Cursor Agent SDK

0:00 By way of introduction, you've got two folks from Cursor on the line today. So, my name is Tim. I'm with Cursor's field engineering team. And uh I'll be walking through our presentation. I've also got my colleague Jake. Jake, I'll let you introduce yourself. Good to be here. I'm Jake on the AI deployment team. Be here monitoring chat and doing my best to answer all your questions. So, feel free to drop your questions in there.

0:23 Cool. Thanks, Jake. Um awesome. So, we're going to jump into walking through the Cursor SDK. And uh one thing that I'm going to note just from a logistical standpoint is we've got both a chat. We should also have a Q&A feature open. Um it's a lot easier for us to track questions and respond in line when we're using the Q&A feature. So, please do uh if you're able to, put your questions into the Q&A. And uh as opposed to the chat, and that'll be easier for Jake and myself to track

1:05 and monitor as we go. Um awesome. So, I'm going to pull up my screen and then we are going to get going. Okay. Jake, I'm going to confirm with you that you can see my screen where we've got the slides for the Cursor Agent SDK. We are good to go. Awesome. [snorts] So, I'm going to jump right in uh to our presentation. Uh like I said, please put questions in the Q&A and we'll answer them as we we go along. Um we are super excited to have released this SDK. It's something that uh our customers have been asking for for a

1:53 long while. And we have finally built and and released it and it's been really exciting to see the the things that customers are building on top of the the the cursor agent SDK. So, um what is the SDK? It basically gives you access to build on top of cursors agents. You can script them uh into to any TypeScript service. You can build your own custom agents for your own uh purposes and you can bring them pretty much anywhere. So, inside of scripts, CI

2:25 pipelines, uh back-end services, and and so forth. Today, what we're going to cover is exactly what the SDK is, why we built it, uh how to build with it. We'll go through a couple of examples um and then talk with any time allowing about some of the guardrails or things that you should be uh aware of, but by the end of the session, our hope is that you come away from this presentation with a better understanding of exactly what the SDK is, uh why we have built it, so what kinds of use cases does it unlock, and

2:59 and how you can go ahead and and build with it. And like we said a few times now, please put questions in the in the chat because uh we want to make sure that this is as helpful as possible for everybody here. Um so so what is the SDK? Well, again, it's giving you access to the same cursor agent uh that you're used to working with today, maybe in the CLI or in uh the the cursor desktop app or in the cursor cloud. It's a single interface that you can use with access

3:28 to all of the cursor runtimes. So, you can run you can build and run uh local agents with the SDK. You can build and run cloud agents with the SDK. You can also build and run self-hosted cloud agents with the SDK as well. A couple of core concepts to be aware of uh when you're starting to to learn the SDK and work with it. The first is, well, what is an agent or or you know, the the agent that's going to uh the object that's going to be uh returned when you when you create and initialize an agent with the SDK.

4:00 Uh what's going to happen or or what is the interface going to look like when you actually start to submit prompts uh and and actually run the SDK. And then how are you going to work with those messages and and streams that are emitted? Why did we build this? So, there's a lot of things uh certainly more than than three things that the the SDK can give you here. Um but we wanted to give customers even more flexibility to extend the the cursor agent platform into any direction

4:31 that they'd like. We've also seen uh of late a pretty exciting explosion of use cases and different types of teams that are using cursor for all different types of purposes. So, these are agents that are not even specific to coding, but maybe they're uh specific to things like designing, planning, doing QA. Uh so, above and beyond the typical type of agent experience that you're going to get in the cursor CLI or the cursor desktop app or cursor cloud. Uh and we wanted to also give customers the ability to

5:02 customize uh the behavior. So, cursor cursor is obviously giving you today a lot of shrink-wrapped type of agents and products off the shelf. So, you're probably familiar with things like bug bot, which is cursor's coding review agent. We've released another dedicated security review agent uh that hopefully a bunch of you have started to to build and and test with.

5:24 But, we also know that teams have sometimes their own use cases, their own specific configurations or customizations that they want to be able to control. And that's a great use case for the SDK cuz it gives you exactly that. You can extend and customize cursor's agents to to your heart's content.

5:45 So, a couple of different uh ways in which you can actually start to to work with this. Um you can have an agent.prompt, which is going to send a message to an agent. Just going to be a kind of fire and forget, you know, one one shot a a particular uh particular prompt. You can create a durable agent, which will then you can continue to send uh send prompts to and have more back and forth types of of conversations. And if you need to or want to, you can always use the agent.resume command to or or method uh

6:18 to to pick up existing agents, so you're able to preserve context across um uh different surfaces or different uh types of use cases. If you if you need something that um you you want to preserve contact context around. Uh I mentioned this earlier, but again, it's going to be a single interface that the Cursor SDK provides that allows you to interact both locally or in the cloud. And when we get into the cloud, it's either in the Cursor cloud or if

6:48 you're using self-hosted cloud agents, you can do that um you you can do that as well. And so, the idea here is that there's going to be a pretty uh you know, straightforward surface area of the API that you need to learn. And it is runtime agnostic. So, you specify if you want this to run locally or if you want to to run this in the cloud. There's often times going to be use cases for both. Um and it again, it's going to be the exact same SDK and you just pass in the runtime that

7:18 you want to specify. Cool. So, now we can actually start to build with the Cursor SDK. And there's a couple of different examples that uh that we can we can build with. So, the first thing that I'm going to do, I'm actually going to at this point stop sharing my slides. And >> [sighs] >> we are going to jump into uh Cursor itself. And we'll be able to get a look at some of the different types of agents that we can uh that we can build with. So, I just had Oh, hit the share.

8:00 Okay. So, now we're looking at my cursor interface. One quick side note, for those who are used to working inside of the the the cursor IDE, you'll notice that this is a slightly newer interface. We released this about a month ago, so um it is still relatively new, but this is what we're calling cursor three, where we're building in or stripping away rather some of the the legacy IDE features that are lesser used in a more agent-forward world and making it a lot

8:28 easier for folks to build in a more agent-directed way, which is more and more common, especially over the last four or five months as models have become quite a bit more powerful. Um So, here I'm in an e-commerce repo that I have been building in. And one of the really nice things about cursors cursor's SDK is we actually have a built-in skill into cursor itself. So, I'm going to invoke this skill using that {slash} command. And typing SDK is going to allow me This

9:02 is the recommended way or one of the recommended ways in which we can start to build with cursor. So, I'm going to invoke this command and say build a um let's say build an agent using the cursor agent SDK, which is you already know how to do because I'm using that that skill directly. So, it'll know exactly where to look and what to do here.

9:28 But, I'm going to say build an agent using the cursor SDK to summarize this repository. So, a really simple and straightforward example here. Um but again, just giving you familiarity with that built-in SDK command, which is a really, really helpful skill that Cursor provides off the shelf that allows you to do uh exactly this. So, let's see. It looks like there our agent is getting to work.

9:54 Again, this is pretty simple um and pretty straightforward. It should be familiar. Just we're we're prompting an agent and uh the agent itself is going to use the SDK to build a a dedicated agent. So, a little meta in in this way. And um so, yeah, we're going to go ahead and build this agent. While I do that, now that we're about 13 minutes in and our agent gets to work, I'm going to pause. Jake, any initial questions? I know it's just we provide an initial

10:22 introduction here. There are. Um yeah, so we'll kick it off. Will we support more programming languages? Yes. Uh we my my belief uh and I hope I'm not speaking two out of turn here, um but I believe I have seen a roadmap item to add Python support. Um if there are other languages that you would like to see, please do put them uh in the chat. We'd love to to gather that as feedback to to send to our product and engineering team. But I believe that we will have support uh

10:55 forthcoming for additional languages. Awesome. We got a few more if we have time, otherwise we can Looks like we're still working on hitting there. Yeah, go for it. How useful How is this useful for non-coding use cases? Yeah. So, we're I'm going to get into this a little bit um in the uh in in some of the examples that we've built out. Um but the way that I like to think about it is you basically now have access to uh to the Cursor agent platform to build purpose-built agents for anything that you'd like. So,

11:31 you know, you're you're probably uh you you you could build a coding agent with the the SDK. Um you may not need to, of course, because that's what the the primary uh you know, agents inside of cursor are going to give you uh off the shelf here. Um but we're seeing some pretty interesting use cases for uh teams that want to build a dedicated SDK for things like planning, things like designing, uh things like doing QA. Um there's also uh you know, some additional flexibility

12:06 that the SDK might provide where, you know, you can replace some or or or embed rather, I should say, embed the SDK into existing workflows that you have. And so, uh you're you're you're only constrained, I think, by your imagination in terms of what types of agents do you want to build? Um I'll I'll I'll talk briefly about some internal examples uh that I've seen a couple of my my colleagues build uh using the SDK. Um where they've built uh basically some um

12:39 >> [clears throat] >> uh like a a desktop app um where it's just like a character on their desktop that lives there throughout the day that they can chat with uh for and this is going to be used across all different teams internally. So, um we've had folks in our finance team use the SDK to build uh you know, dedicated agents for their specific workflows, which again are going to be more focused on finance use cases relative to to coding. We have our

13:06 customer success and support teams who have built similar types of desktop apps where they can basically just click a button or click an icon and start to interact with an agent that is built on top of the agent SDK, uh but we've given it access to a lot of our internal tooling so that they can ask questions about uh the their customers and their accounts um to help streamline a lot of their workflows and you can think of pretty much any team inside of cursor.

13:35 There's there's use cases that we've built for for their workflows and their teams that extend far beyond coding. So again, there that's a higher level overview, but I'd say we're seeing use cases again across finance, across legal, across support, across design and so it's really nice to be able to customize agents for those those kinds of use cases.

14:01 I'll answer what I can in the chat. We'll jump back in in a bit. Yeah, sounds good. Okay, so I've gone ahead and I have built out built out in >> [clears throat] >> repo summarizer and again it was as simple as invoking this SDK command. You're seeing the ways in which now I can start to use this. So I'm going to go ahead into this particular file here which is going to pop out and you can see here the the code that if you ever want to just drop down into or understand exactly what it's doing and

14:35 how it's working, you can see here. We are importing our agent object or entity from the cursor SDK package and it's going to go through and give a nice summarization of our repository here and you can see my agent has helpfully given me some instructions for actually running this. I'm not going to run this exactly right now to need to go grab a cursor API key that I could embed into this particular repository, but then I can go ahead and say hey run my run my new feature agent or run

15:13 my summarization agent and then it'll spit back a summary of this repository. A pretty elementary example here. But hopefully it helps you get the wheels turning in terms of how I can build and start to run these agents. In this case, locally. Uh, we could also just as easily do this, um, do this in the cloud. Now, let me pause one more time my share and jump back to the website because there's another or jump back to my browser, rather, because

15:41 there are some exciting use cases, um, that we've built out and that are a little bit more interesting than just a repo summarizer. So, come here and share. Open up my Chrome. And, uh, what we're looking at here is something I again built by another one of my colleagues, which is this, uh, this notion of a cursor foundry. You can think of this as a software factory. And this, I think, also alludes to one of the earlier questions around what types

16:13 of use cases can I build out with this, um, this SDK. And so, here in the uh, what with this foundry entails is, you can see here, um, we've actually built out this notion of a software assembly line. And, uh, this is very much a prototype, uh, kind of a proof of concept. Um, but what it does is it allows us, because we have this SDK, we can customize agents for all of these different types of purposes. So, we've got a planning agent, uh, which when we go ahead and

16:47 launch a particular type of feature, it's going to go ahead and build out a plan, uh, for a, uh, you know, whatever the the kind of case is, uh, that we that we want to build. Um, I can go ahead and launch a bunch of different agents. And again, this is going to use the SDK under the hood, uh, and all of these these agents are built, um, in a uh, [clears throat] in in a manner where they are specialized with a specific purpose. Um, and so, what we can do here now is stitch together this assembly

17:18 line. So, we've got planning agents, we've got design agents, we've got development agents, QA agents, and so forth. And as we continue to build, they'll be able to hand off to each other. And so, we can start to automate or stitch together and have these nice clean hand-offs from a plan written by a planning agent to a design or a designing agent that's going to go ahead go ahead and build out a design and then pass it off to a development agent and ultimately QA. And so, again,

17:50 this is just one type of agent that you can one type of assembly line or ways in which you can have multiple agents that are going to work together in these coordinated ways. I'll also click over here. One thing that is really helpful and I would encourage everybody to check out. This should be linked in our documentation. Jake and I can also put this into the chat as well. But, we have this cookbook, which is something that you can go ahead you can actually clone this repository, you can

18:19 play with these examples, and you can see some of the different types of use cases for for using the agent SDK. Again, some of my one of my favorites is this this Kanban board where we're able to basically see and get kind of status updates on a number of different cloud agents, group them in interesting ways, and build our own custom experiences, dashboards, reports, alerting, things like this in in using these new UIs. The the final use case that I'll share

18:56 here before opening it up for for questions is I also like to think of the SDK as enabling you to build custom UIs around the the cloud agent ecosystem in cursor. So, um um you you know, you're probably used to coming to cursor.com/agents or the cursor desktop app and firing off a bunch of cloud agents for, you know, whatever your your prompting, whatever uh kinds of purposes you you have in mind.

19:29 Um but because you've got this agent SDK now, you can build custom UIs that are going to wrap this. So, for example, uh one of the things that I do as part of my work on the field engineering team is when we're working with new customers who are evaluating cursor, uh we get uh a number of different um you know, security questionnaires. And so, we've built an internal app on top of the SDK that allows us to have a nice clean interface for managing these

19:58 types of questionnaires. So, anytime I get a new questionnaire, I can take the document, I upload it uh to our new to to this new UI, it will spawn a cloud agent to go and um you know, basically do a first pass on on a questionnaire uh using our things like our security documentation, our privacy policies, all of the documentation that lives in our trust center.

20:23 So, we can actually have these agents get um pretty far by um >> [clears throat] >> you know, just by having uh them do the initial pass to respond to uh you know, a questionnaire really of of any kind. I'm equipping the the cloud the the agent that's built with the SDK with all of that internal tooling or internal documentation, rather. Uh and it's uh yeah, it's something that saves us a ton of time and and that's another nice interesting use case for the SDK, which is again to to basically build um new

20:57 custom UIs and then under the hood, you're you know, we you're able to embed the SDK uh to have cursor agents doing whatever kind of task uh you need or want them to do. So, um that's another another type of use case that we're seeing quite a bit of. Uh I'd seen some hand raise some hands raised. Uh and Jake, I'm imagining that there's some questions in the chat. So, I'll pause and uh yeah, open it up for for questions here.

21:24 Love it. We have probably more questions than we're going to get to in the next 5 minutes, but the um engagement is amazing and want to make sure there's a way that we can get everyone's questions answered at some point. Um is there public roadmap feedback mechanism? That's a good so with this topic in particular. Yeah. Um I would say to reach out to your Cursor account team. So, the folks like uh like Jake, you know, if you if you're working with somebody on the Cursor account

21:51 team, please reach out there. Uh we've also got an active forum um and and support channels where you can uh can send those along, too. Um we also have our product managers. You'll often find them on uh you know, in different forums like uh like Twitter, for example. Uh and you can go ahead and um they're very responsive uh to uh to getting feedback and seeing what types of use cases are out there. So, there's a lot of different ways to reach us um and I'd

22:18 encourage you to to choose whichever is going to be easiest for you. Awesome. Um from Kamal here for long-running agents, can we cache the context to reuse tokens using sub-agents? Yes. Uh so, the the uh the SDK does support uh sub-agents. You just you'll you'll pass that into your um agent.create definition uh when you when you create the the agent, you can, you know, have specific named sub-agents with specific uh types of things and that can be a great way to further customize uh the the agent's behavior.

22:55 Great. And from Peter, as an enterprise customer, do you need a Cursor seat license to use to get a token and use the SDK. Uh you should not need an enterprise license. Yeah, you you can um if you're on a personal plan or uh whatever the case may be, you you should be able to go and generate a key that allows you to use this um whether you're an enterprise customer or not. Yeah, and I and I think I don't believe you actually need a seat either. So, or a seat for the So, yeah.

23:25 Yeah, if you are in an enterprise team, I again, I think that will depend on your specific uh contract, but it should be just drawing from the team's uh pool pooled usage there, but um you can triple-check that with your account manager. And then, can I programmatically inject custom skills or MCP servers into an agent instance at runtime? At runtime, yes. So, you I mean, you can um instantiate with uh or you can inject inline um with uh in a in agent with MCP uh servers.

24:04 Uh I will do a quick check on uh skills. It should be um quite straightforward to do. And um the idea here is that again, it's going to give you access to the full agent loop, the full uh agent runtime, and so um you should be able to uh to to configure this again with hooks, uh skills, MCP servers, uh absolutely.

24:30 Yeah, and it sounds like those are there's a lot of questions in the similar vein. Another one from Danny, are there plans to fully integrate the MCP protocol directly into the Chorus SDK, allowing agents we build today to seamlessly interact with external enterprise data sources or third-party dev tools? Yeah, we should be able to Let me see.

24:50 And again, um there might be some specifics here that you are uh if I'm not you know, fully fully grokking your uh your uh, your your question. Um, but and I'll I'll uh, we'll we'll put this in the chat as well. But, um, just to show you our in in our documentation we specify all of the different ways in which you can uh, load in uh, things like MCP servers uh, into your into your cloud agents. Or sorry, into your agents that you're building with the SDK.

25:27 And here's a good one from from Deborah just to kind of help establish a baseline for everyone. She asked or they ask what makes the SDK different than just having a skill or command to reference? Yeah. It's a good question and I think that the SDK just gives you even more flexibility and even more uh, customizability. It also allows you to bring the cursor agent experience anywhere. So, if you're using a skill, you're using that in the context of the cursor desktop app, the cursor CLI, you

25:58 know, cursor.com/agents if you're using our cloud agents. And so, it will um, be being able to or or having the SDK allows you to bring the the uh, the cursor agent pretty much anywhere. It allows you to to script on top of it uh, and do things in a more custom custom way. Um, and uh, build agents for your for your own purposes. So, in that way I'd say it's a you know, super set of uh, just a a skill which you're going to be using inside of a cursor surface.

26:32 Awesome. Yeah, and it seems like there's a lot of questions in that vein as well. So, hopefully we we we ticked a couple of those off. Um, yeah. I'm just going to go through cuz I know that we are at time. Um, again, we have a short session. We'll likely schedule another one of these uh, and uh, and talk to our team about doing one for even longer cuz I know that there's a ton of uh things here, but uh assuming multi-agent, not just composer, yes, you have the exact same set of model neutrality and optionality that uh

27:00 you get with Cursor. So, you can specify agents in the SDK to run with any model. Um and uh I've never heard of /sdk before. Can you summarize this? Yes. Um /sdk is uh probably the recommended way or one of the recommended ways inside of Cursor to start building with the the agent SDK. So, I definitely recommend using that command. Um There's Wow, so many awesome questions. Is this Is this recorded? Yes, it should be recorded. Um and our team will likely uh disseminate this out.

27:36 Um Uh from Patrick, I may have missed this. Is agent accessible via any UI, via any API call? So, there are a couple different ways, Patrick, in which you can do this again. Um with the agent SDK specifically, is a really nice way where you can just wrap the Cursor a Cursor agent with a new with your own custom UI. We also do have our own REST API. So, depending upon your use case, you might reach for the REST API, you might reach for the agent SDK. Um but we we do

28:06 support both. Um Will the agent SDK help next year the Utah Mammoth win the Stanley Cup? Can't guarantee that. >> [laughter] >> One question. Uh can I easily convert existing sub-agents into SDK agents? Uh this is from Bobby. Good good question, Bobby. Um I don't know that there's any built-in mechanisms like or a command, but that would be a fun skill to write uh to basically say, "Hey, I want to create a skill to take a sub-agent definition and then turn that into an agent SDK." Um that's something that we

28:41 could uh we could also do. Um Uh I see a couple more questions. I think we answered that. What makes it different from a scalar command? Added flexibility. Uh your your ability to script an agent and bring it anywhere uh for any specific type of use case. So, just above and beyond a scalar command, it's just more flexibility in terms of what it can do and where you can use it.

29:12 Um cool. There's a lot more questions. I think Jake will be able to to capture these and and uh we'll talk to our team about uh cuz they're again a shorter session today and would love to get all of these answered. So, um uh thank you for joining. Uh please look for more SDK focused workshops.

29:31 We'll definitely run them. And uh yeah, very appreciate the questions. We'll try and capture these and provide additional detail as a follow-up. But, yeah, thank you all. Thank you all for joining. Thanks, all. Take care.