Cursor Admin Guide

0:00 We'll get going. Great. So, welcome everybody to today's webinar on Cursor admin guide. Uh we're going to walk through a number of different topics here today, but our main focus is going to be how you can how we can support you rolling out Cursor at enterprise scale. Uh and so, this is going to cover a number of different topics, uh making sure that you're able to understand the analytics that Cursor provides in our admin dashboard, some different ideas for calculating uh ROI, which is top of mind

0:28 for everybody as AI as AI gets more and more adoption, uh and also helping to just manage users and uh and things like spend, security, um again, as you're rolling out Cursor across enterprise scale. As Noah mentioned, we do have a Q&A open in the chat, so please do uh put questions there, and there should be ample time at the end of the webinar for us to address those. And Noah, if you do, as we go along, see any themes that are emerging, uh we can uh also redirect and answer some of those in in

1:01 real time, if it if it makes sense. Awesome. So, we're going to start by talking about the admin journey, and again, some of the topics that we're going to cover today, things from security and enterprise readiness, how to measure and track adoption using some of the analytics tooling that is available in Cursor's uh in Cursor's dashboard, different ideas for measuring ROI, and then uh how you can continue to use Cursor to drive value across your

1:29 organization. So, let's start with enterprise readiness. We're not going to spend too much time here, but I think it's important to touch on uh as, you know, people expect certain um cer- certain compliance and privacy and security anytime that they roll uh roll out Cursor across their enterprise.

1:49 So, in an enterprise context, a few things that um that Cursor has put in place, there's going to be privacy mode on by default and it can't be turned off. We have zero data retention agreements in place with all of the major model and so, none of the code uh that you uh work with in Cursor or prompt to the models is going to be stored downstream, uh and it's never going to be used for training.

2:13 And then, all of the enterprise security and compliance regulations that you would expect, things like SOC 2 Type 2 attestations, single sign-on, MDM enforcement, Cursor supports as well. There's also a number of different features that are built into the product, things like hooks. Um this is a really, really popular feature and is something that we highly recommend uh enterprises take a look at and and use.

2:37 These are ways in which you can actually augment and customize Cursor's agent loop to ensure that uh the way that the agent works is going to conform to different security uh regulations, things like scanning for secrets before a prompt is actually submitted or executed on, doing checks and running scripts to make sure that um any particular prompt is going to be safe to execute. And uh so, there's a number of different ways in which you can can use and customize hooks, but we definitely

3:06 recommend taking a look at this because it can be a really helpful way to, again, customize Cursor's agent loop and execution uh as you as you roll it out. There's a special .cursorignore file that you can use to specify different files that you never want Cursor to index or be able to view.

3:24 And then, in the dashboard itself, there's audit logs that allow you as an administrator to track things like off- authentication, user changes, uh different settings, changes to your environment, so that you're always able to see what changes are happening to your Cursor environment and and who is doing what.

3:42 And then, finally, uh as AI usage is continuing to grow and and explode in really exciting ways, Cursor gives you the ability as an administrator uh to set spend alerts, group limits both at the individual, group, or team level, uh and so that you're never going to be surprised by by runaway spend. So, these are all again, a lot of the basic building blocks that enterprises expect when they do a large-scale rollout of Cursor to make sure that it's going to

4:09 be a secure and predictable and maintainable type of environment across large teams. There's a lot more documentation, certainly, and things that we can provide. So, if you do have questions on these areas, please do uh put them in the chat. Noah and I will answer them as we go. Um but now, I think we're going to shift into some of the more exciting ways in which an administrator can manage and work with Cursor. Things like, how do we measure adoption? How do we drive

4:34 adoption? And understand uh how Cursor is being used across large teams. Now, as Cursor's product has continued to expand and grow, we're continuing to add more and more analytics features into the product itself, into the admin dashboard, to help you as an admin understand how Cursor is being used across your team and organization. So, one example is you can see exactly how much uh how how much usage you're seeing across the different surfaces Cursor

5:07 provides. Most people are familiar with the Cursor desktop app or the IDE, but Cursor also has a new cloud agent platform. It we have a CLI, uh and some other ways in which you can work with uh with with Cursor across a lot of different surfaces, so we're able to meet developers exactly where they're at, and you as an administrator are able to easily see exactly how and where Cursor is being used across these different surfaces. You can also view usage by model. This

5:36 is a really important thing for administrators to understand exactly what tools, what models are being used through Cursor. Uh and there's a lot that we can dig into here to help understand exactly uh you know, how to or how it relates to to spend and token efficiency as different models have different strengths and different cost structures, and uh being able to view exactly how models are being used across your Cursor deployment uh is something that can be really helpful. So, we definitely

6:03 encourage you to understand or to look at uh the the team activity analytics that Cursor is continuing to enhance and provide in our admin dashboard. >> [gasps] >> We've also recently uh launched a a new product, BugBot, uh which is a a sorry, a code review agent um that we're excited to see getting a lot of adoption. And within Cursor's admin dashboard, there's a number of different things you can do. Again, from an analytical perspective, you can see how many PRs BugBot is operating on. You can see the

6:37 number of different issues that it's being resolved or that it's that it's resolving. We're very proud that BugBot seems to be a very, very high signal-to-noise type of product. Um and so, we're continuously seeing, like you see here on the screen, uh issue resolve percentages north of 70 and and 80%, and you'll be able to see this in uh your dashboard as well.

6:59 You can also see here at the bottom of the screen, I'm not sure if you can see my cursor, but this broken down by the type of incident, too, whether it's high severity, medium severity, or low severity. And you can also, in the admin dashboard, specify what repos you want BugBot to operate on, if you would like BugBot to auto-fix issues that it finds or just surface them to the team for extra manual review. So, there's a lot of different ways in which you can configure how BugBot operates, uh and

7:31 again, as you see here, the number of issues and the types of issues that it is resolving. I mentioned earlier that cloud agents is a very exciting new platform and surface for Cursor, and we're really, again, excited to see the internal and external adoption uh as we've rolled out this cloud or background agents product more broadly. As we've rolled it out, we've given you insight in the Cursor dashboard as well to see uh all of the analytics that we

7:58 provide, so you can see how cloud agents are being used across your organization. And just as we were talking about earlier, cloud agents can be deployed or activated or spawned from a number of different surfaces from Cursor web app, the IDE itself, things like Slack, and there's a number of different types of cloud agents that can be spawned. We were just looking at BugBot. We've also released a new automations product, which help you spawn cloud agents uh on a schedule or in response to certain

8:27 triggers. And these are all really helpful ways to help you understand what types of cloud agents are running, how they're being spawned, and what type of adoption you're seeing across your organization. We've also very recently uh expanded cloud agent support to run in your own environment in a more self-hosted way. So, this unlocks use cases for organizations that might have more stringent security requirements or compliance requirements, or might have internal package

8:56 registries and so forth that make it such that they need to run uh a a cloud agent in their own environment to make sure that it has access to your entire development system. And that's another really exciting thing that you can do as well. Again, there's all of the same admin controls in the Cursor desktop uh or sorry, in in the Cursor dashboard to configure uh so that you can turn on self-hosted cloud agents and see exactly how they're being used

9:22 across your organization. Moving on to some of the more advanced features, again, we're sticking on this theme of analytics cuz this is something that we see admins uh a having a lot of questions about, and it really helps understand exactly how Cursor is being used across large large organizations. So, Cursor's agent mode has a number of different modes. It's not just the the agent itself, but we also have a plan mode, which is really popular and we see

9:52 as a best practice for working with Cursor's agent. And you can see across your organization how many plans are being created. This can also be broken down by an individual user level, so you can see who across your organization is using plan mode most heavily.

10:07 And you can also see what specific models are being used to create plans. There's a lot of different best practices here, and we do have webinars that are actually uh specifically on this particular topic when it comes to how to best deploy models, but we do see a a best practice of using plan mode. So, we like to see these numbers be really really high. And these plans often are really good fit for state-of-the-art frontier very powerful models where you're able to

10:38 leverage in your plan building all of the best kind of reasoning capabilities from the best models across uh the the frontier labs. So, helpful to see insight into this into your or in your your cursor dashboard as well. Another thing that as agent capabilities and agent extensibility continues to grow and enhance, we're seeing really you know helpful analytics around the number of different skills or the number of different hooks or the number of different MCP uh servers that are being invoked. And

11:14 again, all of that information is surfaced to you in your admin dashboard. So, you can control skills, hooks, etc. You can push them out as a you know from a central uh admin perspective. You can push them out to your entire team. You can centrally administer different MCP integrations that you're setting up in Cursor and and make them available to your team. And you can see exactly how often these are being invoked. And this can be really really helpful to understand what

11:42 school what what skills rather what commands, what MCPs are seeing the most heavy usage. Uh and if there's particular gaps or things that are not being um not being utilized, it helps you get visibility and and start to figure out why. Just talking about this on the previous slide as well.

12:04 Again, sort of the same theme here, helping you understand all of the different integrations and agent extensibility features that are being utilized across your organization. Now, something that Cursor's also continuing to invest in is helping you understand what type of work is being done with Cursor's agents. And so, we have classifiers that are helpful to understand uh you know whether or or what percentage, what is the breakdown within your team of how much agent work

12:35 is being dedicated towards new feature development or bug fixes or ongoing maintenance work. You can also see a breakdown of you know the the percentage of uh agent modes, right? We talked about plan mode, ask mode, agent mode writing code, and you can see the intent distribution in your your admin dashboard.

12:57 And we're also able to break down on a by category basis the type of work that your team is asking agents to do. You're also able to gauge the task complexity and the prompt specificity. I'll I'll I'll highlight again prompt specificity is I think a particularly useful one because we see agents doing better quality work and more token and more cost-efficient work when the prompts are more specific and more detailed and more targeted. And so, if you do see in your admin

13:29 dashboard that you're seeing a higher percentage of maybe low or medium prompt specificity, that could be a signal that training could be beneficial, which our teams are happy to deliver, uh to make sure that you are prompting Cursor and and and using Cursor's agents in ways that are going to be the most token and spend efficient and get you the outcomes that you're you're looking for. So, there's a lot of really helpful information that is packaged into all of

13:58 these different um uh analytics features, and we definitely encourage you to to take a look at. Another thing that you'll see here on the left, task complexity. Um this is another good thing to correlate with the types of models that you're using. Cuz we tend we we see as a best practice that you want to make sure that you're using some of the most powerful models and the most expensive models for high high complexity work.

14:27 And it could be you know again something that could be worth looking into if you're seeing very high usages of you know very expensive models paired with tasks that are lower or trivial in terms of complexity. And so, making sure that you're choosing the right model for the job and understanding at an organizational perspective or in in aggregate can be really helpful to make sure that you're getting the most out of Cursor and and your agents.

15:00 This is again another relatively new feature that we are super excited about, which is a marketplace that we've released with this notion of plugins. And plugins are ways that you can package up different skills, rules, sub agents, MCPs from a number of different providers that are available in the Cursor marketplace. Any of these different third-party plugins they're reviewed by our team and so they're they're vetted and we're very very excited about the different ways in which you can bring these integrations

15:32 in a packaged plugin into your Cursor environment. And you can as with everything that we've talked about here so far, you can manage these centrally. So, you can make available these plugins to your team from a centralized place. Uh and and get the benefit of all of the skills, the sub agents, the MCP integrations uh to help you again extend your agent deployment and and make sure that you're going to be getting the most out of Cursor.

16:01 There's also you can see here down at the very bottom the ability for you to create your own private team marketplaces. And so, if there are specific internal tools, internal skills, internal MCPs that you have developed and you want to manage and deploy those in a centralized way, that's something that Cursor also supports. So, you can basically create a manifest, upload that into Cursor, and then pri- and then distribute private team marketplaces that are going to be unique for your

16:32 organization. So again, this is something that is a a relatively new feature to the Cursor platform and ecosystem and we've seen a lot of adoption around in recent weeks and months. And highly encourage you making use of the marketplace to to make sure that you're getting the most out of your Cursor deployment.

16:56 One thing that I'll highlight here, we talked about this a little bit back when we were reviewing cloud agents, um but there's also a new type of cloud agent called an automation. You'll see within the marketplace a number of different featured or templated automations that you can deploy with a single click. These are automations that often times are built and used by our team internally and you're able to customize these according to your own needs

17:25 or create your own from scratch. So, we definitely encourage you to to review these automations within the marketplace as well. All right. So, we've talked about a lot of the different ways in which you can monitor and and measure analytics and usage.

17:42 And one of the questions that Noah and myself and our team at Cursor often gets now is how can we actually start to measure ROI? And there's a number of different phases or there's sort of a journey for measuring ROI. Uh and this is you know changing quite rapidly as models are continuing to enhance their capabilities, AI usage continues to grow at sort of mind-boggling clips, which is which is really exciting, but it does bring in these questions of how can we

18:11 make sure that we're actually measuring ROI appropriately. So first, some of the things that we see is you know an anecdotal or qualitative type of uh type of ROI. Are engineers and developers are they excited to use these tools? Are they integrating them into their workflows? Are they finding specific ways in their workflows in which these these new tools are working for them? Do they feel more productive?

18:38 Are they moving faster? And once you have some initial signal, you can start to move into more quantitative ways of measuring productivity using Cursor and and AI generally, which is are we seeing are we actually shipping faster? Are PRs being created at a higher velocity?

18:56 How is code quality? Are we seeing a decrease, increase, or status quo in terms of the number of bugs that are being created? What percentage of code is being turned over? And one of the things that we're really excited to start to see happening more and more is as we get to phase four, they're actually being impact measured in real dollars and cents. So, we're starting to see teams realize that as AI is helping them collapse you know deployment of new products for

19:23 example, from things that previously might have taken weeks or months into you know a matter of days, uh they're actually able to start to see that they're pulling forward things from their product roadmap. They're pulling forward the revenue gains from those new products. And there's a number of different things in which you can start to measure the financial impact of of your AI deployment.

19:49 And that's something that again we're really excited to see because we want to make sure I think everybody wants to make sure that the tools that they're rolling out are actually helping their top and bottom lines. And so, these are the different phases. Again, it is a a journey, but as AI is getting more deeply and deeply embedded into the workflows of engineering teams, I think this is something to to shoot for is to understand what is the ultimate financial impact of the new

20:16 products, the the increased velocity that we're able to to see. So, we've talked about velocity and we've talked about quality. We also talked briefly just about translating those into actually understanding what is the financial impact, but these are a couple of different ways in which you can start to immediately quantify the ROI that you're getting from your deployment of AI.

20:43 And one of the things that we're starting to see is that with Cursor, there are really meaningful and noticeable impacts and increases in both the velocity and value of of your your deployment. So, there are, you know, velocity increases upwards of 30% We're actually seeing these numbers continue to increase as the power of of models increases in in turn and you're able to see that the value because of the the additional model the step change increase in model capability and as AI workflows get

21:22 more and more embedded for engineering teams, that the value is is continuing to increase in turn, which is super exciting. So, couple of different tips and tricks as a Cursor admin to identify opportunities to continue to get the most out of your Cursor deployment. We see that you know, within organizations, there tends to be a distribution that emerges among specific engineers or teams in your organization that are using AI the most heavily. And so, when you do start

21:56 to see in your dashboard requests that are very high volume usage from individuals or teams and also a you know, a high acceptance rate, these are power power users within your team that you can actually start to learn from. Figure out what's working for them, how they're using Cursor, how they're using AI in their workflows so that you can distribute those learnings across the team. When you're seeing you know, high volume of requests and maybe a low acceptance rate of code committed, that can signal

22:26 an opportunity for training enablement either from your power users or from folks like Noah and myself. And then when you do see certain teams or certain individuals with a lower request volume, that can be a signal to to just drive more adoption and awareness, which again, we're very very happy to help you. But all of these this usage data and this acceptance data is going to be available in your Cursor dashboard. So, it's going to be readily available to you and you can

22:57 continue to monitor usage across your team. There's a number of different resources that we provide here and so, do consult Cursor docs, reach out to your account team, reach out to folks like Noah and myself so that we can continue to help you get the most out of of Cursor. And I think we've got a little bit of time left in the session. That concludes the slides that I'm providing here, but Noah, I'll kick it over to you if you want to surface any questions that

23:32 have come through in the chat that we can answer. Yeah, there's one in here about the APIs and what metric data is available there. Don't know if you want to talk through the difference between the dashboard and the APIs and how people can think about leveraging either one of those. Yeah, great question.

23:52 And that would have actually been a great slide to include. So, we ran through a lot of the high level analytics all of those different images that I shared over the course of the presentation are images and and analytics that are immediately available in the Cursor dashboard. But as the question indicates and as Noah was just alluding to, Cursor does have an analytics API and an admin API that you're able to use to pull all of that information. You can build your own reports with it. You can integrate it

24:21 into your own analytic systems. There's a wealth of data available through those APIs and so, I would definitely encourage you to to make use of it, build your own custom reporting and take a look at what data is available in our API documentation. That's a great question and a great call out.

24:42 There's a another question in here about monitoring chat history of developers. Mhm. Do you want to speak on that a little bit? Yeah, so this is something that you know, there's this notion of of shared knowledge or shared transcripts that as an administrator you can enable. And there are places in the Cursor dashboard where you can you know, chats or or knowledge or transcripts can be flagged to be shared at a team level. This can be helpful to again, you know, especially with a power user maybe learn

25:18 exactly how they're working with their agents, how they're prompting their agents. again, you can you can toggle and and view in the admin settings. Seeing Noah providing links for admin analytics and in the API, which is great. Yeah, um there's one in here that I'm going to type up an answer to, but somebody asked to optimize our pooled enterprise credits, does Cursor recommend using model access controls to restrict certain expensive frontier models to specific user power user groups or is

25:57 it better to rely on auto model router to manage these costs centrally? Yeah. This is this is a really good question and I think it's going to be you know, it'll be really dependent upon your particular use case and your environment. I would definitely encourage you to reach out to your account team you know, at Cursor to help with maybe custom trainings. Um I think that you know, Cursor Cursor does provide our admins with the ability to restrict model access limit model access. And one thing that we do see is

26:36 you know, cost can be a consideration there. Um Now, there's also another side of the coin is where you know, teams do want to make sure that they're or admins do want to make sure their teams do have access to the best models. And so, there is a push and pull between you know, what is the actual ROI or the the balance between the the cost of the state of the art frontier models, which do tend to be more expensive and making sure your team is able to harness their

27:06 power. So, I think actually that's where some of the analytics that we did talk about where you're able to see what what you know, is the distribution of models that are being used across your team and then you can also see as we we did look at you know, what is the the nature of the tasks that you're giving to your agent? Are they high complexity or low complexity? Are they um are are the prompts that you're giving across the team highly specific or

27:35 minimally specific? And those are things that can also be you know, tuned. You can train the team, we can help train your team to make sure that you know, when you are using a state of the art frontier model, you're doing so with high prompt specificity and we're ideally going to be leveraging those both in plan mode where token use will be more efficient.

27:55 Um and on tasks that require a state of the art frontier model. Cuz it often times won't make sense and you can start to run into issues if we're using really really expensive models for trivial tasks. Um one of the themes that you'll continue to hear us talk about is making sure that you're using the right tool for the job and I think that's something that the analytics we've talked about here can provide insight on. So, this is a long way of answering the question do we recommend you know, using

28:22 these model allow lists? Again, I think it's going to really depend on your individual use cases and there's a lot of other things to consider, but yeah, there's a balance between making sure your team has access to the best models and making sure that you're also controlling spend appropriately.

28:44 Noah, also feel free if you do have additional insights or thoughts um based on your work and in particular I'd say anything that I say that you might disagree with too, feel free to to chime in too, but um yeah, I think that these are great questions so far. Yeah, and to answer there's another question in here um that says is there an MCP server we can use instead of the APIs? There's not today. All of that is called through the APIs, but you can have even Cursor's platform call that with some of your

29:24 authentication information there. >> [clears throat] >> I see we have a feature request in here, which I'll get that submitted to our product team. Um I answered this question through the chat as well, but just to cover it live. Um these analytics can be made available to your team overall. Uh, and if that's something that you want to restrict control to, uh, in the team settings in the admin page, there is going to be a toggle that is titled restrict analytics

30:02 to admins. And if you turn that on, um, your team will not have access to the analytics. It can just be available to you all as an admin. This is something that you want to democratize across them, uh, you can always change the configuration there and allow other people across the team to see the analytics, too.

30:39 Um, there's another question in here, Tim, that just came across. Yep. Says, can you explain the automation and setup or provide URL to what is best practice practices we can use for context caching enforcement ensure 90% more efficient cash reads are utilized for repeated repository indexing.

31:04 Surgical reference audits and flags for wasteful prompts that scan the entire repo instead of at file references to reduce token bloat. Yep. Yeah, there's a there's a lot here, um, and certainly we we can definitely provide URLs for uh, automations setup, cloud agents to make sure that you are able to, uh, you know, have all of the relevant documentation. Uh, Noah, if you're able to to drop, uh, links to, uh, to automations documentations in the

31:37 in the chat, we can certainly, uh, point folks to that, um, after the the webinar. Uh, there there is a lot here, um, like wasteful prompts to scan the entire repo and using file references to reduce token bloat. Yeah, I think that these are, I mean, and we talked about this a little bit, um, making sure that, you know, your teams are using, uh, and prompting kind of according to to best practices. Um, it's where those analytics around very high

32:09 specific prompts versus lower specific prompts can be helpful for you to gain insight into those types of things, uh, and help you identify areas to further train the team. Um, and it is going to be really important because it is going to reduce, uh, token usage when you are providing higher specific prompts and, uh, just reducing, like the more vague the prompt, um, the more research and uh, and lifting you're asking the agent to go do, and that's going to often lead to more token consumption.

32:38 Um, it's also going to lead in in a lot of cases to lower quality output relative to a a high highly specific prompt. So, these are all good things for you to be able to see in your analytics dashboard. Um, model access restrictions using enterprise controls to disable thinking mode. Yeah, for, um, there's actually some some interesting work, uh, too, that Cursor's doing in this area, uh, where we're continuing to invest in making sure that you as an administrator are going

33:07 to have a lot more flexibility in terms of what features are are rolled out to which sort of product groups or which specific teams. Um, this is a a very active area of investment for us, and so if you do have specific feature requests, uh, please do reach out to your account team. Um, we do a lot of analytical work in Cursor.

33:37 Like a way to isolate that work specifically for analytics oriented views. Um, Now, I'm not sure if you I'm not 100% sure uh, what this question is asking, I but I I do think that this could be another good plug for, you know, the analytics APIs that we were, uh, referencing earlier, where if you are doing a lot of analytical work or you do want to access the analytics that Cursor makes available to you, um, would highly encourage you to use that API, um,

34:17 and build with that API in Cursor itself, cuz that's going to help you get uh, do a lot of, um, a lot of great analytics. How does Cursor calculate requests? Um, yeah, so this is, um, going to relate to some recent changes. I think that might be out of the scope of this particular webinar, but it's going to there there has been a recent change, uh, into the way in which, um, requests, and this is for legacy, uh, legacy request-based plans, um, Noah, it seems like you're going to

35:03 answer that live if you wanted to take a stab at it. Yeah, um, so at a high level, uh, depending on which model is used, that will depend on how many requests are used. Uh, for um, some models that don't require max mode, that usually uses one or two requests. For models that do require max mode, uh, essentially what happens there, um, is we take the API cost and map that back to the number of request that it would, um, count for. Believe request

35:39 essentially equate to roughly 8 cents, um, there, and so that's where you get some of those like 30 request uh, chats is it's likely a max mode model that is being mapped back to the, uh, kind of associated value of a request there. All right. Um, well, we are approaching time. I think, first of all, just want to thank everybody for, uh, the questions that they've, uh, provided into the chat.

36:23 Thank you, Noah, for helping moderate and provide commentary here. Um, we are always running more and more of these workshops, and so we encourage you to continue to, uh, go to cursor.com/workshops to see, uh, additional upcoming webinars with a lot of different topics across the Cursor ecosystem. Reach out to, uh, enterprise@cursor.com or your existing account team with any additional questions that you have. Um, and yeah, really appreciate the global audience that we've had today. Uh, thank

36:54 you, everybody, for joining, and we hope you have a great rest of your Tuesday.