An AI Future Without the Lock-In — Remy Guercio, Tailscale

Hello everyone. Thank you for Hello everyone. Thank you for joining me joining me joining me for my last appearance on for my last appearance on for my last appearance on this stage at this year's AI Engineer World's this stage at this year's AI Engineer World's this stage at this year's AI Engineer World's Fair. So, Fair. So, Fair. So, yeah, thank you for being here. yeah, thank you for being here. yeah, thank you for being here. My name is Remy. I My name is Remy. I My name is Remy. I work in the work in the work in the strategic strategic strategic projects team at Tailscale. And, more projects team at Tailscale. And, more projects team at Tailscale. And, more specifically, I'm specifically, I'm specifically, I'm working on our AI working on our AI gateway, Aperture. gateway, Aperture. gateway, Aperture. Today I want to talk a little bit Today I want to talk a little bit Today I want to talk a little bit about what it about what it about what it means or what it means or what it means or what it takes to takes to takes to build the future of build the future of vendor-neutral AI. I vendor-neutral AI. I think we've seen a think we've seen a think we've seen a lot of that kind of lot of that kind of lot of that kind of bonding over the bonding over the bonding over the last 12 months. last 12 months. last 12 months. If If If you look briefly at the last you look briefly at the last you look briefly at the last 12 months, I think they 12 months, I think they 12 months, I think they can be described with the phrase can be described with the phrase can be described with the phrase “ “ “ token maximization.” And I don't mean token maximization.” And I don't mean token maximization.” And I don't mean that in the that in the that in the negative sense negative sense negative sense that you might hear outside of the AI that you might hear outside of the AI enthusiast community. I enthusiast community. I mean, a year of trying to mean, a year of trying to mean, a year of trying to figure out what you can figure out what you can figure out what you can do with do with do with context windows context windows context windows for a million tokens and for a million tokens and for a million tokens and the ability to the ability to the ability to get context from get context from get context from everywhere, right? everywhere, right? Not only because Not only because MCP is everywhere, but also MCP is everywhere, but also MCP is everywhere, but also because of the because of the because of the programming tools, programming tools, programming tools, right? Which are great at right? Which are great at right? Which are great at finding context, finding context, finding context, using the CLI, using the CLI, using the CLI, passing that passing that passing that context into the model, context into the model, context into the model, injecting it into everything injecting it into everything injecting it into everything necessary, and moving necessary, and moving necessary, and moving on. Isn't that right? That on. Isn't that right? That on. Isn't that right? That was pretty cool, was pretty cool, was pretty cool, right? With all of this, right? With all of this, right? With all of this, we've really made we've really made we've really made full-fledged full-fledged full-fledged agent agent agent programming possible, right?

programming possible, right? So, that's pretty cool. So, that's pretty cool. We will see not only We will see not only We will see not only agent agent agent programming, but also programming, but also programming, but also other agent use cases other agent use cases other agent use cases that will continue to that will continue to that will continue to gain momentum as gain momentum as gain momentum as we develop. we develop. So, So, token maximization is cool, token maximization is cool, token maximization is cool, but it has its but it has its but it has its problems. Isn't that right? problems. Isn't that right? problems. Isn't that right? Obvious problems, Obvious problems, Obvious problems, right? It's very right? It's very right? It's very expensive, isn't it? This is an expensive, isn't it? This is an expensive, isn't it? This is an extremely expensive extremely expensive extremely expensive undertaking. I think the last undertaking. I think the last undertaking. I think the last 12 months have been 12 months have been 12 months have been about, "Hey, we have an about, "Hey, we have an about, "Hey, we have an unlimited budget unlimited budget unlimited budget for AI." But in the last for AI." But in the last for AI." But in the last few months, that few months, that few months, that has changed in what it has changed in what it has changed in what it means to use means to use means to use artificial intelligence. artificial intelligence. artificial intelligence. Well, it's also, I Well, it's also, I Well, it's also, I mean, you know, it's mean, you know, it's mean, you know, it's expensive if you expensive if you expensive if you do it right, do it right, do it right, right? But it also, by right? But it also, by right? But it also, by default, default, default, incentivizes what I would incentivizes what I would incentivizes what I would call call call token waste. I, well, I'm token waste. I, well, I'm token waste. I, well, I'm working on an AI gateway. working on an AI gateway. working on an AI gateway. We use We use We use our AI gateway. I see all our AI gateway. I see all usage data through usage data through this gateway. You wouldn't this gateway. You wouldn't this gateway. You wouldn't believe how many believe how many believe how many sessions last hundreds sessions last hundreds sessions last hundreds or thousands of or thousands of or thousands of messages where messages where messages where no one clears the context. no one clears the context.

no one clears the context. They just They just They just keep going and keep going and keep going and going. I going. I going. I mean, sometimes it's mean, sometimes it's mean, sometimes it's useful, but, you know, useful, but, you know, useful, but, you know, it's pretty weird. it's pretty weird. There are sessions for There are sessions for tens of millions of tens of millions of tens of millions of tokens where the cache is tokens where the cache is tokens where the cache is constantly filled constantly filled constantly filled to the maximum. This is getting to the maximum. This is getting to the maximum. This is getting really expensive. And really expensive. And really expensive. And one more point, right one more point, right ? So, you know, it was an ? So, you know, it was an ? So, you know, it was an exciting time, exciting time, exciting time, right? People are making right? People are making right? People are making new things, building new things, building new things, building new things, but this has in new things, but this has in new things, but this has in many cases many cases many cases led to very led to very led to very vertically vertically vertically integrated integrated tool chains. Cloud Code is obviously one of the is obviously one of the is obviously one of the big examples. big examples. big examples. CodeX, right? Everyone is CodeX, right? Everyone is CodeX, right? Everyone is building something, starting building something, starting building something, starting from the model itself and working their way up from the model itself and working their way up from the model itself and working their way up , , , creating agents and creating agents and creating agents and their their their toolchains. So, there toolchains. So, there toolchains. So, there are many problems here. are many problems here. are many problems here. And so, you know, my And so, you know, my And so, you know, my question, right? Regarding question, right? Regarding question, right? Regarding this: how do we this: how do we this: how do we solve these problems? solve these problems? Specifically, how do we Specifically, how do we overcome overcome overcome vendor lock-in and how do vendor lock-in and how do vendor lock-in and how do we actually we actually we actually improve improve improve the use of AI so that the use of AI so that the use of AI so that we can we can we can apply it apply it apply it to more than just some to more than just some to more than just some cutting-edge tasks, and cutting-edge tasks, and cutting-edge tasks, and not just waste not just waste not just waste money and tokens. But money and tokens. But money and tokens. But first, I want to first, I want to first, I want to talk about how talk about how talk about how not to solve not to solve not to solve this problem. And this is an this problem. And this is an this problem. And this is an extremely common extremely common extremely common thing. Over the past 6 months, thing. Over the past 6 months, thing. Over the past 6 months, I have spoken to I have spoken to I have spoken to hundreds of people about hundreds of people about hundreds of people about their spending on AI and their spending on AI and their spending on AI and how they are how they are how they are using it. This is using it. This is using it. This is one of the benefits of working one of the benefits of working one of the benefits of working on an AI gateway. People on an AI gateway. People on an AI gateway. People are turning to you.

are turning to you. are turning to you. They say, "I don't They say, "I don't They say, "I don't understand what's understand what's understand what's going on. I know going on. I know going on. I know we we we use AI a lot. I use AI a lot. I use AI a lot. I get a get a get a bill from Anthropic every day. I bill from Anthropic every day. I bill from Anthropic every day. I get a get a get a bill from Codex every day." Well, bill from Codex every day." Well, bill from Codex every day." Well, you know, it's not just like that you know, it's not just like that you know, it's not just like that , I , I , I see it all the time. So, what should I see it all the time. So, what should I see it all the time. So, what should I do? They usually do? They usually do? They usually come with a come with a come with a ready-made solution. And ready-made solution. And ready-made solution. And it sounds like, "Oh, it sounds like, "Oh, it sounds like, "Oh, everyone seems to be everyone seems to be everyone seems to be using using using cloud technology now." cloud technology now." cloud technology now." So we're just So we're just So we're just going to going to going to focus on focus on focus on one supplier one supplier . Maybe for two, right . Maybe for two, right ? Maybe for two. And ? Maybe for two. And ? Maybe for two. And they say, "I they say, "I they say, "I need to get this under need to get this under need to get this under control." How can I control." How can I control." How can I control this? How do I control this? How do I , well, oh, I understand. Yes, , well, oh, I understand. Yes, , well, oh, I understand. Yes, I'll say that everyone will I'll say that everyone will I'll say that everyone will use it, use it, use it, right? We're just going to right? We're just going to right? We're just going to use this use this use this thing. This is not the best thing. This is not the best thing. This is not the best idea. It may idea. It may idea. It may seem seem seem obvious from the outside that this is not the obvious from the outside that this is not the obvious from the outside that this is not the best idea, but best idea, but best idea, but when you are in a panic because when you are in a panic because when you are in a panic because you spent a lot of you spent a lot of you spent a lot of money and don't know what to do money and don't know what to do money and don't know what to do . And the finance . And the finance . And the finance department asks you department asks you department asks you questions. The IT department questions. The IT department questions. The IT department asks, "What do you asks, "What do you asks, "What do you do?" We said to do?" We said to do?" We said to use AI, use AI, use AI, but perhaps not to but perhaps not to but perhaps not to such an extent. And it's a such an extent. And it's a such an extent. And it's a completely natural completely natural completely natural reaction to say, " reaction to say, " Okay, I'll Okay, I'll Okay, I'll consolidate everything, right?"

consolidate everything, right?" consolidate everything, right?" Okay, I'm going to Okay, I'm going to Okay, I'm going to control this. I'm control this. I'm control this. I'm going to have, you know going to have, you know , control over this. , control over this. , control over this. Control over this. Control over this. Control over this. I think that's what we're I think that's what we're I think that's what we're going to see going to see going to see over the next 12 over the next 12 over the next 12 months and what people are months and what people are actually trying to do. At least actually trying to do. At least on the front lines, this is what on the front lines, this is what on the front lines, this is what I call " I call " profit maximization," right? profit maximization," right? We will not We will not We will not maximize tokens maximize tokens . We will . We will . We will maximize maximize maximize return on return on return on investment. And it's investment. And it's investment. And it's really just a really just a really just a necessity, right? And necessity, right? And necessity, right? And by "maximizing by "maximizing by "maximizing profitability" I do profitability" I do profitability" I do n't just mean n't just mean n't just mean cutting costs. I cutting costs. I cutting costs. I mean actually mean actually mean actually proving the value of what proving the value of what proving the value of what you do. And you do. And you do. And not only that, but also not only that, but also not only that, but also opening up new opening up new opening up new opportunities. So opportunities. So opportunities. So if you only have two if you only have two if you only have two leading models and leading models and leading models and their entire ecosystem, their entire ecosystem, their entire ecosystem, that creates a certain that creates a certain that creates a certain level of cost that level of cost that level of cost that makes some things makes some things makes some things just just just unviable, unviable, unviable, right? You won't be able to right? You won't be able to right? You won't be able to scale scale scale AI in your AI in your AI in your company if that's company if that's company if that's the only way you the only way you the only way you use it. So use it. So use it. So when we think about when we think about when we think about maximizing ROI, we maximizing ROI, we maximizing ROI, we first need first need first need to understand to understand to understand the components of the components of the components of internal internal internal AI implementation. I'm not AI implementation. I'm not AI implementation. I'm not talking about "hey, we talking about "hey, we talking about "hey, we built a new built a new built a new AI feature in our AI feature in our AI feature in our app." I app." I app." I mean the internal mean the internal mean the internal processes of your processes of your processes of your organization. So there are organization. So there are organization. So there are four components, right?

four components, right? four components, right? There is a big language model There is a big language model There is a big language model or models, right? This is the or models, right? This is the or models, right? This is the first level. There is a first level. There is a data connection level. For many, data connection level. For many, it's an MCP, but in reality it's an MCP, but in reality it's an MCP, but in reality it's a CLI, API, or just it's a CLI, API, or just it's a CLI, API, or just any way to any way to any way to get data, get data, get data, such as reading such as reading such as reading files. For your files. For your files. For your agent system. There is agent system. There is agent system. There is an interface. I an interface. I an interface. I think this is think this is think this is sometimes forgotten, but it all sometimes forgotten, but it all sometimes forgotten, but it all started with the chat started with the chat interface. interface. interface. TUIs appeared, we TUIs appeared, we TUIs appeared, we got IDEs with chat. got IDEs with chat. We went from TUI We went from TUI to using the to using the to using the phone, and now phone, and now phone, and now the trend is Slack, that is, the trend is Slack, that is, the trend is Slack, that is, interacting with agents interacting with agents interacting with agents through messengers, right through messengers, right ? This is another key ? This is another key ? This is another key component that can be component that can be component that can be highlighted. And lastly, highlighted. And lastly, sandboxes, that is, the sandboxes, that is, the sandboxes, that is, the environment in which the environment in which the environment in which the agent operates, right? agent operates, right? agent operates, right? So those are your four So those are your four So those are your four main components. main components. Maximizing ROI Maximizing ROI means taking advantage of means taking advantage of means taking advantage of every opportunity every opportunity every opportunity to make a choice at to make a choice at any of these levels, any of these levels, any of these levels, right? It is not necessary right? It is not necessary right? It is not necessary to have a unique to have a unique to have a unique solution at every solution at every solution at every level, but it is important to level, but it is important to level, but it is important to be able to be able to be able to make choices at make choices at any level when any level when any level when needed. And it can be needed. And it can be needed. And it can be different for different for different for different teams or different teams or different teams or people, right? This can people, right? This can people, right? This can be quite different across the be quite different across the organization. I can't organization. I can't imagine an entire imagine an entire imagine an entire development team development team development team writing code through a writing code through a Slack thread. Maybe it's fun, Slack thread. Maybe it's fun, but I can't imagine it,

but I can't imagine it, but I can't imagine it, right? Um, well, you know, the right? Um, well, you know, the right? Um, well, you know, the entire marketing entire marketing entire marketing team can team can team can work exclusively in a work exclusively in a work exclusively in a Slack branch, asking Slack branch, asking Slack branch, asking questions about data or questions about data or questions about data or doing other doing other doing other similar tasks. So it's similar tasks. So it's similar tasks. So it's really important to be able to, really important to be able to, really important to be able to, you know, sort of you know, sort of you know, sort of differentiate between these differentiate between these differentiate between these things. So, I work at things. So, I work at things. So, I work at Tailscale, right? I'm working Tailscale, right? I'm working Tailscale, right? I'm working on an AI gateway. It’s perhaps on an AI gateway. It’s perhaps on an AI gateway. It’s perhaps no surprise that I’m here to no surprise that I’m here to no surprise that I’m here to talk a little about talk a little about talk a little about why an AI gateway why an AI gateway why an AI gateway will help you will help you will help you do just that. And I do just that. And I do just that. And I don't necessarily mean don't necessarily mean don't necessarily mean using using using our gateway, although that would be our gateway, although that would be our gateway, although that would be nice, but nice, but any AI gateway is any AI gateway is any AI gateway is important for this. important for this. important for this. So, our AI gateway So, our AI gateway So, our AI gateway is called Aperture. And is called Aperture. And is called Aperture. And what it does—and what it does—and this may not come as this may not come as this may not come as a surprise—it a surprise—it a surprise—it sits between sits between sits between your agents and the your agents and the your agents and the interface of those interface of those interface of those agents, and agents, and agents, and all of your all of your all of your endpoints: LLM endpoints: LLM endpoints, MCP endpoints, MCP endpoints, API endpoints, API endpoints—whatever endpoints—whatever endpoints—whatever you have—and allows you have—and allows you have—and allows you to monitor and you to monitor and you to monitor and track what's track what's track what's going on. This could going on. This could going on. This could be from a be from a be from a security perspective, from a security perspective, from a security perspective, from a cost perspective, or from a perspective cost perspective, or from a perspective cost perspective, or from a perspective that we often that we often that we often do: wanting do: wanting do: wanting to know exactly what a to know exactly what a to know exactly what a user user user or agent is doing. It's or agent is doing. It's or agent is doing. It's actually really cool actually really cool actually really cool when you have logs and you when you have logs and you when you have logs and you can tell can tell can tell your agent, "Hey, your agent, "Hey, your agent, "Hey, maybe he maybe he maybe he ran out of resources."

ran out of resources." ran out of resources." He just, you know, didn't He just, you know, didn't He just, you know, didn't get an answer. get an answer. get an answer. He seemed to be out of He seemed to be out of He seemed to be out of control. Maybe control. Maybe control. Maybe it happens it happens it happens less often now, but it used to less often now, but it used to less often now, but it used to happen quite happen quite happen quite often, and it often, and it often, and it happens more often, you know, with happens more often, you know, with happens more often, you know, with cheaper models. cheaper models. cheaper models. They usually aren't very They usually aren't very They usually aren't very good at stopping good at stopping good at stopping themselves, but they are great at themselves, but they are great at themselves, but they are great at analyzing where they analyzing where they analyzing where they made mistakes after the made mistakes after the made mistakes after the fact. So fact. So fact. So when you have something like an when you have something like an when you have something like an AI gateway, it's very AI gateway, it's very AI gateway, it's very easy to go back to easy to go back to easy to go back to what you were doing and what you were doing and what you were doing and ask, "What happened ask, "What happened ?" Why didn't it ?" Why didn't it ?" Why didn't it work the way I work the way I work the way I expected? Well, so you expected? Well, so you expected? Well, so you know, if you've know, if you've know, if you've used an AI used an AI gateway, I... How many gateway, I... How many gateway, I... How many people have an AI gateway, have people have an AI gateway, have people have an AI gateway, have deployed one, or are deployed one, or are deployed one, or are using one right now using one right now ? Yes. So, if you've ? Yes. So, if you've used it before, you're used it before, you're probably used to it being like, probably used to it being like, probably used to it being like, "Okay, I'm "Okay, I'm "Okay, I'm putting all my putting all my putting all my keys on this gateway. keys on this gateway. keys on this gateway. Now I need to Now I need to Now I need to create synthetic create synthetic create synthetic keys for everything keys for everything keys for everything I have. Right? So it I have. Right? So it I have. Right? So it turns out that I turns out that I turns out that I traded some keys for a traded some keys for a traded some keys for a bunch of others. We bunch of others. We bunch of others. We can do something can do something can do something special. Aperture special. Aperture special. Aperture is built on Tailscale, is built on Tailscale, is built on Tailscale, and if you're not familiar with and if you're not familiar with and if you're not familiar with Tailscale, it's an identity-based mesh network Tailscale, it's an identity-based mesh network . It is . It is built on built on built on WireGuard. And what you can WireGuard. And what you can WireGuard. And what you can do with Tailscale—every do with Tailscale—every do with Tailscale—every connection that's connection that's connection that's made through the made through the made through the tailnet, which is the tailnet, which is the tailnet, which is the Tailscale network, well, you Tailscale network, well, you Tailscale network, well, you install a install a install a tailnet agent wherever you tailnet agent wherever you tailnet agent wherever you need it. It could need it. It could need it. It could be containers, be containers, be containers, laptops, GPU servers, it laptops, GPU servers, it laptops, GPU servers, it doesn't matter. You doesn't matter. You doesn't matter. You install it install it install it everywhere, and every everywhere, and every everywhere, and every connection that is now connection that is now connection that is now established through established through established through this tailnet receives the this tailnet receives the machine or machine or user identification data from the other user identification data from the other user identification data from the other side. This could be a tag side. This could be a tag , right? That is, it could , right? That is, it could , right? That is, it could be, for example, the be, for example, the be, for example, the machine ID. This is a machine ID. This is a machine ID. This is a PR check bot, right? And PR check bot, right? And PR check bot, right? And then it gets a then it gets a then it gets a separate node ID that separate node ID that separate node ID that points to an instance of

points to an instance of points to an instance of that bot to that bot to that bot to check the PR, right? You check the PR, right? You check the PR, right? You can do that, can do that, can do that, or it could be the or it could be the or it could be the user, right? It's user, right? It's user, right? It's just me just me just me using Cloud Code using Cloud Code using Cloud Code on my laptop. on my laptop. on my laptop. This can also be This can also be This can also be transmitted along with transmitted along with transmitted along with this. Also, this. Also, this. Also, if you if you if you use something use something use something like Entra or Okta, like Entra or Okta, like Entra or Okta, we can actually we can actually we can actually sync all the sync all the sync all the groups. So, you can groups. So, you can groups. So, you can see the group that see the group that see the group that this person is associated with. this person is associated with. this person is associated with. So someone on the other So someone on the other So someone on the other side can see all of side can see all of side can see all of this. We have a very cool this. We have a very cool this. We have a very cool library. It library. It library. It is called TSNet. This is an is called TSNet. This is an open source library. You open source library. You can can can use it use it use it to create such to create such to create such applications. That's what applications. That's what applications. That's what we used to we used to we used to create Aperture. That's what's create Aperture. That's what's create Aperture. That's what's really really really happening here. So happening here. So happening here. So the system can see the system can see the system can see the identification of that the identification of that the identification of that user or user or user or agent on the other side of agent on the other side of agent on the other side of the thing or the other side of the the thing or the other side of the the thing or the other side of the connection. And so, connection. And so, connection. And so, again, these are just again, these are just again, these are just small components of the small components of the small components of the internal internal internal AI deployment, right? AI deployment, right? So there's the LLM aspect, the So there's the LLM aspect, the data aspect, the data aspect, the data aspect, the interface aspect, and, essentially interface aspect, and, essentially , the , the sandbox environment in which it sandbox environment in which it sandbox environment in which it all works. So all works. So all works. So when it comes when it comes when it comes to using Aperture, the to using Aperture, the to using Aperture, the LLM part is pretty LLM part is pretty LLM part is pretty similar to what you similar to what you similar to what you would expect from would expect from would expect from other AI gateways, except other AI gateways, except other AI gateways, except here I here I here I already have the already have the already have the gateway set up. There's Anthropic, OpenAI, gateway set up. There's Anthropic, OpenAI, gateway set up. There's Anthropic, OpenAI, Vercel, even Anthropic OAuth, Vercel, even Anthropic OAuth, Vercel, even Anthropic OAuth, Bedrock, Codex—everything you Bedrock, Codex—everything you Bedrock, Codex—everything you can imagine.

can imagine. I add all these keys I add all these keys to the gateway, and now when to the gateway, and now when to the gateway, and now when I need to I need to I need to connect to the connect to the connect to the LLM provider as an LLM provider as an LLM provider as an agent or user, agent or user, agent or user, I don't need to I don't need to I don't need to create additional create additional create additional API keys or API keys or API keys or IDs on IDs on IDs on separate screens. As separate screens. As separate screens. As soon as I connect soon as I connect soon as I connect to the gateway, it knows to the gateway, it knows to the gateway, it knows who I am and what I am and says, "Oh who I am and what I am and says, "Oh , you have access to these , you have access to these , you have access to these models and this models and this models and this provider." You provider." You provider." You can can can use GLM 5.2, use GLM 5.2, use GLM 5.2, and you get some and you get some and you get some budget for GLM 5.2, right? You budget for GLM 5.2, right? You budget for GLM 5.2, right? You get an get an get an unlimited budget unlimited budget unlimited budget for the cheap model and a for the cheap model and a for the cheap model and a very limited very limited very limited budget for the advanced budget for the advanced budget for the advanced model, right? These are the model, right? These are the model, right? These are the things. There are no things. There are no things. There are no additional additional additional settings here. Just settings here. Just settings here. Just add your API key and add your API key and add your API key and go. So, go. So, go. So, again, we've taken the entire again, we've taken the entire again, we've taken the entire LLM level and LLM level and LLM level and put it in put it in put it in one place that you one place that you one place that you can scale can scale can scale across your organization. across your organization. across your organization. Again, it doesn't Again, it doesn't Again, it doesn't matter if matter if matter if you want to change models or you want to change models or you want to change models or providers—you providers—you providers—you can do it. can do it. can do it. Next is the Next is the Next is the data aspect. In addition to being a data aspect. In addition to being a data aspect. In addition to being a gateway to LLM—and that's why gateway to LLM—and that's why gateway to LLM—and that's why I I I called it an AI called it an AI gateway at the beginning. I know this gateway at the beginning. I know this gateway at the beginning. I know this sounds a bit vague, sounds a bit vague, sounds a bit vague, but I deliberately chose but I deliberately chose but I deliberately chose this term this term this term because we are both an LLM because we are both an LLM gateway and an MCP gateway, and gateway and an MCP gateway, and gateway and an MCP gateway, and even inject even inject authentication for authentication for standard APIs, standard APIs, standard APIs, re-exposing re-exposing re-exposing them as MCP. So I them as MCP. So I them as MCP. So I deliberately chose a deliberately chose a deliberately chose a broader term.

broader term. Thanks to this, as you Thanks to this, as you can see, in can see, in can see, in Aperture as an Aperture as an Aperture as an administrator you administrator you administrator you can configure can configure can configure any MCP servers you any MCP servers you any MCP servers you want. We have want. We have want. We have built-in servers built-in servers built-in servers that we have tested that we have tested that we have tested to make sure they are to make sure they are to make sure they are easy easy easy to work with. If you have to work with. If you have to work with. If you have ever ever ever deployed MCP or deployed MCP or deployed MCP or used MCP in used MCP in used MCP in your organization, you your organization, you your organization, you know that there is a know that there is a know that there is a standard, but everyone standard, but everyone standard, but everyone adheres to their adheres to their adheres to their own version of it. own version of it. own version of it. You know, you can You know, you can You know, you can determine exactly when an determine exactly when an determine exactly when an MCP server was created MCP server was created MCP server was created just by looking at just by looking at just by looking at how exactly they decided to how exactly they decided to how exactly they decided to implement implement implement authentication. It's a authentication. It's a authentication. It's a great great great archaeology lesson when you're archaeology lesson when you're archaeology lesson when you're setting up MCP setting up MCP servers, but we're servers, but we're servers, but we're trying to make trying to make trying to make the process very the process very the process very simple by walking you through it all simple by walking you through it all simple by walking you through it all . So, . So, . So, as an administrator, you as an administrator, you as an administrator, you can say, "I can say, "I can say, "I configured these configured these configured these servers, and these are the ones that are servers, and these are the ones that are servers, and these are the ones that are allowed." You allowed." You allowed." You can can can use the use the use the same same same identification information to identification information to identification information to specify that the specify that the specify that the marketing team marketing team marketing team accesses Google accesses Google accesses Google Calendar and Gmail, and the Calendar and Gmail, and the Calendar and Gmail, and the engineering team engineering team also accesses GitHub. Right also accesses GitHub. Right ? You can even ? You can even ? You can even specify, for example, specify, for example, specify, for example, that a certain part of the that a certain part of the that a certain part of the engineering team engineering team engineering team gets gets gets write or write or write or read-only access to GitHub. You can read-only access to GitHub. You can read-only access to GitHub. You can do many do many do many different things here. And what's different things here. And what's different things here. And what's really nice, really nice, really nice, again talking again talking again talking about decoupling about decoupling about decoupling systems, is that all the systems, is that all the systems, is that all the authentication authentication authentication happens here. So happens here. So happens here. So if you have people who if you have people who if you have people who want to want to want to use Cursor, use Cursor, use Cursor, Cloud Code, Code X, or some Cloud Code, Code X, or some Cloud Code, Code X, or some special agent special agent special agent framework—something new framework—something new framework—something new and popular that they and popular that they and popular that they discovered or created.

discovered or created. They just need to They just need to access that access that access that endpoint, and if endpoint, and if endpoint, and if they're on the Tailnet, they're on the Tailnet, they're on the Tailnet, everything authenticates everything authenticates everything authenticates as them, or as an agent, as them, or as an agent, as them, or as an agent, or as a tagged or as a tagged or as a tagged device. There is no device. There is no device. There is no need to constantly need to constantly need to constantly re- re- re- authorize yourself in authorize yourself in authorize yourself in each service or each service or each service or new tool. new tool. new tool. We make the cost of We make the cost of We make the cost of transitioning to new transitioning to new transitioning to new tools, tools, tools, workflows, and workflows, and workflows, and interfaces much interfaces much interfaces much lower thanks to lower thanks to lower thanks to this approach. It is this approach. It is this approach. It is because of these minor because of these minor because of these minor inconveniences that people inconveniences that people inconveniences that people often hesitate often hesitate often hesitate to try new to try new usage scenarios, unwilling to experiment experiment to make something to make something to make something cheaper or faster. cheaper or faster. Now let's talk about Now let's talk about the interface. So, in the interface. So, in the interface. So, in Aperture we have a Aperture we have a Aperture we have a built-in chat built-in chat interface, surprise! interface, surprise! interface, surprise! It seems like It seems like It seems like everyone has a chat everyone has a chat interface these days, and we interface these days, and we interface these days, and we have one built in too. have one built in too. have one built in too. However, it is However, it is However, it is built on built on built on APIs that you APIs that you APIs that you can also can also can also use on your use on your use on your own. So our own. So our own. So our main goal with Aperture is main goal with Aperture is to give you the same to give you the same to give you the same tools and tools and tools and interfaces, or interfaces, or interfaces, or tools to tools to tools to create create create interfaces, that interfaces, that interfaces, that we have. So, do we have. So, do we have. So, do you want to build a you want to build a you want to build a Slack bot, or a Slack bot, or a Slack bot, or a chat interface, or, chat interface, or, chat interface, or, let's say, something new let's say, something new let's say, something new that no one has that no one has that no one has thought of yet, like a thought of yet, like a thought of yet, like a mobile app or a mobile app or a mobile app or a voice interface, voice interface, voice interface, right? You can right? You can right? You can do it. The main idea do it. The main idea do it. The main idea is that you can is that you can is that you can do this do this do this using Aperture. We using Aperture. We using Aperture. We want to support want to support want to support all these different individual all these different individual all these different individual use cases use cases . But again, you . But again, you . But again, you see there are see there are see there are connectors here. They are all connectors here. They are all connectors here. They are all configured configured configured directly through directly through directly through Aperture itself. They are not Aperture itself. They are not Aperture itself. They are not configurable in the configurable in the configurable in the chat interface. You chat interface. You chat interface. You get them simply get them simply get them simply through Aperture. And lastly,

through Aperture. And lastly, through Aperture. And lastly, we talked a little we talked a little we talked a little about interfaces. It is about interfaces. It is about interfaces. It is very important that you very important that you very important that you can can can use all of use all of use all of this. By the way, another this. By the way, another this. By the way, another thing that's thing that's thing that's coming up soon: if you've coming up soon: if you've coming up soon: if you've ever tried ever tried ever tried to set up Open Web UI, to set up Open Web UI, to set up Open Web UI, or Lever, or something or Lever, or something or Lever, or something like that—you know, that like that—you know, that like that—you know, that type of interface, right? The type of interface, right? The big problem with big problem with such an interface is such an interface is such an interface is that you always that you always that you always lose lose lose identification information, and identification information, and identification information, and suddenly everything starts suddenly everything starts suddenly everything starts going through the UI only, right? going through the UI only, right? going through the UI only, right? Everything looks like it's Everything looks like it's Everything looks like it's coming coming coming straight from the system. straight from the system. straight from the system. We are going to add We are going to add We are going to add full end-to-end full end-to-end full end-to-end authentication. So authentication. So authentication. So you will be able to see all you will be able to see all user and machine IDs user and machine IDs if they if they if they use any of use any of your interfaces. your interfaces. your interfaces. You can You can You can configure everything configure everything configure everything to skip to skip to skip authentication authentication authentication while maintaining while maintaining while maintaining the same visibility you the same visibility you the same visibility you expect from an AI expect from an AI gateway. And here's the gateway. And here's the gateway. And here's the last moment, right? last moment, right? last moment, right? I think this is exactly the I think this is exactly the I think this is exactly the area where people area where people area where people work quite work quite work quite well. In fact, well. In fact, well. In fact, the choice is already quite the choice is already quite the choice is already quite large. People large. People large. People are trying to are trying to are trying to do this. I think it's because the do this. I think it's because the definition of a " definition of a " sandbox"... I'm sandbox"... I'm sure if I sure if I sure if I asked everyone asked everyone asked everyone present, everyone present, everyone present, everyone would have their own, individual would have their own, individual would have their own, individual definition of what a definition of what a definition of what a sandbox is, or sandbox is, or sandbox is, or what they need from it what they need from it what they need from it , or what , or what , or what they expect from it they expect from it . So, you know, we . So, you know, we . So, you know, we are now at a are now at a are now at a stage where "let stage where "let stage where "let a thousand a thousand a thousand flowers bloom," where there will be a sandbox supplier for flowers bloom," where there will be a sandbox supplier for flowers bloom," where there will be a sandbox supplier for every whim, every whim, every whim, need, or desire need, or desire . But, . But, you know, people are doing a pretty you know, people are doing a pretty you know, people are doing a pretty good job of good job of good job of integrating them, integrating them, integrating them, but again, but again, but again, it's very important. We're actually it's very important. We're actually it's very important. We're actually going to going to going to add a kind of

add a kind of add a kind of orchestration layer orchestration layer orchestration layer for sandboxes into the gateway. for sandboxes into the gateway. for sandboxes into the gateway. Again, Again, Again, I try to use I try to use I try to use the term "AI gateway" very the term "AI gateway" very the term "AI gateway" very loosely, deliberately very loosely, deliberately very loosely, deliberately very loosely, because in order to loosely, because in order to loosely, because in order to make make make the best choices for the best choices for the best choices for all the different all the different all the different endpoints that your endpoints that your endpoints that your environments have, LLMs, environments have, LLMs, environments have, LLMs, data connectors, and all data connectors, and all data connectors, and all that, you really that, you really that, you really need to have a place need to have a place need to have a place where you can see them where you can see them where you can see them all, and also a place all, and also a place all, and also a place that brings them that brings them that brings them together so that you can together so that you can together so that you can use use any of them. It's like a any of them. It's like a any of them. It's like a universal level of universal level of universal level of translation in a translation in a translation in a sense. But yeah, that's pretty much it sense. But yeah, that's pretty much it sense. But yeah, that's pretty much it . And as I . And as I . And as I said, I remind you that said, I remind you that said, I remind you that maximizing maximizing maximizing return on return on return on investment is investment is investment is really really really maximizing choice maximizing choice maximizing choice to prevent to prevent vendor lock-in, vendor lock-in, right? So, even if right? So, even if right? So, even if you don't use you don't use you don't use Aperture, I highly Aperture, I highly Aperture, I highly recommend that you recommend that you recommend that you start looking at AI start looking at AI Gateway, take a look at Gateway, take a look at Gateway, take a look at it. This is not it. This is not it. This is not only useful in terms of “oh, I only useful in terms of “oh, I only useful in terms of “oh, I don’t have to don’t have to don’t have to set up three set up three set up three different different different providers or providers or providers or switch between switch between switch between them”, it’s useful them”, it’s useful them”, it’s useful beyond that, right? This is beyond that, right? This is beyond that, right? This is great just for great just for great just for understanding what understanding what understanding what agents do, agents do, agents do, even if it even if it even if it only applies to you only applies to you only applies to you as an individual. We have as an individual. We have as an individual. We have so many people who, so many people who, so many people who, just as individual just as individual just as individual users, users, users, hide their hide their hide their usage behind an AI usage behind an AI gateway to have gateway to have gateway to have an idea of ​​what an idea of ​​what an idea of ​​what their agents are doing, their agents are doing, their agents are doing, because if you run because if you run because if you run something like OpenClaw something like OpenClaw something like OpenClaw through Tailscale on your through Tailscale on your through Tailscale on your tailnet and it's tailnet and it's tailnet and it's just running in the just running in the just running in the background, they're background, they're background, they're like, "Oh my God, it like, "Oh my God, it like, "Oh my God, it did something." I'm not did something." I'm not did something." I'm not sure what exactly it sure what exactly it sure what exactly it did. It did. It did. It did something it shouldn't have did something it shouldn't have did something it shouldn't have done.

done. done. With something With something With something like an AI gateway, you like an AI gateway, you like an AI gateway, you can go can go can go back, look, back, look, back, look, check, and check, and check, and diagnose exactly where diagnose exactly where diagnose exactly where the error occurred the error occurred the error occurred and what happened. So, and what happened. So, and what happened. So, you know, it's not you know, it's not you know, it's not just about being tied to a just about being tied to a just about being tied to a supplier, supplier, supplier, right? Actually, it's right? Actually, it's right? Actually, it's just experimentation just experimentation just experimentation and understanding in general and understanding in general , you know, it's really important , you know, it's really important , you know, it's really important to use something to use something to use something like an AI LLM gateway. like an AI LLM gateway. like an AI LLM gateway. So, that's all I So, that's all I So, that's all I wanted to say, wanted to say, wanted to say, briefly and to the point. briefly and to the point. briefly and to the point. Thank you very much for Thank you very much for Thank you very much for coming. I know this is the coming. I know this is the coming. I know this is the last last last report for today, report for today, report for today, so I hope so I hope so I hope everyone enjoyed the conference everyone enjoyed the conference everyone enjoyed the conference very much. very much. very much. I personally really I personally really I personally really enjoy the AI ​​Engineer enjoy the AI ​​Engineer enjoy the AI ​​Engineer World's Fair, and I hope World's Fair, and I hope World's Fair, and I hope you do too. So, thank you.