Building with Claude on Google Cloud

1. SPSpeaker0:00

   [upbeat music] Hello, everyone. Thank you for being, uh, for being here today. I'm pretty much excited to have, uh, this session with you. I know we are almost at the end of the event, so thank you for bear with us. So just to start, let me to introduce myself. I'm, uh, Ivan Nardini. I'm a developer advocate at Google Cloud working on building content in partnership with Anthropic. And, um, in order to start this presentation today, I just want to ask a very simple, uh, a very simple question. So how many of you in the last week used any AI tool, uh, to code or build application? Okay, the majority. Um, how many of you used the same AI tool to build and deploy application on Google Cloud? Yeah, just a few of them. So the, the goal here today is just try to make it better. So in this, uh, in this demo, what I'm gonna show you is, uh, how you can use Claude on Google Cloud to build and, uh, deploy application end-to-end. And, uh, in order to do that, I'm gonna wear five different, uh, hats. And, uh, we will start from, um, imagine the use case, right? Imagine the scenario. You are an enterprise context. In an enterprise context, you probably have a team like the one that we are visualizing here that, um, it's engaged to build a new feature or a new product. So starting from the left, uh, you probably have a PM. Uh, some of you in this room are PMs. And, um, the PM might have an idea of how to improve a particular product or how to implement a feature. And starting from this idea, he shared this idea with a UI/UX design, which allows him to design the idea, to visualize the idea. And, uh, after the idea start getting a shape, then, uh, the idea is sent to a software engineer that essentially start developing the core logic, uh, behind this idea in order to ship and make the application accessible or the feature, uh, the future, uh, feature accessible to everyone. And, uh, but before to ship it, in order to be confident on what you are releasing, you probably have to pass, uh, a security review. So that's why you have a, a sh- a role like the security engineer that will allows you to be confident in the release. And finally, once the application gets, uh, released, um, you probably have a data persona, like a growth marketer or a data analyst that analyze, uh, the data that are collect through the app in order to generate insight and produce feedback to the PM, um, in a way that the, uh, product or the feature itself gets, uh, i- improved. So these are kind of, uh, you know, all the personas that, in a simple way, we can imagine are involved in, uh, the software development lifecycle. Now, with respect to these personas, uh, Claude Code, the Anthropic's coding agent, a- augment all of them, providing several components that we are gonna use, uh, today. So as I said, in the, uh, for the remaining part of this presentation, I'm gonna put the hat of all these personas. I'm gonna show you how you can use these, uh, Claude Code components together with, uh, Claude on Google Cloud in order to ship, build and ship a very simple, uh, feedback application that at the end of the presentation we are gonna use to rate, uh, my, my performance here. But before to start building, uh, of course, you need to, uh, set up, um, the, uh, Claude Code that we are gonna use. And, uh, I'm so excited and so proud that we work together with Anthropic to make this process of setting up Claude Code in order to use model on Google Cloud in a very simple way, in a very straightforward way. So you have multiple methods to use, uh, Claude models on Google Cloud in Claude Code, but the simplest one, the faster one, is using, uh, a, the application default credential, which, uh, automatically fi- uh, finds, uh, your credential, for example, the user one, and based on the environment that you're gonna use. And, uh, as you can see in this representation, so recently, Claude Code also introduced this, uh, wizard that will simply allows you to, uh, detect your project and your region where the models are served and, um, uh, check which models are available in your project and, uh, like let, uh, you to pin them in order to start building your application. At this point, like, probably you are familiar with this, and you're wondering, "Okay, but what's i- what is different to use just Claude Code with, uh, with, uh, Claude models? Why using Claude on GCP, on Google Cloud?" So there are many reason, uh, why you want to do, why you want to do this. So first of all, because you pay for what you use, so the con- the usage of a Claude models on Google Cloud is per token, so you don't receive a message, uh, message cap. And also, if you're building a, a enterprise application that needs to go, uh, to production, you can always, uh, access to what is called provisioning throughput, which essentially will reserve some, uh, um, qu- uh, throughput for you in order to build this kind of application. Um, the other important reason why you want to consider, uh, like, uh, Claude on Google Cloud is, as I said, the setup is pretty straightforward using the, uh, the ADC. You don't have a API to rotate or, uh, you know, uh, environment variable to set in some sense, so it's a, it's a happy journey i- in, uh, in, uh, with respect to this aspect. Uh, you can access model in your project, uh, with your own, uh, um, you know, policies, uh, set. And, uh, also, like, the data stays in your project while you are interacting with the Claude Code.And model are served, uh, in multiple region, so you have global endpoint, you have regional endpoint, depending on, you know, the availability that you, that you need. And as a Google Cloud, talking about availability, we are very great, uh, availability service that, uh, standards that will allows you to, uh, use Claude and on one of the most performing infrastructure that you can find, uh, in the, in the market. So these are all some of the main reason why you want to consider Claude on, uh, Google Cloud, especially in an enterprise context. So now that you have, uh, like few reasons of why using Claude on Google Cloud, we are ready to, uh, build. And so, as I said, I will start wearing the hat of a PM. So imagine that you just join, uh, the company or maybe you're already part of the company. You have, um, you, you want to improve our services, you want to implement a new, uh, f- features, uh, with respect to a particular product. What was happening in the past is that you have the idea, you go to a UI/UX designer, and, uh, you ask him to prototype and visualize the idea. Now, with the, with the Claude and CoCode, all you, you... what you can do is just, uh, um, drawing a picture like the one that you see, uh, here while you're drinking a coffee maybe in, uh, San Francisco, and, uh, then let, uh, Claude doing, um, uh, implementing the idea for you. So let's see this, uh, uh, in action. So this is, uh, the CoCode UI, like you will probably familiar with that, and you are familiar with the Claude MD, which is essentially gives some instruction. Here, we just say, um, that we are a PM, we wanna, uh, we want to have, uh, starting from the picture, we want to render, uh, a prototype of, uh, the app, the wireframe that we are gonna then use and pass to u- the UX designer. And in few minutes, you can see how Claude was capable of rendering it, and, um, just starting from a very simple, uh, picture or drawing, uh, that you, you, uh, did while you were drinking your coffee. So pretty, pretty straightforward. But imagine, uh, how much time you save in doing this because compared to what you were doing in the past with the back and forth, uh, to, in order to get this first prototype of your idea. Okay. So at this point, the PM gives, uh, like, uh, creates a prototypes and, uh, pass these prototypes to the UI/UX, uh, developer. And, uh, at this point, he needs to implement a more solid, uh, interface in order to use it in, uh, in production. So in, uh, in this particular use case, what we want to create, it's at least three like, uh, pages, from the landing to the thanking, uh, um, message, like message page, and a dashboard that, uh, will allow me to show you in real-time what can be the feedback that I will receive from the room. So in this case, there are many ways you can, uh, you can implement this. But in this case, I want to use an additional capabilities of CoCode, which is, uh, the plan mode. So with the plan mode, what we do, we put Claude in a mode where it thinks, uh, before to, um, like it thinks and propose what it's gonna do before to implement, uh, any code. And this is very important because it gives me like, uh, a degree of freedom of deciding to change something based according to my preference or according to some standard that probably I will get access to an MCP server using, uh, Figma, for example. So now that we have, uh, in mind what we are gonna build, let's see this in action as well. So we started from, uh, the wireframe from the PM. S- similar prompt, I enabled the plan, uh, the plan mode. And so as you can see, compared to before, um, in this case, I'm simulating the receiving some instruction from Figma using a design doc. But a- as you can see, compared to before, it creates a plan of what it's gonna do with respect to all the components that I defined in the slide. We look at them, we are happy, we accept, and CoCode will, uh, implement all of them. And at the end, what we get is, uh, these, uh, optimized version. So as you can see, we start from here, and we get this. Very, very straightforward. But you can see how we are shifting from a prototype to something that can be used, uh, in this session in a very simple way. Okay. So this is the part that probably every view, uh, or every, uh, of you in this room, like you do every day, right? Um, let's, uh, let's wear the third hat, which is the one of the software engineer. And, uh, the software engineer, he receive this, uh, front end, like, uh, all the components that I was showing you before. And, uh, maybe he doesn't know anything, as probably, uh, some of you in this room, it doesn't know how to deploy this application on, uh, Google Cloud, right? So how you can, how you can do that? How you can, you know, uh, hand to have this clear picture of what are the components on Google Cloud that you need to use in order to deploy a very simple application like the one that, uh, I show you today. Luckily, it's not a problem because as a Google Cloud, and we invest a lot of time to integrate with this large, uh, by coding ecosystem that now is, um, growing around, uh, models. And, uh, we have, we, in the last few months, we released two, um, important components. So the first one is, uh, the developer knowledge API with this, with the, the associated MCP server, and the second one is the Google Cloud Skills.So with the knowledge, uh, with the developer knowledge API, you get access to the, a fresh documentation from Google Cloud that can be directly consumed in Claude Code, uh, through the MCP server, and, uh, it will have Claude Code to figure it out, wha- uh, what is the best, uh, architecture, what is the best implementation, uh, to deploy a certain application on Google Cloud. This is very important because, again, what, what we are saying here is that you don't need to know, uh, like how to deploy an application on Google Cloud. You can just leverage Claude Code and this MCP server that we expose now on, on Google Cloud side to build application like this one. So in this case, uh, probably what we wanna, we are gonna do is that we are gonna deploy the feedback API on a serverless function like a Cloud Run. We will connect with a web, uh, web, um, oriented DB, like a Firestore, to corre- to collect the raw responses that we, we will give through the feedback app. And then because we want to have that data analytics part, we will, uh, build, uh, uh, implementation in a way that we can store those raw response in a, an analytical data warehouse like BigQuery. And we, in BigQuery, we will post-process, and we consume this information in a dashboard, uh, in a, in a, in a dashboard like the one that you can find in, uh, in Looker. But again, it, you, you can build this using, uh, um, a Claude Code in combination with the MCP server without kn- uh, without you having a prior knowledge of how to do that. This, like the MCP server and the, uh, developer knowledge API that I just mentioned, it got paired with also the skill part. And the skill part is if with the, uh, MCP server you will be able to design the architecture, with the skills, you will be able to cover the single blocks of this architecture. For example, we release a, a sim- a simple skills that, um, enable Claude to deploy, uh, on Cloud Run, to deploy an API on Cloud Run, or, uh, to connect like, uh, Cloud Run with Firestore. So it's more about, uh, the implementation itself rather than, you know, giving, uh, the overall picture. So once you get Claude Code enabled with the, the MCP server, uh, we, uh, of the documentation and the skills that I just mentioned, you can just directly implement, uh, the architecture that I was showing you before, and you can do this in parallel. So another components that you can use in Claude Code is a subagents. And as you can see here, we are gonna spin up three different subagents, one for the API, one for the ingestion pipeline, and the other one for the dashboard. And you can parallelize the implementation of each of these components just like, uh, you run a team, uh, a team sprint, uh, in the, in your normal, like, usual development, uh, life cycle. So let's see this in action as well. So here we are again. Uh, first of all, I just want to show you that we have enabled the MCP server of the documentation, and we have the skills, some of the skills that we pre-built. One, one time more, I provide a very simple prompt. The first step is designing the cloud-native backend. So it will, uh, start, like, it will provide me, um, uh, draft of the architecture. In this case, I could use again the plan mode, but for simplicity, I didn't. And then use the skill, one of the skill that I provide in order to implement the API. Let's say that we are, uh, happy with the API spec. Then we have the architecture, we have the API spec. The next step is running, uh, multiple, uh, agents in parallel in order to implement the three, uh, the three components of the app that I was showing you. So it's, uh, it's pretty quick. As you notice, Claude Code, uh, like it also manage the testing part one-- after it finish the implementation, and at the end, you will get your, um, your app, which is now, uh, ready to, uh, deploy on Google Cloud. Okay. So at this point, uh, we have, uh, we have the code of the app, uh, that is, uh, ready to be deployed. But because we are deploying on, on cloud, and we want to open this up to a larger audience, we want to deploy it in a confi- uh, confidently. Like, uh, so this is, uh, this is when you want to consider to run a security review. Now, depending on, uh, your company, uh, you can, uh, you can have different security requirements. So for example, you may want to check if your, uh, application is solid with respect to the most, uh, common, uh, OWASP issue. Uh, or because you are deploying this application on cloud, and one, one thing that you need to consider is probably you will use what is called a service account, and you want to limit, um, the, the service account when it calls a particular API, like the one related to reading and writing a DB with respect to certain role. So you are, you, you ensure that you, you are limiting what it can happen when, uh, when the application runs some operation on the, on the cloud itself. So again, these are just a couple of examples of, um, uh, what you want to consider in a, a phase like this one. And of course, this representation is a strong simplification of what can happen in the, in the real life. Um, it's just, uh, one of the possible scenario, uh, that you can have. And you can see why it's also a simplification because we are letting the security engineer not only to ap-approve if the app is, uh, secure enough to be deployed, but it will also deploy the app in this case. ButAgain, this is just a demo, so we have this degree of freedom. With that being said, let's run the final demo, and let's, uh, get the app, uh, running. So in this case, as you can see, I use a pre-built security review that you can find in, uh, Cloud Code. Uh, s- very simple, um, very simple prompt also in, in this case. And, uh, what is happening is, uh, Cloud Code run the first test. Essentially, it double-check that, uh, everything is aligned. It found, like, a, a possible issue, and, uh, it automatically fix it, so in a way that the app now is secure. And once it is secure, it deploy the back, uh, the, the backend API, and it deploy the API itself. So what we get at the end of this, it's an endpoint with our app. The app is live, so I will, uh, quickly unlock my laptop, and I will ask the backstage to share on my laptop. So the, uh, the app, as you can see, is, uh, up and running on GCP. Uh, for people that doesn't know GCP, this is called Run. It's a serverless, uh, ser- uh, service that you can use in order to deploy app. And, uh, the app at the end looks like this one. So if you remember, I show you the original, uh, I, I show you the feedback, uh, frame at the beginning. So what we can do now live, I can just give me a score. What do you think the session is going so far? Give me-

2. SPSpeaker20:11

   Oh, five.

3. SPSpeaker20:12

   Five? Ooh, okay. Thank you, man. I really appreciate that.

4. SPSpeaker20:16

   [laughs]

5. SPSpeaker20:16

   Okay. Cool session. Let's be, uh, let's be simple. I submit, and then, uh, in real time it updates the number of response, the score, and, um, you know, the visualization. And also, just for fun, I build a feedback analyzer. So once I click this, it will, uh, run. It will call, uh, Cloud Code, uh, Claude on Google Cloud, and, uh, based on the feedback and the comments that you sent, it will generate this, uh, this summary. So pretty, pretty straightforward. I will ask to go back on the presentation. Thank you. Okay, so at this point, we have the app. We collect, uh, we collect, uh, good feedback. Thank you again, man. Uh, but the development lifecycle is still there. Uh, we have the last step to cover, which is essentially people there start, uh, using our app. If you saw one of the KPI that I had on the dashboard was the time, the response time, so how long is, uh, was taking you to just in, uh, uh, like providing a feedback. So these are the kind of information that you can use in order, uh, to, like, uh, through this data, uh, you can collect this data, analyze them, and generate, uh, insights in a way that it, they can be used in order to improve, uh, the, uh, application. Now, running, um, if you're new on, uh, Google Cloud, there are several services that you can use in order to analyze this data that comes from the app. So a couple of them, one, as I said at the beginning, is an analytical data warehouse, which is BigQuery, and, uh, for the reporting part, you have a tool like Looker. But again, as we said before, you don't need to know how to use BigQuery in order to analyze those data as well as w- how to build a dashboard in Looker because we also provide, uh, a MCP server for doing this. Now, for the sake of time, I'm not gonna demo how to use an MCP server to query BigQuery or building a, a dashboard, but, uh, I want to quickly share with you, um, where you can find, uh, this information in order to do that after this session because we are gonna release the code. So I will ask, uh, to shift back on my laptop. Thank you. So, um, okay. Loading time. So with respect to all the MCP server that are available on, on Google Cloud, we, uh, recently announced the, uh, Agent Platform. So in the Agent Platform, you have, uh, services that, uh, is represented by the Agent Registry. And in the Agent Registry, you have the list of all the MCP server that we natively support on Google Cloud. So for example, we have the developer knowledge service that I just show you, and we also have the BigQuery, the BigQuery MCP server that you can use in order to, uh, query the data that we just collected from the app. Um, it's very, like, this registry is relevant because it tells you how to set, uh, how to set the MCP server on your side as well as it gives you some observability feature and the description of all the tools that you will find, uh, for the MCP server. So you know how Cloud Code will be able to use, uh, this, uh, this, um, this server to, uh, query your data. With respect to the, um, Looker part, also, like, uh, we released the MCP toolbox of DB. This is, um, open source, um, like a model context protocol server th- which include an integration with Looker. And, uh, it's very well, um, we have a very, very, very well documented quick start on how to set up with the Cloud Code and start using it in order to consume that data from BigQuery and build your dashboard. So I leave you this as an exercise, like, uh, we then, we are going to release the code, so you can go home and integrate these two parts. It's pretty straightforward, but the dashboard that you, uh, you can create, they are pretty, they are pretty powerful, and you will see, uh, how nice they can be. Okay, back to the presentation. I think I'm just in time, so time to, uh, time to wrap up. What I try to explain you today is essentially, uh, two things. So first of all, I was trying to, and I hope I did, I did good enough, I was trying to show you how, like, all the components of, uh, Cloud Code, including skills, MCP server, subagents, they can really, like, speed up the process of software, uh, development, as well as how you can use Cloud Code with, uh, uh, Claude models on GCP in a very seamless way. Like, if you saw, we run several session across, like, uh, multiple, uh, personas, and, like, it was, uh, the experience was just, uh, straightforward. It was just, uh, in- incredible. So this is, uh, what you can get if you combine, uh, you know, Cloud Code with the Claude models on, uh, on GCP. As I said, um, the re- the code, um, is gonna be available right after the session. We have a great quick start, and, uh, we have, um, a very well, uh, maintained documentation both on Google Cloud side and Anthropic side, so I highly recommend to just go there and check out. And then I hope I cover everything, but if you still have questions or you want to provide additional feedback, just feel free to reach out. Uh, these are, uh, my social media, uh, point. So with that being said, thank you so much, and it was a pleasure being here today. [audience applauding] [gentle music]

Episode duration: 26:29