Build Next Level Projects with Claudes Agent SDK

I'm going to show you how you can add cloud code into any app using the Claude agent SDK. This isn't just another SDK where you get access to some kind of API endpoint and a model. It's so much more than that. It's Claude Code as an SDK. So, you get an entire toolkit to fully customize Claude code any way you want right within your apps. In this video, we're going to dive into using the Claude agent SDK. I'll first explain how it works and then we're going to get our hands dirty and write some code to give you an idea of what's possible with all of this. I want to show you this project that I'm working on. So, this is my own AI coding workflow manager tool. And my issue was that anytime I'm running multiple cloud instances, running my agents, I'm always switching back and forth. So, when one task is done, I'm always opening up multiple terminals trying to figure out where I'm at in this process. And it makes it hard to manage. So when I was learning the agent SDK, I really wanted to push the limits and see what I can do with this. And I decided to build out my own interactive AI coding terminal that's attached to every ticket in this conbon board style application. So essentially, I can open up a ticket, fire off a prompt right from my ticket here, and while that's running, I can see it on the left side. I can easily jump into another ticket, work on a task, and then once the next one is done, I can open it up and can continue from there. So, it makes it really easy and I never lose where I'm at. I even added in the ability to send off tasks to a remote environment. So, if I ever need to leave my computer, I can fire off a prompt and my agent will spin up this remote environment and it'll continue working even if my computer's off. This part actually uses Warp's Oz orchestration platform in order to spin up these remote environments, but more on that later. If you've ever built out any AI powered apps, then you understand exactly how difficult something like this would be to build. And this is exactly where the claude agent SDK can help us. Now the anthropic team explained why they launched the agent SDK in an article and their logic was that they saw a lot of people using cloud code for non-coding tasks. So things like notetaking, video creation, working with Excel files and so on. So they decided to take that same engine that powers cloud code and release it as an SDK. So you can just keep building on top of it and it's become extremely useful for that. Now, it's important to understand the difference between the Claude agent SDK and the Claude client SDK because these are two different things. So, the client SDK is a wrapper around the cloud API. Maybe you've worked with OpenAI's API. Essentially, it's the same thing. So, you get this API endpoint and you send off a request and you get back a response with access to a model and that's it. So, nothing on top of that. You have to build in everything else from scratch. Now on the other hand you have the clawed agent SDK and this is a readytouse agent. So essentially when you fire off a prompt this is a fully functioning agent. So it already has the agent loop. It already manages context. If you want to add in memory there's just a few configurations you have to add in. And it comes in with built-in tools. So if you want to read and write files, you want to run bash scripts, you want to perform web searches, all this is already built in. and you just configure what tools and how much permission this agent has. So the client SDK gives you access to a model and everything else you have to build out on top of it. The agent SDK, you get a readytouse agent out of the box. Okay, so let's actually get into building something out here. So we're going to go ahead and work within this simple file here and I'm going to show you how to configure your first agent. So we have a few different options for how we can install this, but in my case I want to stick to the TypeScript SDK. So, I'm going to go ahead and copy this command right here and go ahead and just install this. Once we've installed the SDK, we can get started by importing this query method from the Claude agent SDK. Now, this function is just our main entry point and it initializes that agent loop. In here, we'll kick things off by passing in a prompt. And in my case, I'm just going to tell it, please review the images in this current directory and create a report of all the data in them. Now, I'm going to explain what's happening here in a second, but ultimately it's just a folder with images of receipts. Next, we want to pass in our options. Now, this is a series of values that configure our agents behavior. So, the settings we pass in here will determine what tools our agent can use, which model to use, and any other configuration we need specifically for this agent. Now, within my options, the first thing I want to pass in is the current working directory. So, this just tells the agent which folder we're working within. Now, in my case, this is just going to be a hard-coded value. It's going to point to a folder on my desktop, and that folder contains those images of those receipts. So, from here, I'm just going to set the model that I want to use, and I'm also going to pass in my Claude API key. Now, if you don't have one, just go to your Claude developer console, and you should be able to create one under the API keys tab there. So, go ahead and create it and pass it in here. The last thing I want to pass in is going to be an array of allowed tools. So out of the box, the Claude agent SDK already has these built-in tools for us. And here we can just add them in. And this tells the agent what it can use. So we have tools for reading and writing, performing web searches, running bash scripts, and much more. So for this demo, I'm just going to pass in all the tools that we already have built in. And in this case, this agent is just going to have access to everything. So this way you can get an idea of what's available. We've configured the query method, but now we need to handle the response. And this is going to be an async iterable. So we're just going to write a for loop to process all the messages as they stream in. So as the agent is working, we're going to see a series of messages like the tools the agent is using, the agent's thinking and reasoning process, as well as the actual messages that the agent wants to return back to the user. So it's up to us to extract which messages we want the user to see and get these values. So if I run this and log out all the messages, we're going to see something like this. It's just going to be a bunch of messages as objects with the details of the message contents. So here we see the assistant messages where we can see the type value and from here we can get access to the actual contents of the message. The last message will always be with a type of result and this is just that final assistant message telling us that the task has been completed and it's like that final message that you see in any chat box. So here we can see that this assistant reviewed all the images and it even created a report based on the transactions that we have. So that's actually pretty cool. So I want to return the actual message value and not this object here because I want to stream this back to the user. So in this case I'm going to access the message content value and this is going to give me that message text. I can do the same for the final result by checking for the type of result. And now when I run this I can actually see the message contents being returned. And this is much cleaner. And these are the values that I'm going to stream back into that user's chat interface. All right. So, I've showed you how to connect a basic agent. And now I want to show you how powerful this actually is. So, I took this code and I connected it to my own chat interface. So, I built out the UI, added in the streaming, and I want to run a task here to show you what this agent can do. So, here I have a couple of images in my desktop, and these are just pictures of receipts. I just Googled up some images of receipts and as you can see, they're not the highest quality, but there is information on these receipts. So, we're going to have this agent go ahead and look through this folder, analyze these receipts, and we're going to create a report in a markdown format. So, I'm going to tell this agent, process the receipts in the current folder, review each image and its contents, and create a markdown table with the information out of these receipts and put them into a receipts.md file. So, we're going to go ahead and generate this. Now, really quick, we're going to go back into our configuration settings, and I'm going to set permission mode to bypass permissions. And I'll also allow dangerously skip permissions and set that to true. And the reason why I'm doing this is because certain commands require us to approve things and interact with the UI. And I just want my agent to not have any guard rails and just be able to complete the task. So, I'm going to go ahead and just run this and we're just going to see how the agent streams things in and what the final results going to look like. So, let's take a look. So here I can see the agent's thinking. I can see that it's reading certain files. It's going through them and it's actually returning some messages as it's going through its reasoning process. Now I just did a little bit of configuration. I didn't fully build everything in, but this should give you an idea. So if we want to open these up, we can see the exact inputs that it ran. And these are those message streams that the agent has for us. Okay, so it looks like we have a final success message. We can see the write command. And it looks like it put it into the receipts.md folder. And if I open this up and take a look at it here, we see the final report. So I see a full table here. It looks like all the items are listed out and it actually organized it in a cool way. So I know that this is a simple example, but this just shows you the power of the agent SDK. The fact that it was able to call tools on its own. It was able to read the images without us having to custom build everything in and it generated a full report off of minimal configuration. And this is just a slight version of how I was able to make that full coding agent. So if I wanted to have it build me some kind of app, I can actually start that process from this minimal chat UI and it'll actually function. It'll write the code, run the commands that it needs to, and it has full control, the same control that claw code has itself. I want to switch gears for a second and talk about spinning up agents in a remote environment since this was a key part of the AI coding platform that I demoed at the start of the video. So, my idea was to have a tool that I don't just use on my machine, but something that I can take on the go. So, if I'm ever getting on a flight, I want to be able to spin up an agent, close my computer, and have that agent keep running and essentially offload the work from my computer to this environment. Well, as I was working on this, I actually stumbled upon the Warp Oz orchestration platform and it actually does pretty much everything that I need it to do. So, I want to show you this and I want to explain how this works. If you haven't heard of Warp, it's a fully native AI development environment, and they recently announced their orchestration platform for cloud agents called Oz. Now, Oz lets you spin up multiple cloud coding agents in these remote development environments. So, you fire off a prompt, an agent clones the linked up GitHub repo, it begins work remotely, and once it's done, that code is ready for you to review. And this can come in the form of a pull request or you can just bring that code down locally and continue working on it. Now to trigger these agents, we have multiple options. We can do this from the warp CLI. We can do this with integrations like Slack or Telegram. But the method that works perfectly for me is their TypeScript SDK. So this way I can actually build this and trigger an agent from my own UI. So I can do it programmatically and fire off these remote tasks anytime I need something sent off. So, I want to jump into the code and actually show you how I did this and how it all works. Okay, so jumping back into the code, the first thing we'll want to do is install the Warp Oz agent SDK. Once we have that installed, we're going to import the Oz API class, which is going to set up our client instance. And this is going to be our entry point. Now, there are three key methods that I needed in order to make this work within my own UI. So, first we have the run method, which essentially lets us fire off a run. So this is where we add in our prompt and it triggers that run in the remote environment. Then we have the list method which allows us to see our runs and this is crucial for being able to see that status. So if I want to see where a run is at in my dashboard, I'm able to list those out, see that current value. And last, we have the retrieve method, which just lets me get the details about a specific run. So these three methods give me pretty much everything that I need in order to control these runs from my own UI. So to spin up an agent, I'm just going to go ahead and trigger the agent run method. And I'm going to provide it with a prompt. So in my case, I'll just say create a proper readme file explaining this project and what it's about, and generate a pull request when it's done. Now, I can pass in more configuration settings like the model we want to use or the name of the run, a base prompt, but in this case, I'm just going to pass in the minimum settings. And at the bare minimum, we at least need an environment ID here. And this just tells the agent what environment we're working within. Now, we don't have an environment ID yet. So, what we're going to do is quickly create one. We'll grab that ID and we're going to pass this in here. So, there's multiple ways of creating environments. We can do it from the Warp CLI or we can do it from the Oz web platform. So, we're going to use a web platform just for simplicity. So, we can go to oz.warp.dev. And once you sign up for an account in there, you're going to see this section for environments. Now, I'm going to go ahead and create a new environment. We're going to give this a name. Now, for the repo, I already connected my GitHub account and I'm just going to work within this repository that I attached. So, you just want to give it some permissions. And for the Docker image name, well, I'm just going to click on suggest there. And that's just going to simplify everything. Okay. So, our environment's created. We can go ahead and copy this ID. And let's just paste it in. And that's all we need in order to fire off a run. So, let's just go ahead and test this. We're going to run this file and see what happens. So, when I fired that off, I can immediately see the run right here inside of my UI. So I can see the status of it. If I want to intervene, I can go ahead and open up the run, click on the session, and open this in my warp terminal. And I can keep adding prompts if I need to work within it. And this lets me interact with it. Now, when the run is done, we're going to see this pull request here. So if I open up this repo here, we see the pull request. And that run was completed. So it's pretty cool. Now it's ready for me to review and I can either work on this locally, accept the PR, and that work was done all in a remote environment through my agent. All right, so that's it for this mini crash course on the Cloud Agent SDK and the Warp Oz orchestration platform. Now, I know I only showed you two edge cases here, but I hope this gives you ideas of what's possible and what you can build with it because it is incredibly powerful, and I'm excited to see what you all come up with. If you have any ideas, share them down in the comments here. I would love to take a look at those. Oh, and there are 1,000 free warp credits available down in the video description. So, be sure to check that out if you want to try Warp. So, with that being said, that's it for the video.