So from the end user's perspective, it feels very conversational, but underneath it's.03:01Build a very explicit and controlled API workflow.03:20One of the most immediate benefits is on the read side where we can use ACI with MCP to be able to retrieve information. So when an end user can use natural language to answer common questions like what tenets exist, what endpoint groups are present.03:26Or whether there are any active critical faults. The same pattern works for health checks, node visibility, interfaces, and even BGP peers. So instead of logging into the APIC UI and clicking through multiple screens or writing a one off.03:45Rest API query, you can simply just ask it directly in chat. That makes discovery and troubleshooting much faster, especially when you're trying to move quickly during an incident. That same model also works for provisioning. We can expose tools for creating.04:04Core policy objects such as tenants, vrfs, bridge domains, subnets, application profiles, and even contracts.04:24These are just a few examples. On top of that, we can have composite tools that can even handle multi step workflows such as building out a three tier application or a web app stack. That's where MCP becomes especially valuable in ACI because so many changes are order dependent.04:33In their relationship heavy. So instead of manually stitching together each and every object, the end user can express intent at a higher level and let the workflow handle that orchestration. Let's now jump into a demo and see how it works.04:53In this demo, we're gonna be using cursor as our AI client. So we're logged into cursor. We've got a new agent pulled up, and we're gonna go to the plus sign and then down to MCP.05:13And here we can currently see that it says there's NO MCPs found. So what we need to do now is go into our code and then update the section for our MCP server. So we'll go into the cursor IDE.05:27This will take us and open up the cursor IDEE in our code folder. We'll then go to the.cursor directory and here there's an mCP.json file. And here we're gonna set up our MCP.05:42Server configuration and specifically for a cursor, it uses this MCP servers to set up the config. We're then gonna name our server cisco-ACI, and then we're gonna add the command and argument to use Python to start our MC.05:59DP server.06:19So we'll go ahead and save this file. The next thing we'll need to do is set up authentication from our agent to the APIC. So we're gonna run through the setup.sh script. We're gonna give it the URL to our APIC.06:20And then we'll give it the login credentials and then this will set up our authentication from the agent to the apic. From here, we can now go back to agent mode.06:39We'll go back to our plus sign to MCP. Now we can see we have our MCP server that we named Cisco-ACI, and we also have the option to toggle on and off the connection. From here, we can also go to manage.06:58Which will give us some additional information about the types of tools that this MCP server has access to. So here we can see a list of a number of tools such as ACI tenants get, application profiles, get bridge domains get. So when we put into the proper.07:16Prompt show me a list of tenants. This is the tool that it uses to be able to retrieve the information from the APIC.07:36So here you can see all the different types of tools that it can use to get information, but there's also additional tools to be even, be able to create objects such as creating a an application profile. So these are the tools that it will use. We put into the prompt the different types of actions.07:44That we want to do or to be able just to retrieve information. So from here, we'll now go back to our agent.08:04And now that we have our connection set up, we'll go ahead and ask the agent to test the connection to ACI. So what the agent will do now is go through those list of tools that it has access to. It finds a tool called ACI.08:24Simple test. So it will then use that tool.08:43To be able to test the connection to ACI. So here we can see that we have a successful connection and we can see that the MCP server is working. So this lets us know that we have our our connection set up between our agent and the APEC.08:47So here's where we can ask additional questions such as list tenant information. So the agent will go through that same process. It will go and discover what tools it has available such as ACI tenant sket to be.09:05Able to retrieve that information from the APIC.09:24So we can see it use the tool, all ACI tenants, and we can even expand this out to see the information that it's retrieved. Then it takes that information and then it brings it back and builds it out in the form of a table.09:28Another thing we can get information on is system health or fault information. So we can say list any faults and system health information.09:47So the agent now goes with that list of tools it has access to and it's now gonna use multiple tools to be able to retrieve all of that information and bring it back. So here we can see it's used three tools to pull this information, one for ACI faults, one.10:05For a critical ACI faults and then fabric health.10:23Then it brings them back in the form of a table for fabric health, critical faults, major faults and warnings, and then it gives us a summary of everything that it discovered related to faults and system information. So these are just a couple of examples of how to retrieve information.10:27The other thing we can do is provision resources. So e.g., we can say create a new tenant named demo one. So the agent goes through that same process, it finds the tool that it needs to be able to create a tenant, and then it will create the new tenant.10:45Demo one. It also recognizes that there's additional things that would also have to be set up and it even asked us, would you like to include creating a VRF or an application profile? So let's say let's create.11:05A VRF for a tenant demo one named demo one_VRF.11:25So now we can start to build these objects and then stitch them together. So we can see it found the tool and then it created the VRF demo one_VRF. So this is one example of being able to create objects and then start to stitch them.11:39Together. But another thing we can do is use multi workflow.11:59Templates to where we can build out a tenant and then the VRF and that application profile and the remaining resources. And we can also use that using voice to text, so within cursor, we can even use the mic option here, so instead of typing.12:05Out all the information we can use the microphone to just simply say what we want to build, e.g..12:25Build me a complete web application stack called finance one in a new tenant finance one. It'll have a single VRF, two bridge domains for web and app tiers.12:35EPGs and an HTTP contract between them. Build this and then summarize what you've built and how the tiers connect, right? So we'll send that prompt to the agent. The agent will then go.12:52Go through that same process of discovering.13:12What tools it has access to, and so it'll pull together the tools to create the tenant, to create the VRF, the bridge domains, the EPGs and and contracts, and then stitch this all together end to end.13:16Alright, so now we can see that it's discovered all the tools that it needs and now it's it's building out the tenant and all of the underlying resources. Then once it's done building, it'll then summarize everything that it has created.13:35Then also give us some, some diagrams of how the communication works between all of the, the different tiers here.13:53So we can see it built out our new tenant finance one, the VRF called finance one_VRF, E two bridge domains and application profile, two EPGs of filtering contract.14:02It also tells us what subnets were used for the build, how the tiers connect, and then how the traffic flow works with the contracts. So from here we could even go and log in to the apic, and then validate our build for our new tenant.14:18So if you go to tenants and here we can see we have our new tenant finance one with two bridge domains, our VRF and two EPGs.14:36Here we can go in and validate, we can see our application profile, and then our two EPGs.14:52And then under contracts we can see that it also created the, the contract for finance one web contract and and the filters that we asked it to to build out. So this is one example of doing a multi workflow template.15:07Where we can ask a, a prompt to build out a new tenant and multiple underlying provisioning resources. So in this demo, we talked about how to set up MCP with our AI client. We talked about how to retrieve information and then we all.15:26So went over example of how we can provision resources using ACI and MCP.15:46Thanks for watching.15:54
