Free YouTube Transcribe

Video transcript

Gateways are All You Need — Karan Sampath, Anthropic

AI Engineer · 3,677 words · 17 min read

Want to search this transcript, jump the video from any line, or download it as TXT, SRT, or VTT?

Open in the transcript tool

Full transcript

0:07[music]

0:14>> All right. Um so everyone, I'm uh Karan

0:18Sampath. I'll be talking to you about

0:22how we Anthropic think about MCPs in the

0:24enterprise. I've alternatively titled

0:27it, I think more catchily, is why we

0:28think gateways are all you need. Um so

0:31before I go into the talk, uh I'm going

0:33to quickly tell you a bit about me. I

0:34was uh I'm a forward deployed engineer

0:36at Anthropic, um

0:37first one outside uh the US. Uh a lot of

0:40my work includes working with

0:42enterprises on things like MCPs, and I

0:44also work on our internal use cases. In

0:47this talk, I'm going to be

0:49positing to you what we think the

0:51problems with enterprises, what

0:53enterprises face with MCPs today, why we

0:56think gateways and the the necessary

0:58implications that come out of it are the

1:00best way to fix a lot of these problems,

1:02and what is that how does that align

1:03with our future vision for agent tech

1:05deployments. Um so before we go on, I

1:08know a lot of you already know this. I'm

1:09going to quickly run through this, but

1:11just a very quick overview of MCPs as it

1:13is relevant today. The first is of

1:14course we all know it's an open standard

1:16that was created. Uh most of us know

1:18that at Anthropic. Um there is an

1:19official registry today which contains

1:21over thousands of servers, and this is

1:23growing rapidly over the last year. Um

1:26and we see this happening that

1:27individual companies all the time are

1:29building new servers uh very quickly,

1:31and are trying to ensure that they stay

1:33ahead and trying to adhere to this

1:35protocol nowadays. But still, even if

1:37this happens, we see a major problem

1:40when enterprises try to use it.

1:42Enterprises struggle to deal with what

1:44they believe to be table stakes. Things

1:46like observability. When they want to

1:48know who's using my MCP, who's using

1:50these tools, how do I ensure that the

1:52correct people are using it, um and how

1:54do I know how to develop on certain

1:56parts of the MCP protocol, which parts

1:58of my tools aren't working properly.

2:00These are something that something

2:01that's simply completely opaque to

2:02enterprises today.

2:03Second is access control, something I

2:05just touched upon already, right? How do

2:07I ensure the correct correct users have

2:09access to these servers? Things like

2:10ensuring that certain servers are scoped

2:13correctly. Tools are only made are only

2:15allowed for certain groups of servers.

2:17If you're working in observability MCP,

2:19you might want the entire company to

2:20have

2:21view into like why which things are

2:23failing, but only certain people to have

2:25the ability to actually change things

2:26and update new dashboards, right? These

2:28are the kind of things that currently

2:29are quite hard to do with MCP servers.

2:31And honestly, not something that we as a

2:33community have worked on enough. And

2:35finally, security. And I like to think

2:37of the three of these is almost like a

2:38three-headed hydra in some ways, right?

2:40Um security is a wide range of things. I

2:43like to think not only in terms of the

2:44MCP server, where you know you

2:46enterprises want to know how do you

2:48verify whether a server is safe, it has

2:50uh the correct protocols and the correct

2:52ways of ensuring data exfiltration

2:54doesn't occur, things like uh you know

2:56the tools on on can't be used in a

2:59harmful way both for infrastructure

3:01externally, and but your internal

3:02infrastructure as well. But secondly,

3:04how do you ensure remote clients that

3:06are perhaps untrusted in nature can

3:08access your private data as an

3:10enterprise, right? These are all really

3:12hard problems that you've kind of solved

3:14in previous paradigms with APIs, but we

3:16really don't have a good way of solving

3:18it uh at the moment.

3:20So just going back to the registry

3:22point, it's really useful to think of

3:24where we are and where we want to go,

3:25right? Registries are really useful, and

3:27I don't think there's any discussion

3:29that here where we are having where we

3:30think they're not useful. We're really

3:31proud of it and we're really happy we we

3:33helped develop this. But it's really

3:35important to realize that registry

3:36registry simply aren't complete for an

3:38enterprise. And what's like funny about

3:40this is is that MCPs were specifically

3:42designed um if you attended David

3:45David's talk earlier today, MCPs are

3:47specifically designed because they allow

3:49themselves to be so much more useful for

3:51enterprises. And there's this kind of

3:52gap which the protocol allows for that

3:54we have yet to build into well, and

3:57these include things like

3:58authentication, but also access control,

3:59observability, and credential

4:00management. These are all things that

4:02enterprises need and are critical part,

4:04um but are simply not working. So now

4:07that we have this happening, what is

4:08what do enterprises do right nowadays?

4:11What they do is something like this,

4:12where every single team now with cloud

4:15core, you know, and explosion of coding

4:17across across various surfaces, can now

4:20start developing MCPs. But suddenly find

4:23out that they can't actually get them

4:24deployed. Or even if they want to get

4:25them deployed, the MCPs often can't use

4:28the tools they want to, those tools

4:30can't give them the correct access, and

4:31security teams on the other end are also

4:34pretty justified how they're going about

4:35it, because they're often overloaded,

4:37and they're often unable to see which

4:38MCPs they want to go through, and which

4:41they want to which they want to allow.

4:43And finally, at the company level, CEOs

4:46and C-suites are like, you know, why are

4:47my MCPs not working correctly? Why are

4:49my agents ineffective? Why can't um why

4:52can't your agents actually, you know,

4:54being be the thing that we all thought

4:55it was going to be? And so this

4:57bottleneck, which we're seeing right

4:58here, is something we need to solve. We

5:00need to ensure that security teams

5:03aren't overloaded, that users are given

5:05the freedom to actually develop their

5:06own MCPs, and organizations have

5:08visibility into all of them. And so what

5:11I'm going to tell you is that this

5:14problem where enterprises stay with a

5:15handful of MCP tools is going to

5:17fundamentally restrict the protocol and

5:18hurt agents until we really go and solve

5:21this.

5:22And I think it's worth zooming out at

5:23that moment, right? It's like I think

5:25it's really valuable to think how

5:27important these paper cuts are, and how

5:29valuable it is to try and invest the

5:30time to try and solve it um

5:32super super well.

5:34On to why we think the gateways are a

5:36better solution. I think this is a very

5:38way to good way to build intuition.

5:40The core intuition here is at a point in

5:43which almost all of your teams can build

5:45MCP servers really well, or have the

5:48ability to theoretically be able to do

5:50so because they're simply using coding

5:52agents that can understand the structure

5:53of the servers very well, that are able

5:55to understand what the tool definitions

5:56look like, what you should want, uh what

5:59access controls look like. The really

6:01important thing for security teams and

6:03enterprises that want to allow this to

6:04be decentralized is they need to

6:06establish a root of trust. And so we

6:08think that the goal for a secure this

6:10for any security team is to is to bless

6:13one platform. And this is if I take one

6:15thing away from this talk, I would

6:17really suggest it be this slide, because

6:19it's irrelevant like you can you might

6:20want to use a gateway and I'll mostly be

6:21talking to you through this, but I think

6:23this intuition is really the intuition I

6:25would really want to stress. Because

6:27enterprises that are able to do this, in

6:28our experience, we've done this

6:30internally and done this externally,

6:31too, are able to explore the usage of

6:33MCPs, and thereby explore the usage of

6:36how powerful the agents are. And it's

6:38worth obviously coming back here and

6:39saying like, you know, MCPs if the usage

6:41is exponential, every good MCP you have

6:43helps all of the agents in your company.

6:45And so doing something like this has

6:47knock-on effects beyond just that one

6:49MCP. Um so now now that we think that,

6:52you know, our platform is really useful,

6:54let me tell you why I think gateways and

6:56I'll define it for you are a good way of

6:58doing this. So a gateway, as a black box

7:01definition of it, is simply a middleman

7:05or

7:05sort of middle layer in between your MCP

7:08servers, and they can be numerous into

7:10the hundreds, and any MCP client. Notice

7:13the diagram here is notice is extremely

7:15simple, and it leaves a lot to be filled

7:17in. But what I really want to talk about

7:19is what we want to get out of the

7:20gateway. Things like authorization,

7:23authentication,

7:24observability, ensuring that you have

7:26correct connectivity between

7:28clear secure connections between your

7:30MCP client, which might be untrusted,

7:32and internally. And also finally, you're

7:35able a easy way to host and deploy any

7:37new MCP server. What this allows for is

7:40that any new MCP server now does not

7:42have to deal with any of these five

7:43things. And a team that wants to wants

7:45to add a new server only needs to care

7:47about what is the business logic. So

7:49your legal team which wants to review

7:51contracts only needs to care about,

7:53okay, if a contract comes in, this is

7:55what I want to be seen, this is how a

7:57redline should happen, this is what this

7:58is how you should escalate your

7:59different people. They don't need to

8:01ensure, you know, who who accesses this,

8:03how often can it be accessed, how do I

8:05know it's being accessed correctly, how

8:06do I know it's like scalable, and if I

8:08want to have new agents come in, how

8:10These are all things you need to worry

8:11about. And that's a really really useful

8:13middle layer to have, because actually,

8:15in the world we live in, your legal team

8:17can build the MCP server on their own.

8:18They aren't forced to go back to a new

8:20technical team and build it, right? And

8:22so if you really do want to live in that

8:23world, a gateway is a fundamental piece

8:25of that infrastructure.

8:28What does a gateway contain? Now there's

8:29like multiple definitions of gateways,

8:31we're going to give you examples, but in

8:32my mind, uh there is there are normally

8:35these components which usually exist

8:37within it. You know, I'm not going to be

8:38like this is hard and fast, this is

8:40neither exhaustive nor nor the only nor

8:43like the without these like there's not

8:46all required. But what you would really

8:47like to do to achieve the goals I just

8:49achieved I just talked about, is you

8:51kind of want a way to do auth, you want

8:53a way to do access control using roles,

8:55you want a way to route using a proxy

8:57such that any MCP client can only see

8:59your gateway, and and the gateway then

9:01can route to the individual MCP servers

9:03which treat the gateway as the only

9:05trusted endpoint. You want a way to

9:07ensure you have a tunnel which is a

9:08secured connection. You want to have a

9:10sub registry, which is your MCP servers

9:12internally. And finally, you would want

9:14to have any additional tooling. Tooling

9:16like a CLI for your gateway, such that

9:18anyone who wants to create a new MCP

9:20server in your in your company can

9:22easily create one because the gateway

9:24has a quick and easy CLI that not you,

9:27not that team, but but the agent that

9:29team is using, that cloud core or

9:31something else, can easily understand.

9:33Um these are all parts that when put

9:35together become really powerful, because

9:37someone who is just wants to create a

9:40new MCP server can use the gateway CLI

9:42and just easily integrate into these

9:44five components, and then completely

9:46focus on their MCP server, right? And

9:48that's why we really think that in

9:50making that kind of one-time investment,

9:52which doesn't hopefully doesn't require

9:53a lot of maintenance, and something you

9:55can easily do with agents would lead to

9:57several knock-on benefits.

10:00Um just like and just like a

10:02higher-level view of what this gives

10:03you, right? So, we already talked about

10:05this in terms of a listing of what we

10:07think is required, but just looking at

10:09once we have this, you have a vision of

10:12a gateway which can give you access

10:14authentication very easily. You might

10:15have your own IDP which you can plug in.

10:17You can have delegated identity in terms

10:19of users and agents. This is something

10:21we think is going to be really important

10:22into the coming year, where we think

10:24agents

10:25are going to require newer and novel

10:26definitions of identity that, you know,

10:29you can uniquely define and think about

10:31and scope for your enterprise using a

10:33gateway. We'll ensure that you can have

10:35one access control panel for all of your

10:38for all of your agents and and MCPs, and

10:40your access control can be scoped

10:42depending on like whether a team is

10:44accessing it, whether a user is

10:45accessing it, whether uh whether another

10:48employee is accessing it. And finally,

10:49observability. And I think observability

10:51here is nuanced in the sense that you

10:53not only want usage metrics to reflect

10:56what MCPs are

10:58being used, but you also kind of want to

11:00know how how are your tools being

11:01defined. In a world where the MCP

11:04protocol is itself being developed so

11:06rapidly, you kind of want to see how to

11:08how what are your load-bearing tools and

11:10how do you better adapt them to meet the

11:12needs of your various agents. So, I

11:14think this is really really cool and

11:16really useful. It kind of, you know,

11:18sometimes some people often say that it

11:19sounds too good to be true, but I really

11:21encourage you to try and take a look at

11:23this

11:23and try and and try and see where you

11:25can go from there. Um but let's say you

11:28have a gateway now, right? We've already

11:30talked about like the minimum

11:31requirements. I think there's a lot of

11:33exciting worlds that you get on very

11:36easy follow-ons or almost free lunches

11:38you kind of get from there. The first is

11:40it's very easy for you to add any new

11:41surface. So, you can easily have your

11:44MCP servers now plug into cloud.ai, they

11:46can easily plug into cloud core, they

11:47can easily plug into cloud core work

11:49because why? Because all of them are

11:50kind of listening to the same gateway

11:53and you kind of you only need to do it

11:54one time. Compare that to the world

11:56where you have 40 different MCP servers

11:58and some MCP servers are better

12:00configured for only one of these

12:01services in one of these clients. This

12:02is really important because this kind of

12:04ensures that you can be kind of

12:06invariant to any new surface that comes

12:08up. And I know like this is something

12:10that sounds interesting coming from me,

12:12but like it's really really useful for

12:14any new enterprise. It's really really

12:16useful for any enterprise to do this.

12:18The second is you have far more secured

12:20connections that can be built in.

12:21Currently, you know, your MCP servers

12:23often have access to sensitive data

12:25internally and you any enterprises are

12:27worried about, you know, "If I send this

12:29out to another client, is that data

12:30going to be exfiltrated? What happens to

12:32my data? How do I ensure that, you know,

12:34my like literally years and years of

12:36work doesn't go away?" You can now

12:38invest in much more secured connections

12:40between your MCP servers that are

12:42between your MCP server and your client

12:44that is completely encrypted

12:46and can ensure that you have root trust

12:48in how in what kind of data be sent. And

12:51this is really important. I think like,

12:52you know, we've often done you've often

12:54made MCP servers with kind of play data

12:57and things that, you know, enterprises

12:58can't really use, but for enterprises to

13:00actually derive the exponential value

13:02they can get from this, this is

13:03something we need.

13:04The third is faster iteration. The real

13:07value in like ensuring your teams can be

13:09decentralized completely is going to be

13:11that they can iterate and develop their

13:13own workflows much faster. If the legal

13:15team is able to very quickly just change

13:17their legal MCP rapidly and iterate on

13:19it without repeated security reviews,

13:22that has a massive effect which can just

13:26which just builds upon itself. So, this

13:27is something that shouldn't be

13:28underrated. The fourth is that you get

13:30very much more standard primitives. So,

13:32any new MCP server that you want to

13:34build kind of has to adhere to

13:36primitives that are that are within your

13:38enterprise and how your enterprise wants

13:40to go about it. A common request, you

13:42know, a common problem enterprises have

13:43is like, "How do I ensure said AI agent

13:47meets my standards and operating

13:48procedures?" A gateway is a way of

13:50encoding that. A gateway is a way of you

13:52being like, "These are my standard

13:54procedures, these are the primitives I

13:55want to see, these are the tools I

13:57expect, and these are the things I don't

13:58want to see." And that's a very quick

14:00way of encoding how you as an enterprise

14:02want to behave. Uh the fifth is

14:04pluggable credentials, right? What we

14:07currently have is certain MCP servers

14:09that only accept one type of user

14:11authentication, right? And it's very

14:13hard to kind of make sure it it it

14:15becomes

14:16like, you know, you can have a

14:18company-wide one or a team-wide one or

14:20something with service accounts. These

14:22are all things which are often useful in

14:24different cases, and that becomes easier

14:26with a gateway because a gateway

14:28allows any new MCP server to easily

14:30support new credentials and swap it in

14:32and out in an intelligent manner. And

14:34then finally, it's scalable, right? Like

14:37if you have 40 MCP servers that need to

14:39scale to from tens to hundreds to

14:41thousands to hundreds of thousands of

14:42agents, that's a really hard surface to

14:44maintain. A gateway which is able to

14:46take all these requests in and farm it

14:48out in an intelligent manner is a far

14:50better place for your teams to focus on.

14:53Finally,

14:54you know, this is not something that's

14:56we think is going to be like we're we're

14:57telling you out of the blocks. We have

14:59examples of this both in the open

15:01source, you have provided something

15:02in-house, but this is also something we

15:04can help with. This is something I do in

15:05my daily basis. If you're interested,

15:07I'd love to talk to you about this after

15:09this.

15:11Um but just finally zooming out, right?

15:13I want I want to take these last few

15:14minutes to think about where we're going

15:16with this. I like pitched a version of

15:18the MCP world to you that, you know,

15:19sounds cool, probably works, but does

15:22this really work in the hundreds of

15:23things you've been hearing over this

15:25conference? I think what we want to

15:26ensure is the larger picture where

15:29things are going is we want to separate

15:30the agent harness from where your data

15:32lives. We have many more surfaces coming

15:35up. We see an explosion of agents, but

15:37those agents shouldn't be tightly

15:39coupled to where how your data

15:41structured and how your MCPs are

15:43structured. An example of this is just

15:46the pure definition of an agent should

15:48like when you have tens of thousands of

15:51the orange box on the left, you don't

15:53need to keep you don't want to ensure

15:55that it has to be opinionated. You don't

15:56want you don't want the things on the

15:57right to be opinionated on how this

15:59works. We have this working we have an

16:02example of this which was released

16:03recently, right? Just as an example of

16:05this right now, if you have an MCP

16:07gateway, you can use you can easily

16:09connect it to cloud managed agents which

16:10was released recently, but you can also

16:12build it into use it internally

16:14with your own cloud agent SDK as well,

16:16right? And this is just an example from

16:17the cloud.x ecosystem. But think about

16:19the value of this. As an enterprise, you

16:22now have the ability to quickly decide

16:24which agents you want to keep in-house,

16:25which agents you want to have outside,

16:27but that becomes an invariant decision.

16:29The gateway remains regardless. What I'm

16:31trying to tell you is that the gateway

16:33is an investment which will allow you

16:35give you the flexibility to try to try

16:37and meet the wide-ranging agent needs of

16:40the future, and that's going to be

16:41really really exciting.

16:43And that allows you to not only

16:46not have to think that much about agent

16:48design at the moment, but it also allows

16:49you to really invest very strongly in

16:52opinionated MCP gateway primitives.

16:56So, just in summary,

16:58I would say three main things here,

17:00right? The first and the most important

17:01takeaway is to invest in common infra,

17:04to not try and roll your own MCPs, and

17:06to ensure your teams can build their own

17:07MCPs. Second, we really think gateways

17:10for secured connections allow you to

17:12build that root of trust. And third,

17:14that moves towards a world where we

17:16think the agent harness is better be

17:18able to separated from your data layer.

17:21So, yeah, thank you so much. If you have

17:23if you're interested in any of this, I'd

17:24love to talk

17:26and that was the talk. Thank you so

17:27much.

17:28>> [applause]

17:37[music]

This transcript was generated from the captions YouTube publishes for this video. Get the transcript of any YouTube video atfreeyoutubetranscribe.com: free, unlimited, no sign-up.