One Year of MCP — with David Soria Parria and AAIF leads from OpenAI, Goose, Linux Foundation

Hey, everyone. Welcome to the Latent Space Podcast. This is Alessio, founder of Kernel Labs, and I'm joined by Swix, editor of Latent Space.

Hey, and here we are joined finally in the studio for the first time. Uh, welcome back, David, from Anthropic/MCP.

Yeah. Hey. Well, like, nice to finally talk to you in person. Like, last time, like a year ago, it was over VC, and this is way fun.

I watched it back. It was eight months. Uh, it's, it's been a crazy eight months, and, uh, I think we just celebrated, like, the one-year anniversary of MCP.

Yes, we did.

At least the public announcement.

Yeah.

Um, and also last night or yesterday was the a-Agentic AI-

Foundation launch

...Foundation launch.

Wow.

Yeah.

That was nice. Was a nice event. It was nice to see the Anthropic office.

Yeah.

I've not been.

You like it?

Uh, yeah. It's very cool.

Very good food. I, I would say- ...um, in terms of my food bench, uh, Anthropic does rank over OpenAI.

Yeah.

At least that, that, that's what we have going for us.

Awesome, man. Do you wanna give just a quick overview of what's happening with MCP, and now you're donating it to the foundation, and then we'll do kinda like a one-year recap of the protocol itself, and then we'll have the rest of the leads from the foundation join us to do more of the high level?

Yeah. That, yeah, that sounds good. Um, yeah, I mean, the-- where, where, where we at at the moment, we have done, like, a year-- like a year ago, we launched it, and then we had this, like, crazy adoption over the last year now, which it feels like an eternity, honestly.

But we had this, like, crazy, um, growth and, like, adoption, um, you know, through initially through, like, Thanksgiving and Christmas, very early. It was a lot of, like, builders building MCP and then, you know, you had, like, the first big clients coming in, like Cursor and VS Code.

And then, like, you had this, like, inflection point around April with, like, Sam Altman and, uh, Satya and, um, Sundar and all, um, posting about, like, MCP and that they're gonna adopt MCP at Microsoft, at Google, um, at OpenAI, and that was really, like, the big inflection point.

So-

Yeah

...um, but in all of the time, you also had to do a lot of work on the protocol itself, right? We, like, we, we moved-- We launched originally as, like, basically local only. Um, you could, like, build local MCP servers for cloud desktop.

Um, but then we, like, in March this year, we moved into, like, how can you do remote, uh, MCP servers to connect, like, really about, like, to a remote server and introduced, like, the first, uh, iteration of authentication.

And then in June, we revisited that and, like, improved it quite a little bit so that it works better for, um, you know, uh, for enterprises in particularly. And we were very, very lucky that in that time from March to, like, June, we were, like, able to, like, have absolute industry-leading experts that literally work on OAuth itself-

Mm

...to help us with some of the, the pieces, right? Um, and how to get it right. And then we focused a lot of-- on, like, security best practices and this type of work. And now we, like, have-- I feel we have a really solid f-foundation, and we're doing-- We just launched, like, in, um, end of, like, November, the, the recent iteration of the protocol, finally, like, the next bigger improvement to the protocol, which is, like, long-running tasks to really allow for, like, you know, deep research type of task and, like, you know, maybe even agent-to-agent communication.

And so I think we're just stepping into, like, this territory now with, like, okay, we have really solid foundations. We have, like, one more big primitive we wanna have. We wanna make, like, a little bit more scalability, things work, and then we're, you know, gonna get into a phase where it's probably becomes a bit more stable.

And so, yeah, it's been, it's been an absolutely crazy year, man.

You did say the agent to agent, so there is a A2A protocol. Um, I'm curious when the Agentic Engineering Foundation got formed or just Agentic AI Foundation, was there any discussion about any of these other protocols being a part of it?

Or, you know, Sean Wa- wrote a post called Why MCP Won-

Yeah

...um, already. So, uh-

One of my favorite posts of the year.

Maybe it was already... It was a-- It was, and it was before Sam and all the other guys.

Yeah. Yeah. You were right all along.

Well, I, I mean, I, I think it was just obvious that was going to happen.

Yeah.

So.

Yeah. Um, so we did-- we, of course, have conversations a-around, like, what else is in the market, like, there are payment protocols that are interesting and so on. But when we wanted to, to start a foundation, we wanted to make sure, first of all, two things.

We wanted to start small and make sure, um, that the group that is founding this, like, for, for us, it's the first time we at Anthropic have an open source foundation, so this is all new to us. So we really wanna start it small and making sure we're learning along the way and being able to, like, like, shepherd this in, in the way we feel is best for, for the industry together with, with OpenAI, um, and Block.

But the second part of that is also, like, we really felt like we wanted to see things that, um, have a lot of adoption, are de facto, like, um, at least on the protocol side, like a de facto, uh, standard.

Um, and, and I don't think any of the other protocols-

Mm.

It feels like they're not just there yet. But of course, if they get there, then we're, like, super open, as long as they're, like, complementary to, to what's, what's in the foundation. On the application side, we're a little bit more flexible, and we're, like, more open.

But on the protocol side, I think we really wanna make sure that we're not, like, offering, like... That the foundation doesn't encompass, like, five protocols for the same, like, communication there. And so, yeah, there was discussion, but I think for now, we just wanna s-start it small.

Is there a role, like a double hat that you have now with the foundation, uh, or are you more focused on MCP?

I am still mostly focused on MCP. It's a bit of a double hat, so there is this, like... I, I think people need to understand, like, the foundation part is mostly just an umbrella to make sure the projects under it stay always neutral.

And I think that's really the most important part you wanna get a lot, um, you know, wanna understand. Because the rest of it is like, okay, how do we use the budget of the foundation for events and things that are, like- Quite dry.

And then the technical parts to, like, MCP, they stay actually the same. Like, on the go-- on the way we govern MCP, nothing has really changed. And so that's really still my job as the lead core maintainer of, like, shepherding, um, the processes, shepherding the, the protocol forward.

And then beyond that, now the additional double role is, like, I'm also gonna be on the, the technical steering committee of the foundation, which will, like, make sure to, like, figure out what are the projects we wanna have in the foundation.

So if someone comes with a project to us, the people that have projects in it will decide, is this something we would want? Is this something that we feel is, like, well-maintained, has a lot of adoption, is not gonna go away?

We wanna make sure the foundation is-- have, like, super interesting and important projects and not, like, a dumping ground like have, you know, how some foundations might have ended up with.

That's true. Uh, so we're gonna meet some of the others later, but maybe we'll just focus back on the sort of MCP development.

Yeah.

You've covered a lot. There's been four spec releases.

That's a lot.

Yeah. Yeah. So people may have missed some of them is what I'm saying.

Yeah.

Right? Like, and, and I think it's really interesting how, uh, you-- we've continued to work on, like, really important parts. Like, I, I always think, like, it's very hard to follow up a, a major success with a sequel.

Yeah.

'Cause the sequel usually, like, is-- it's hard to repeat-

Yeah

... that, uh, impact. But I think, like, every single time you've actually managed to, like, focus on something important.

Yeah.

So maybe we can cover, I, I guess, uh, maybe we'll start with the, the March/May one, uh, which is HTTP streaming, which is good, um, and the auth spec, right?

Yeah.

Any other-- I don't know if you wanna highlight any, any others, but we'll just catch people up on that stuff.

Yeah. So that, that was-- I think that was really a-- that was such an important one. Like-

It was the number one requested thing.

Yeah. It was like, it really opened up this, like, remote thing, and we were-- we already knew actually in December and November that the next big thing will be like, how can you do this over the-- over remote?

And authentication is quite important. One of the things I think people very rarely notice when it comes to MCP, MCP is very, um, very prescriptive in, like, each layer. Like, other protocols are not like that. For example, like, we like-- you wanna do authentication.

If the client and the server don't know each other, you need to do OAuth, right? And so we were very early, we wanted to have, like, one way to do something, and so we really focused on, like, what does this mean?

Like, how do we get it over? How do we build a protocol that is-- has this, like, these streaming properties that we require, and then how do we do authentication very early? Authentication, in the first iteration, I think we did an okay job, but we got some aspects wrong, and most of them honestly were just me not understanding enterprises well enough.

But then again, I think the, the strength that we have with MCP, and I think the, the one thing, if anything I'm proud of, is, like, building a community of people that can come together-

Perfect, yeah

... and help me figure shit out because, you know, I have my set of experiences of, like, what I'm good at, um, and enterprise authentication, it turns out, is not one of them, right? But there are way better s-suited people for that.

And so that's when we, like, after that March-

I, I saw, I saw you post that, but I didn't really dig into the details. Was it, like, the, the typical SAML, uh, s-type, type of auth-authentication issue?

The main issue we did is, um, in, in OAuth there are two components. There is, um, there's the authentication server who gives you the token, and then there's the resource server. It takes the token and gives you the resource in return.

And in the first iteration of our authentication spec, we combined them together into the MCP server, which if you were building-

Unusable, yeah.

The-- I-it's kinda usable if you build, like, an, an MCP server, like, as, like, a, like, an, a public server for-- as a, you know, you're a startup, you're building a server for yourself. You wanna bind this to, um, the accounts you already have.

That is completely usable. The reality in enterprises is you don't authentic-- you authenticate with some central entity. Like, you know, you know, you have some ID-

IDP

... identity provider or an IDP, and you-

You go Okta, Auth0.

And yeah. For most people, they-- what they don't even notice that it's happening. All they know is like, "Oh, in the morning I'm gonna go log in with Google for-- and then get access to all my work stuff," right?

But that's effectively the IDP, right? Um, and if you combine these into the same server, you just can't do this anymore. And so all we needed to do is like, okay, we are a resource server. The MCP server is a resource server.

How you get the token from the authentication server, we have opinions on how you should do it, but it's kinda separated. And that's what happened then in the June spec where we separated this out, um, and worked through a lot of these, like, okay, you know, how do you do dynamic client registration and other aspects which also were part of the March spec.

Um, we can talk about that. That's, that's a whole other story of, like, we are actually pushing the boundaries of what OAuth can do with, with MCP 'cause we're trying something very unique with MCP. Um, but yeah, that was, that was the big part in, in March, which we're like...

And that, um, was that authentication spec, the first iteration, then fixing it in June. Yeah.

What's the state of agents authenticating on my behalf? Because even today with the OAuth, I still have to, you know, log into Linear and whatnot.

Yeah. OAuth itself is, for the most part, a very human-centric protocol. It's just, it just tells you how you obtain a token if you don't have a token. Once you have a token, actually, it doesn't matter. You just put it into the bearer token.

And so we, we are not very prescriptive of what, like, agent-to-agent, um, authentication would look like or on behalf of agents. They are ideas that we are looking into, and I don't have all the specifics, but we are not prescriptive in the same way we're prescriptive as with OAuth.

But you can technically, the moment you have a token that might be, like, bound to, like, a workload identity-

Mm-hmm

... or something like that, then you just can pass that still to the MCP server. We're just not telling you how to obtain it just yet, and so we're not prescriptive. And so people do this, and they can do it when, where particularly when they're within, like, an enterprise and have a somewhat closed ecosystem.

But if the client and the server don't know each other, we just don't have a good solution for now.

Yeah. And then, yeah, on the remote thing, you went from local servers like SSC and then streamable HTTP. Any learnings you wanna call out there? Any, uh, yeah, regrets or, uh, learnings for others?

Certainly in transport. The one discussion that has never stopped from the very beginning of the last years are about transport. And we literally just spent the last two days at the Google offices with a bunch of, like, senior engineers from Google, Microsoft, AWS, Anthropic, OpenAI, just like- What do we need to do here to really, really make this solid?

Um, when we looked into MARD, we wanted to get a transport going that basically retains a lot of the properties we had from standard IO because we really-- and I still believe this until today, that the, that MCP should also enable agents, and agents are inherently somewhat stateful, and there's some form of like long-term communication going between, like, the client and the server.

And so we always looked for something like that. We also knew that we looked into alternatives like, okay, what happens if we do WebSockets, for example? And we have found a lot of issues with doing a proper bi-directional stream.

And we were like, "Okay, what is the right middle ground between having something that can be used in the simplest form that people do, like, where they just wanna provide a tool, but, um, then is able to be upgraded to like a full bi-directional stream if you need it because you really have, like, complex agents, you know, communicating with each other?"

That's where Streamable HTTP was born, with that intent. And I think there's some things that in retrospect that we got right i-- and, and something that we got wrong. I think we got right that we are really leaning just on, on standard HTTP in that regard.

We got wrong that we made a lot of things optional for the clients to do. Um, like you can-- the client can connect and open this return stream from the server, but it doesn't have to. And the reality is, is no client does it because it's optional.

Right.

And so a lot of the bi-directionality goes away. And so features like elicitations and sampling are just not available to servers because they don't have that stream open because the client, the client implementer was like, "Ah, that's the minimum viable project from-- product from me.

I don't have to do it." And so that, that became an issue. So I think there are lessons there. The second part of the lessons is that the way we designed the protocol, the, the transport protocol requires some form of holding state on the server side, and that is fine if you have one server, but the moment you scale this horizontally across multiple pods on a con-- like in containers or something like that, well, now if you get like two call and then an elicitation and then an elicitation result, you so- somehow require to, like...

You might hit two different servers, and you need to find a way to have these two servers somehow get this result together, and you in-- effect-effectively need some form of shared, like, uh, Redis, Memcache, whatever you want. Like something usually pops up or something like that.

Want to, to have like a shared state that you can, like, have. And that's kinda okay, and like we have seen this in PHP application and Python application being done, but it, it's, it's not fun if you do this at scale.

Mm-hmm.

And we know from like some companies like the Googles of the world, the Microsofts of the world, they're doing MCP at a scale that I can't tell you the numbers, but it's like in the millions of requests. And so now it becomes a problem, right?

And so now we're sitting here like, okay, how do you build an iteration of the protocol that allows for basically these principles of like make it as simple as possible for simple MCP servers, but allow this full spectrum of like really bi-directional streaming if you need it, but also make it scalable.

And I think we're just about to find the right solutions, but it's, it's just, it's complicated, yeah. Because a lot of the technology today is, is really just... There's very little like that. People either do the, the simple thing and then, then you do like something like REST or you do like a full-

Socket

... biDi stream and then you're just gonna do like WebSockets or like gRPC and so on, and we need kind of both.

What is it like to be in that kind of meeting where you have all these impressive companies and everyone is senior and everyone has an opinion and-

It's much fun. Yeah. I get to work with some of the best engineers in the, in the industry. Like this is-- It's insane.

Okay. Well, who decides? You know? Usually, usually there's-

It-- We're trying to get to consensus. Like the reality is technically I decide in the end of the day, but I think that's, uh, more like a formalism. In the end of the day, what you're trying to do is just to really narrow down of like what's the-- what are the real problems which we all agree on, what are the things where we nes- not necessarily agree on, and what are the, you know...

And then within those bounds, like build the best solution. And it takes a while, and it takes a lot of iterations. But it's, it's so much f- honestly, it's so much fun because you get to see these unique problems from, from the companies.

You, you see some of the identity of the companies in the problems themselves, right? Like, you know, Google has a different set of problems like an, like a Microsoft and, and a lot of it comes from like just the, the ways of building things, and then the problem from Anthropic look different from the problem from OpenAI.

But what I love about all of this is that everybody is the, like, sometimes you step back and, like, you sit in a room with all these competitive companies, but you're actually building something together, and I, I love that.

I've been in open source for like twenty-five years.

Yeah, it's very rare.

I love this kind of stuff. Yeah.

And when a standard works, this is the ideal.

Yes. And these people are all amazing. I just learn from, from all my peers so much. I'm, I'm very grateful to be in this situation. Yeah.

This reminds me of the IETF standards process.

Yeah.

Is there some discussion about how this works as a private group versus something more traditional?

It's an interesting one. Like it does look a little bit like the IETF. The IETF is very-- is slightly different. The IETF is like an open forum where everybody can go, and the result of that, it's like the IETF is very consensus-based and by accident, not by like, not necessarily because they want to be, but by accident, quite slow in the processes, which is very good in many ways.

It cannot be undone, right?

Right.

Once it's, once it's up, it's-

Yeah.

Yeah.

And it, like, for example, when you look at like the OAuth 2.1 spec, it's been in the works for like three years or four years, and they're just not done with it, right? And that's like, that's the length of which IETF standardization works.

Like these things can take a long, long time. And I think that's good for certain pieces, but I think AI at the moment is just so mo-- fast-moving. You just, you, you're somewhat forced to find a smaller group.

And so that's why we run MCP as like a really traditional open source, um, project with like a core maintainer group of like eight people. That basically decide everything and then like input from everybody else. Like we get input and people can make suggestions and we have-- a lot of the changes don't come from the core maintainers, but they are the ones that decide it, and that's like way more-- It's like a middle ground of being somewhat consensus-based, but also somewhat like a bit of a dictatorship, which can be good if you wanna move fast, which MCP wants to do at the moment.

How do you balance the influence of like the model improvements with how to shape the protocol? Because obviously, you know, you have Anthropic and OpenAI, you guys are doing post-training on these models to make them better, tool calling, and you have preferences on the shape of the protocol versus there's people that are not aware of like how you're structuring that.

So yeah, do you like share some of these? Like does the protocol influence some of the model post-training or like vice versa maybe?

I'm not one hundred percent familiar. Like I'm, I'm a product person. I'm not f-fully familiar with everything we do on the research side for sure, but it influences the post-training in the sense that we're making use of things like the MCP Atlas, that we are like having in our model card of like making sure that like we, we're taking this large set of tools in the wild and make sure, um, that our models work with that.

Mm.

But I think the primitives of the protocol, they're actually very rarely influenced by model improvements. I think there's a sense that, that we do anticipate the exponential that the models are on in terms of like improvement and that we're relying to some degree of mechanics that you can put into, into the model training.

I'm, I'm gonna get more concrete here. So for example, people have had long conversations around context build of MCP servers. Um, and that happens because MCP opens up the door to a lot of tools, and if you naively take all the tools and throw them into the context window, you, you just get a lot of bloat.

It would be just equivalent if you take all the skills, take all the markdown files, and just throw them all into the context. You would also have a lot of bloat. Um, but we already knew and, and I think we always knew that we-- that you, that you can do something like progressive discovery, and that's like a-- There's this general principle thing of like you can add, give the model some information, let the model then decide to and gain more information, right?

And of course, here is where we're like, you know, some of the foresight that we see because we are, we're, we're, we are the big model companies, we know that we can train this if we wanted to. And what the training does is just optimizes it.

The model can do it in principle already, right? And it can-- any model can do it. It does any type of tool calling. But if you train the model for it, it's just better at it, right? And so these things then go hand in hand i-in a way.

But in the end of the day, the, the ge-general, general mechanic of progressive discovery or dis-dis-- uh, yeah, progressive discovery, that's just inherent to, to any type of model that can do any type of tool calling in the end of the day, if that makes sense.

Mm-hmm. Yep. Yeah, and I think the, yeah, the context route point is important, and I think now there's the MCP versus code mode thing. And then it's like, well, if Anthropic says code mode and Anthropic made MCP, maybe is that the best way to go or like-

So the protocol was never actually called a code mode. That's the Cloudflare term.

That's it. Well, well yeah, but like people call it-- we call it programmatic-

Yeah

... MCP and other call it code mode. In the end of the day, what it boils down to is just like, okay-- And here's the interesting part, like the... So first of all, MCP is a m- is a protocol between the AI application and like servers, right?

So in-- the model is actually technically not involved in MCP. Um, and so now you have an application go like, "I have a bunch of tools. What can I do with it?" And you can do the naive thing and go like, "Okay, I have tools.

Uh, I'll throw them into tools for the model, and I, I, I call them." But you can be more creative with it. You can go and like, okay, models are really good at writing code.

Mm-hmm.

What if I take this and treat it like just like API calls, and you give it to the model, and now the model generates, you know, code. And what you're effectively doing is this composability that the, the model would have done anyway by like call tool A, you know, get the result, go back to inference to call tool B and then combine it into call three.

Now you-- all you've done is you let the model optimize it in advance and put them into a bunch of code that is just executed in a sandbox and go like call one, put it into two, put the results into three, get a result.

And all you've done is an optimization at the end of the day. Um, but the, the benefits of MCP of having authentication done for you, having, um, something that is, that is suited for the LLM, something that is automatically dis-- that is discoverable and self-documenting, this thing has not gone away.

That's still MCP for you, right? You're just using it a different way. So I'm always a little bit confused when people go like, "But MCP is-- Why, why does it tell me that, that, that does not-- Does it mean MCP is useless?"

No, it's still-- It's just a different use, right? And I think you will see evolutions as we're getting better of like how we use these models and the infrastructure around it gets a bit more mature, and you suddenly can assume that most model like AI applications will have some form of like sandboxing for execution.

You can do a lot more fun stuff like that, but I don't think that the value of like an, a protocol that connects the model to the outside world is, is gone because of it.

Yeah.

If that makes sense. I see it purely as an optimization, honestly, as a token optimization.

Is this a good time to bring up skills?

It's always.

It's awesome. So we, uh, skills is a more recent concept.

Yeah.

Uh, we-- I only bring it up because it's mentally linked in my mind to progressive disclosure and to adding-

Yep

... preset code scripts-

Yep

... and all that. Uh, skills can also create skills, which is just very fun.

Yeah.

Well, I think a lot of people are trying to place MCP versus skills. Obviously, they're not overlapping, but how do you view it?

Yeah, I agree. I, I think that's the interesting part is like they're not overlapping. I think they, they solve different things. I think skills are super great and, you know, they're-- I think that the first that really like they're being built from the principle with progressive dis- uh, discovery.

But I think the mechanism of progressive discovery, that's just universal to any type of thing you can do with a model. But what skills do, they like, they give you the domain knowledge for like a specific Set of tasks, like how you are, how you behave, how should the model behave as a data scientist, or how should the model dis- um, behave as, I know, an accountant or whatever.

But MCP gives you the connectiveness of the actual actions that you can take in with the outside world. And so I think they're somewhat, um, orthogonal in like, in terms of like the skills really gives you this domain knowledge, just like kind of vertical, and then like MCP gives you this horizontal of like, okay, you know, give me that one action.

And of course, skills can take actions. They can take actions because you can have code and scripts in there, and that's great, but it has two interesting aspects that I think people ignore. The first one is you need an execution environment, so you need to, you need to find a way to execute-

Usually your machine. Yeah.

Yes, and that, that's, that's perfectly fine for, you know, if you like run a local, like, you know, cloud code or something, then we can talk about like CLIs, for example. In those scenarios where you have like an execution environment, these things make a lot of sense and, and um, and then it's great.

Um, or if you have a remote execution environment, then it make a lot of sense, but you still don't get authentication in that regard. And so what I think MCP brings is the authencat- authentication piece. It brings the piece that you don't have to the, this, like an external, like person, like for example, if you have like a linear MCP server, they can improve the server.

You don't have to deal with that in your skill, right?

Mm-hmm.

It's not fixed in space. And then the third part is that you don't necessarily need an execution environment because the execution environment is effectively somewhere else on a server. Um, and so if you build a web application or a, like um, like a, a mobile application, you-- these things come w- you know, work better in, in some of these regards.

So I think they are orthogonal in that regard for the most part. And they are... And I've seen some quite cool deployments where people use skills to explore of like different functions, different like, you know, the accountant, the engineer, the, the data scientist, and then use MCP server to connect, um, these skills-

Mm-hmm

... to the actual like data sources within the company. And I think that's actually a really fun model, and I think that's the closest how I think about this.

Yeah. The, so MCP is the connectivity layer, I think-

Yeah

... is the word that you choose.

The communication layer.

Communication layer.

Yeah.

Yeah. So is, is it, um, architecturally, I'm, I'm wondering if it's like the MCP's client inside of each skill or is there a shared client that can discover skills?

We do that as shared. We do that as a shared one. I think you technically want a bit more shared ones because you do... The more shared you have, the better you, the more you can do like discovery things.

You can do things like, okay, I have connection pooling, I have this, I can, I'll do automatic discovery of things. I can even like, you know, in a skill you might just very loosely describe what you want, and I can look into the registry that I have access to and get an MCP server for you, right?

Yeah.

These things can, you do, can do when you do, but I think both works in the end of the day.

Yeah.

But these, this is things to experiment with.

Uh, I do wanna highlight for people who might have missed it. You say, "We do," blah, blah, blah. Actually, I think nobody understands, uh, enough how much Anthropic dogfood MCP.

Yeah.

And I only understood this when I watched John Welsh do his talk-

Yeah

... at AIE, where he was like, "Yeah, we have an MCP gateway. Everything goes through this."

Yeah.

And like, what can you say more about that?

Yeah. I mean, like, you know, we, we use both, right? We use a lot of, like, we use a lot of skills internally. We use a lot of MCP servers internally because like we have, you know, obviously, you know, you wanna make it very easy for people to deploy MCP.

You wanna like have some form of like integration with your IDPs and so on. So we have a gateway that we've built custom purpose for ourselves, and you just gotta like deploy your MCP servers, um-

It, it is all internal apps? Uh-

It's all internal stuff.

Yeah.

Yeah. Some of them are like external things, like, like technically external things, but in the lack of them offering a first party one, we have our own. Like we have a Slack MCP server, which I love to use that have Claude like summarize my Slack for me.

And so there, there's quite a lot of usage for that. Like we, we even have like an MCP server where like we're doing like a, a semi, a biannual like survey, for example, around like how we, how we feel like about the company, about the future, about AI, about safety, these type of things.

And we have an MCP server for that, and I can ask Claude questions around the results, which is really fun.

Is it your team maintaining it?

Uh, no. We maintain a gateway.

Some, some... Yeah.

But like I think one of the fun part is like when we started MCP, it was always, like MCP before we even open sourced it, it was born of the idea of like I'm in a company that is growing crazy.

I'm in the development side of things, development tooling side of things. I will grow slower than the rest. How can I build something that they can all build for themselves? And that's really the origin story of MCP. And so it's fun to see a year later, like that's what's actually going on is like people build MCP server for themselves.

I probably don't even know ninety percent of the MCP servers at Anthropic- ... because, you know, they might be in research, and I might not even see them, or I just don't know because people build for themselves, so.

But do they, do they host it themselves? Is there a re-remote?

They effectively have a command to launch it, and it just launches in like a, in a Kubernetes cluster for them. So it's like partially managed.

Yeah. That's good info for anyone at a large company-

Yeah

... to, to build.

Yeah.

Any platform infra person.

And some, there are platforms that offer that to you. For us, from a security perspective, we wanna build these ourselves.

Yeah.

But like they're like, um, the person who built, uh, um, FastMCP, Jeremiah, like-

Yes. Oh

... he's a company that offers like FastMCP Cloud, which is a little bit like that. You just like two commands, and you have a running instance of an MCP server that, that talks stream over HTTP. And then a lot of inter-- like a lot of enterprises use things like LiteLLM as a gateway, and then they can even do like just launch standard IO servers, attach them to the-

Mm-hmm

... to the gateway, and the gateway does all the authentication, all the hard parts of MCP for them. And so there's a lot of ways to do this, but that's good infrastructure you really want to have is just like make it trivial.

Make it like one command to just launch an MCP server that was a standard IO server and suddenly is a stream of HTTP server with authentication integrated. And you as a, a developer only had to do the standard IO part.

Yeah. Uh, I love calling that stuff out because people will take that and actually put this into their companies. So, um, yeah. Otherwise, also the alternative is chaos.

Oh, it- ... 100%.

Now every... Reinventing everything.

Yeah.

Uh, shout out to Jeremiah. Actually, I, I did invite him to do a workshop on FastMCP.

Yeah.

At, uh, my New York summit

He had, he had, he had recently a very great blog post about, like, a lot of the usage of MCP we're actually seeing is internal in companies.

That's actually what we see at the moment, too.

It's really cool. Like-

In, in what companies?

Internally in companies. In big enterprises, you see MCP everywhere, and it's, it's actually way growing way faster than you would think because it's mostly internal to companies and without people seeing it.

About discovery, so you launched a registry.

Yeah.

There were registry companies, there were gateway companies.

Yeah.

The official registry now has other registries putting their own MCP hitting in your official registry.

We need more registries, man.

I mean ...

Just one more, bro, one more. Just-

Yeah. What, what's the-

One registry to rule them all

... any, any learning from that, like launching a registry for like a new technology and, like, whether or not, you know, people... Like, y- you know, Smithery is one example, right? If you go on the official registries, like, all these Smithery AI-

Yeah

... MCPs that you need to authenticate through them.

Yeah.

So it's kinda like just a pass-through-

Yeah

... registry in a way. How do you see how is this gonna shake out?

I think we saw a lot of these, like, different registries come up, and we've really felt that there is a need for basically like an npm, PyPi kind of approach to this, where, like, there's one more central entity that, that, that is the, the where everybody can publish an m- MCP server to, and that's really where the original registry came from.

And we really wanted to make sure that at least we're encouraging the ecosystem to have a common standard of what these registries can talk to. Because what we want to do, we wanna live in a world where a model can auto-select, um, an MCP server from a registry, install it, and then all the given tasks that you have at hand, and then you just use it, right?

It should kind of feel magic, but for that you need, like, some form of standardized interface. And so we've gotta do... And that was really the, the inflection point of, like, we started quite early working with the GitHub folks, even in April, um, and then I got distracted- ...

uh, with other things, like authentication, um, and worked on that. And so what I wanna see, and I think where we slowly, but this is slowly heading, is a world where, um, we have the official registry where everybody can put their MCP server, but this is the equivalent to an npm, which has the exact same problems of an npm.

Like, um, everybody can put it there. There's basically you don't know what to trust and what not to trust. You have supply chain attacks.

Mm.

These are just fundamental properties of public registries, and that's why we have this concept of sub-registries, which then, like the Smithery's and others hopefully can do, where they can filter and curate on top of it, and that's the, that's really the world we wanna live in.

I don't think we're quite there yet, but we're slowly getting there.

Yeah.

Like, the GitHub registry is, is curated off the... or speaks the same format as the, as the official registry. And so what we want is, like, you as an ent- as a company, uh, can have an internal registry that is a curated form of the a- the o- the official one, plus maybe your own ones, and then that's the one you trust.

And it speaks the same API than the official one, and if you have, like, a VS Code or anything else that wants to talk to a registry, you just connect it to yours and you're, you're good to go, and that's, that's really what we wanna do.

It's interesting because npm, in, in a way, it's almost like a download gateway, you know? It's like I'm not really using npm for discovery that often. I don't go to npm and search for packages. It's kinda like I find them in other ways, and, um-

This is Eden. Yeah.

Yeah. I'm interested if you see, like, discovery as, like, a core piece of, like, the registry, or, like, if you still assume that, like, there's gonna be some other way that the agent discovers.

I th- I do think discovery is important in the, in the, in the mo- in the model world, but I think that that's where it's different from, from npm, because we're building, like, something for AI first, and we can assume there's an intelligent model that knows what it wants.

Um, I think that's something that didn't exist before, right?

Mm-hmm.

If you, uh, maybe, I don't know, if you would build modern package management systems with models at heart, maybe you would do a similar approach of just like, "Here's what I wanna build. Just figure out. I don't care what, what packages you install.

Just do it," right? I mean, that's the equivalent in the end of the day. But again, with the public registry, you should probably not do this because it's a dump-

Right

... it's a dumping ground for everybody. You wanna do it against the, as against the curated, trusted registry.

I like your phrasing that the model knows what it wants.

Yeah.

Uh, because I think there, there's a lot of... there's a dream that peop- that agents can use the MCP directories to discover new, new servers, install it for itself.

Yeah.

That, that seems, like, very AGI if it works.

Yes.

But it may not work, and I wonder what needs to happen in order to do that.

I do think we need, like, a good registry interface on one hand, and then the second part is just, like, we need to build for this and see-

Yeah

... what works and what doesn't.

We need, like, trust levels maybe.

You definitely need trust levels. You need trust levels. You might need some form of, like... Yeah, you need trust levels. You might need some form of signatures. For example, like one of the ideas, I'm not sure if we're gonna do it, it's just a random idea, but one of the ideas I always had is, like, you can attach, like, signatures from, like, different model providers that have scanned this MCP server and say, "We trust this.

Here's the signature from Anthropic that these tool descriptions are safe, and here's the, like, the signature from OpenAI that these are trusted by us," and then you can decide.

Oh, wow.

So I think these-

Distributed code signing.

It made a circular And it's not just really distributed, it's just, like, central in a way, right? But I think this is the kind of stuff you will require. Or, but I think in the simplest form, um, what you can do, where you probably see it first is in, in scenarios like lo- like internally to a company where you have inherent trust because they will use a private registry.

They're effectively using private registries already for npm, they're using for PyPi, and they will also do it for MCP servers. In there you have implicit trust, and then you can just search.

Yeah.

Right? And I think that's really the interesting ground where we wanna, where we wanna experiment, and that's like we have our internal registry effectively because when you launch, like, um, via John's infrastructure, like an MCP server, it gets registered, right?

And so we, we need to go in and experiment with that too.

Okay. I actually wanted to also ask, you, you started running some events over in London.

Yeah.

Uh, you had the Agents Hackathon, and you had DevSummit that you called on your timeline.

Yeah.

I just wanted to get- Anecdotal stories of stuff you learned as you, as you saw the community spring to life

So we had two big summits this year. We had the, the MCP Dev Summit in San Francisco, um-

And the one in London too. Yeah

And the one in London, and I think what you learn is a few things. I think the, the one thing is that you-- that's very hard to get otherwise is just, like, these stories around how people use it internally in their companies.

In there, you see some of the struggles, but you see also some of the success stories. And one of the interesting bits that, that... wh-which I really loved is, like, particularly in London, you had a lot of financial people there because it's, like, clearly a financial hub, and it was actually the, the whole conference was in the financial district.

And learning just, like, the kind of problems of, like, um, things you need to enforce because you have legal contracts, because, like, financial, like, regulations. These were things that were... that I did not know before, and I learned a lot about, like, okay, what does, like, a thing like an MCP, like a communication layer need to look like if you have these, like, constraints that in a normal, like, development world doesn't exist?

Like, I give you an example. Like, um, if you are in f- in financial services and you're exposing some data, you, that data might be coming from a third party, and you must guarantee that you attribute that third party, and that's a legal contract, right?

You must-- Like, they, if the client displays this data to you, it must tell you, "This came from this third party," right? And these are constraints that just, like, in a normal model world don't really exist, right? But, like, in the financial industry, this is, like, uh, legally enforced.

And so these are the things you're like, "Okay, how, how will, how will this work in a world where MC- where for MCP?" And so now that's when we, like, started, like, creating this financial services interest group that Bloomberg is heading up, um, to, like, figure out, like, what are some of the things that you would like a client must do if it wants to speak to a financial services MCP server, for example.

And, you know, what are the things that need to be respected? And I think that's the kind of things you only learn on the ground in the conferences talking to people, right? So I think there was some of these learnings there.

I think the other things that you just see is, like, just how many people are building and just the excitement and, like, the creativity that some people bring to this, like, that I just love, right? Like, and, and from, from areas you didn't expect, right?

Like, I love the guys at Turkish Airlines who just built, like, the Turkish Airlines MCP server. You can search for flights and stuff like that, so that was always fun. Um, I love when, like, people bring some really creative parts to, to the MCP ecosystem.

So I, I love these community when they come together because you're just meeting things that are a little bit outside of your bubble, and I, and you just get some input, and I think there's a lot of learning there.

And some we're gonna repeat it again. We're gonna do it in New York in April, I think, or, or March or something like that, and then we're gonna do it again six months later, and I absolutely love that.

Any good, uh, sampling use cases that you found?

Uh, not so much.

Okay. Yeah.

Uh-

And that, that's always the, like, you know, last time we talked about this-

Yeah.

The future user more-

We should fix sampling a little bit, man. Like, we... I think one thing I learned from sampling, like, everyone wants to use tools with sampling and tools that are not exposed via the MCP server. Like, when you wanna do sampling, you wanna have a set of new tools that you only wanna use during that sample call, and we just had no ability to do it, and then we just fixed this in this iteration.

Um, and so we hope to see a little bit more, bit of more sampling use cases. You will find every now and then an MCP server that does it. But particularly as MCP servers have moved away from more local to be more remote, in remote cases, it's probably always better for you to bring an SDK because you have full control.

You can deploy it, you can deploy, uh, API keys, and maybe even charge someone. In a local case where sampling is really powerful because you're shipping something to a lot of people, and you don't know what is their...

what is the model that they have configured, what is the application they plug it in. Might be VS Code, might be Cloud Desktop, right? And in those cases, sampling is useful. But also, like, clients just don't support it.

So sampling is one of these things I'm like, I'm still sad about it. I still think it's a very powerful idea. But you know, you gotta, you gotta win some, you gotta lose some, you know.

No, no, no. But, but you, you're also, you know, u-upgrading it, and, um, you know, in, in some ways-

My hopes are still up there. I'm just like-

Yeah, yeah

... weird.

Like my... In, in some ways, you know, when you get it right, uh, this will be the real agent-to-agent protocol.

Yes, yes.

Are most of the use cases that you see still data consumption? That's been my use case for MCP mostly. It's like-

Yeah, it's context engineering

... getting, getting data. Uh, well, the, the most action my MCP takes is, like, update the linear task status. Um, have you seen very complex, like, MCP taking action workflows, or are still people mostly using it for context?

Most people use it for context. I think that's the vast majority of usage.

It is in the name, mo-

Model Context

... Model Context.

Yeah, yeah, yeah. And, and, and Nick Cooper from, from OpenAI always, uh, keeps telling me, rightfully so, that the name MCP was probably a little bit poorly chosen, um, um, because it rest-- like it feels it restricts it a little bit, um, which I agree with.

Um, I, I... It's mostly data use cases. I've seen, like, people doing deep research via it.

Mm.

I think people expose agents via it, and so they are, they're a little bit more complex. Uh, but it's, it's not super common. It's... But people like to have experiment with it. Like, they have, like, the deep research use case I think is a good one, um, that is very, very, that is, that's not too uncommon where people do, like, custom research-

Yeah

... um, for it. But beyond that, um, yeah, I, I... Most of it is really data. Beyond data and deep research aspects now you have also this new aspect where people expose, like, UI components through mcpui or what we're gonna call mcpis in the future.

Yes.

Um, and I think that's super, super promising, and I think that's really quite fun. That's actually you see a lot now with ChatGPT apps, with mcpui in general-

Right

... that you see a lot.

Yeah, and you have the tasks in the last, like, release.

And we have tasks, yes.

Um, yeah. Well, uh, I mean, I'm curious because, like, if most use cases are, like, context, and then you build tasks, it's almost like people are not really using it for tasks.

Oh.

So I'm curious, like, how you design it, like what you expect people to use that for.

We design tasks because people come to us and go like, "Okay, we really want long-running operation," which is basically agents. We want-

Right

... like a long, deep, deep research task that, that Finish this in an hour, or we want tasks that like might not finish within a day, right? And people have like awkwardly tried to do this with our tools, and you can, because tools are effectively just an RPC interface-

Right

... at the end of the day. But it gets very quickly awkward because now the model needs to understand, oh, I need to pull this, and it, it, it's just not very fun. And like it's just not a first-class primitive, and you run into a lot of limitations.

But it's, it's come from the fact that people want to have long-running agents, and that's like something we heard from so many areas and, and people trying to do this, that we, like we really felt we needed to do something like Task.

Like, um, on GitHub issues from big companies, everybody was like-

Mm

... we need, we need something that long-running operations is really top of mind. So I really think now we're gonna see a lot of it now, but it's, it's a little bit early to see how good it's gonna go because it just landed in the SDKs-

Yeah

... and it needs to land, uh, in the clients, and then we're gonna, we're gonna see more of it. But I will-- I definitely-- Like, I do-- I think you will see a lot of the custom deep research parts with it in others.

Yeah. I, I'm very bullish on Tasks. I think, uh, it was very important to get right. Uh, basically every orchestration or protocol needs-- has a sync version and an async version.

Yeah, um, exactly.

Has an async version. Any like design choices that you wanna call out that, you know, you-- there were, there were two directions and you picked one, w- uh, in, in just the overall design of Tasks?

Uh, yeah, in design, there was a lot of, uh, con-uh, conversations like some, some were like, "Okay, is this just asynchronous tools? Do we do a different primitives?" In the end of the day, it was important for me, my litmus test for it was always, it needs to be able to like, if I wanna expose something like Claude Code or like any other like, like coding agent as an MCP server, hypothetically, this needs to work.

And a pure asynchronous tool call would just not do this. You want some form of operation that can return, for example, intermediate results in the long term. You want like, okay, I got to this result by calling this tool, this tool, this tool.

I had this other input, I had this other tool, I did this, and now this is the result, right? That's really what you wanna expose. Um, Task is, is early, and it doesn't do that just yet, but it's built in a way that it will be generic enough to be able to support this.

That was the main constraint. The other constraint was, um, making sure, um, it is, um, it's, it's not a copy of tools where you can, you can think about like, okay, we just do tools again, have slightly different semantics.

Um, but instead, what it's doing is like this, like you can create a task by calling a tool with a certain set of like metadata fields, and then it automatically creates a task. So the task itself is just the concept of like a container that can-

As a message, message sent

... do something asynchronously. Yeah. Like, just like we do something asynchronously-

Mm

... from starting here and to ending here, and the thing we're doing is a tool call. I mean, that opens the door to like later plug in other things and maybe even other tasks.

Like observability as well.

Yeah, potentially.

Uh, which is obviously gonna be important.

Yeah. So I think, um, that was really the design goal, which makes it a little bit more abstract, a little bit more complicated to implement, but that goes away because the SDKs just do it for you. In the SDK, and then over there, you're just gonna like async call this-

Yeah

... and you return something.

I mean, there you start to overlap with, uh, other async like tRPC in, in JavaScript land or, uh, you know, uh, whatever, whatever Go protobuf stuff that Go people have.

Yeah. In the end of the day, it's like it's-

Yeah

... it's, it's designed like a, a classic, um, uh-

RPC

... operating system interface. Like you create a task, you pull it until it's done, and then you can make an optimization, which we're gonna do in the next round, which we didn't get around, is like, okay, instead of having to pull every minute or hour or whatever you interval you choose, the server can-

Service call events

... call, call you like, or like, uh, like a webhook or something and go like, "I'm done," right? That's the optimization. But the actual core interface is always that the client can pull, and that's actually how like operating file system operations on an operating system can work is like you pull, it has the file changed, has the file changed?

But you can also use like a modern interface on the Chrome, like INotify or something like that, um, that tell or, or a Urring or something like that to tell you, "Oh, I'm done."

Right. Yeah.

The, the file has changed.

There's a trick I, I learned, uh, where like servers can hold the HTTP connection until it's done, and then they terminate, and that's the signal to, to the callback.

Yeah, yeah. Which we do not necessarily wanna do because it might take a few days, and I don't want to have people-

It's very irresponsible, but it's cool.

Yeah, yeah, yeah.

Yeah.

There are plenty of ways. I think we were just gonna go the webhook way, honestly.

Tasks are really interesting a- and we basically have to invent this, um, when we did this at, um, did like the Devin API, uh, at, at Cognition, and I think that's also like, uh, an interesting reinvention of, of like, well, everyone is going to need some kind of long-running operation.

Yeah.

And, and this is-- Well, when you're calling in agents, you also need this.

Yeah, yeah. But the interesting part for us, I think what, what MCP is always trying to do, MCP always tries to encapsulate what currently people trying to do, and we not wanna be prescriptive what you're supposed to do in a year from now.

We don't know. We don't predict.

Oh.

We, we did Tasks because people are like, "We need this now," right? "We needed this basically six months ago," and we're like, "Okay, I guess now is time to do this," right? Instead of trying to do being predictive of, of the future, which is why we're trying to keep the protocol somewhat minimal and have, I think to some degree achieved this, although other people would think already there's too many primitives in the protocol.

One minor thing, and so let's say, let's say super long-running tasks.

Yeah.

Um, lots of messages go back and forth. Anthropic, uh, actually was a kind of leader in context compression or-

Yeah

... uh, in compaction maybe let's, let's call it, and I think a lot of the other labs are also doing the same thing. Is there a way to handle that or do we just statelessly sort of cut context and it's fine?

Do you need a, a full log of everything that happens-

No

... or you just ask the file? Yeah, right.

No. No, you don't. Like I think we-- I think there-- This is the thing, right? We're very early in the industry still. We're learning a lot about like what does the model need, what does not need, right? Um, and even today, like Some agents start to, like, drop tool call results after a few rounds because they n- don't need it anymore.

And I think that's very, very, very good. And so I think be- besides compaction, you will see, I, um, just better mechanics of, like, understanding what you need and what you don't need. Like, for a long asynchronous task, you might have a, a way where like, okay, maybe for a while the model sees it, but once we get the result, you just drop everything else.

Or you might, might even call, like, a small model, like a haiku model, and go like, "What of this I should retain? Tell me," right? You can like... You might be like the AGI build approach would be just like, let the model figure out what it needs to retain, right?

And so you can, you can see both worlds. And, um, and I think there's just lots to learn. I think there's not the one answer yet because I think we're still figuring these type of things out, and we're just improving.

And compaction, compaction is, is a good step for it, but I don't think it's the last step there either. It is actually the, the most obvious one, but I don't think it's like... I think if you pay more attention to it, if you particularly think about like, okay, what could you train a model to do here, I think we get to much better ways of doing that.

But they're all like independent from how you obtain the context. And I think MCP I always see as like back to like it's an application layer protocol. That's just how you obtain the context, how you select the context that the problem for the application, and that's the problem all the agent applications will have in the end of the day, and there will be a lot of different techniques.

A year ago, everybody would've told you it's RAG style stuff. That's now apparently dead, right? And now we do use models, we use compaction, so I don't know what's gonna happen in a year from now.

Cool.

Around MCPs, another question I had is like how do you see them as being used by developers to build AI apps versus being a protocol for AI consumers to plug things in? I, I think that's one of the main things people get wrong, where it's like, well, I can just use a REST API.

Why do I need MCP? And to me it's almost like it's not really for the developers to use. It's for, like, people using AI tools to just plug things in.

I get the comparison with the REST API is quite a lot, and I, I think there's interest-- It's funny enough because there's two problems in general. Like, the first one is REST does not tell you what to do on authentication.

Right.

The second part is everybody complains to me about, like, tool bloat. But have you looked at, like, the average OpenAPI spec lengths?

Mm.

If you put that into a model, like, you will have a lot of bloat there too, right? Actually way worse. And funny enough, when people try to, like, map one-to-one things, the- often the model gets slightly confused because you have, like, search by name, search by ID, search by something, right?

And, like, suddenly you have, like, five tools that look very similar to each other, and the model goes like, "Which one do you want," right? "I have no clue anymore." Um, so anyway, that side note to, to REST versus MCP.

But I do think MCP, I wanna live in a world where it's, like, very s- like, much like a consumer-focused thing. That's something consumers should know about. For them, what I want is, I want a world where you go to your application, you say, "Do this," and it should just do the thing, and it should just connect with the right services.

That MCP is under the hood is a detail for... that the developer needed to know about, because that's the communication channel that they're talking. But in the end of the day, you just get the, the task done, right?

And I, and I actually prefer a world where nobody of-- Like, my mom should not know what MCP is, right? If she wants to use Claude at the end of the day. Um, but I do think it's very focused on, on that pluggability of, like, an external, like, service, and in that regard, more like on the consumer-focused side.

And there, there's still use cases for developers in general, like first of all, as builders, but also, like, I still love my Playwright MCP server, man.

Yeah.

Yeah, but-

Well, I think Chrome developer, uh, tool, the new Chrome one is, like, the new-

Right

... the new meta.

I, I also understand, like, for developers, right, like that run, like, Claude code locally, you know, like things like-

Yeah

... HIs can be better approached, right? To some degree. And that's okay.

I'm curious about the MCP apps UI-

Yeah

... with what you're talking about, where it's like every client... Like, ChatGPT has their own, right?

ChatGPT.

So it's like if I'm used to the MCP app of this product, but then if I go in ChatGPT, there's, like, a different version that they curate it.

Yeah.

It's kinda like a different experience. So I'm curious how you feel about that. Like, do you feel-- Especially now that you have OpenAI and the foundation, right? Do you feel like all of this will be MCP-backed in the same structure?

Yeah. There's two influences, right? Like MCP-Y existed as a project, um, which had a lot of really good ideas. Um, OpenAI took some of them and really improved upon them. And now one thing we just announced three weeks ago on the MCP blog is that we're actually working with all, all two of them together to build, like, a common standard.

And so we're really hoping that we're getting back to the world where you build for one platform, but you can use it across all of them. You build for, uh-

Write once, run everywhere

... for ChatGPT, and you might be able to use it in Claude or in the Goose-

Mm

... or whatever it might be the, the program of your choice that implements this. But I think the, the, the, the general problem is what we have is I think that there are certain problems, like if you, if you think about a modern AI application, everything is very text-based.

And that's okay. It's nice. But there's things that as a human, you're just way better suited to do-

Yeah

... in, in visual, right? The most basic example is, like, you want to, like, book a flight. Seat selection, right? Like, you now get select, like, if you wanna do seat selection in text, it's like, "Here's, like, the 25 seats you have available."

Like, nobody fucking wants to do that, right? Like, I have no clue where these seats are even.

That shoe-based drawing.

Yeah. Of course you want an application that you can select with, or it might be like a theater that you wanna book for or something like that. It's so obvious that you do want to have some form of, like, an application and a user interface that the model can navigate and that the model can, um, interact with, but you as a human can also interact with at the same time.

And I think that's what we're looking for. And so I think it's just this next iteration of, like, building richer interfaces, because the pure text interface is just somewhat limited, and there's very natural things. And you, like, you see this in music production, you will see it, of course, or you will have, like, certain brands that will deeply care about presenting their interface.

Shopping is a good example, man. Like, um, like Shopping has like 20 years of like AB testing, what's the best way to sell you something, right? And so, and, and shopping interfaces are super complicated actually. Um, and you just want a way for, for displaying that to the user so that it's familiar to them and that they, they can interact with that.

And that's what MCP Apps is in the end of the day.

Yeah. And technical direction-wise, is the iframe the, the way to-

Yeah, it's an iframe. It's, it's, um, you are serving basically raw HTML over, um, uh, over an MCP resource. It goes into an iframe, and then it talks to the outside via, um, post messages over a specific interface.

And so what you can do now, you can-- because it's raw HTML, and you're not real- like loading some external content, you're gonna get, uh... You can analyze it in advance if you wanted to with security. And because you have an iframe, you can-- like the external application, like, can just speak like a very clear bounded or like security bounded, um-

Yeah, and this, this has been in browsers forever. I think the, the... I'm scared of it only because I hate CORS issues.

Yeah.

And iframes always have CORS issues.

Yeah. But this, again, this does not load ex- anything external. Like, it is surely-- like, it should not, right? Like, there probably are restrictions that we'd like then iterate and iterate, and then in five years, maybe it has like- ...

25 CORS headers and whatnot. Whatever, right? But, um, but I think at, at-- we, we're starting small again with like pure raw HTML. You should probably not have external references, so you don't run into these issues. But you're right.

And can I inherit styles?

Uh-

No, because iframe, you know.

Yeah, I think you need to put it in line.

Yeah, obviously, like, like you will want it-- And I feel like this is really minor, but UI people care about this.

Yeah.

Which is you have it, it should look like ChatGPT.

Yeah, yeah.

The inside of ChatGPT should look cloud, inside cloud.

I think that's a very good question. I a hundred percent agree with you, like brands and others will deeply, deeply care about it.

Designers will, will for sure.

Oh, 100%. 100%, yes. And that's something we need to figure out, and that's where we need to get it out of the door and see how people use it and then iterate on it.

That's, that's why I don't think it should be an iframe long term. I don't, I don't know what, I don't know what the solution is.

Okay.

But like, we need like new iframe that-

Yeah

... lets some permeability because of this stuff.

No, no, I think, I think that's sensible, yes.

Um, well, I don't know.

But the other solution to the problem is the AGI build approach of, like, just give it a tool that says, "Give me a style sheet." And-

Well, so okay-

... the model can call you and tell you what you're supposed to look like.

Okay. Should an MCP app be-- know what it's being used, what the parent application is, is? You know what I mean? Like-

It, it might be. Like, the, the, the application auth exposes tools, right, that the model is free to call it.

Right, right, right. Okay, so maybe standardize the interface for people to pass down styles.

Maybe, yeah. Maybe. I don't know. But it's a, it's a very good question. Let me, let me ask the team. I'm just like, I'm mostly right directionally there. I'm like not in the- ... in the weeds of doing everything there.

Yeah. It, it, it seems like a, a little bit of a surprise to me. I, I never really paid any attention to MCP UI, and then suddenly you guys all adopted it, and I was like, "Okay, well, I guess this is a part of MCP now."

Yeah.

And it went from a purely back-end concern to now-

Yeah

... front end.

It's also like Node was technically an extension to MCP. Like, it's not MCP, MCP. Um, that's a pure technicality because-

It's a governance thing, right?

Yeah. It's mostly like if you are a client that can render HTML, then you might wanna consider implementing it, but you, you're still an MCP client if you don't. And the reality is, is like your average, like, CLI agent can't do it, right?

So they will never do it. Um, and so I think that's fine.

Are there any other extensions that are similar?

Uh, we gotta look into like financial services as an extension.

Yeah.

We're like, okay, you might, you might end up in a world where in a year from now there might be clients that have like certifications that they are an amp, like-

Mm.

... and get like, um, a signature that they are like financial services MCP clients, and they can prove it through the server, and only then the server allows connections because it knows they are respecting attributions, these legal contracts that you put into place.

And you will see this everywhere. You will, if you wanna deal in the long run with a public service and public, uh, clients that do, like, deal with HIPAA data, like finan- like healthcare data, you will have to have guarantees.

Isn't that part of just auth or auth is-

Uh, not necessarily. Like, I give you an example. Like, if I have... The client might need to have, have five servers installed, and if there's one healthcare server, that healthcare server might tell you, "You are not allowed in this session to use any of the other MCP servers because this data I'm giving you cannot leave you," right?

"You must guarantee that this data doesn't go anywhere else because it's HIPAA data, because it's financial data," whatever it might be. This is a good example, and that might be some of the enforcements you need to do.

Mm-hmm.

Because you're just like, yeah, you don't want to have your, or your social security number or your healthcare data show up in a linear accident, right?

Yeah. Azam, we're gonna transition and have the rest of the AAIF, um, group join. But any final call to action, like either, you know, people that should join your team, people that should contribute to the MCP spec or anything else?

I think the most important part is still building with MCP on a, on a day-to-day basis for people to just go out, build really good MCP servers. I think we see a lot of mediocre MCP servers, and it's...

I mean, some very, very good ones. Um, and just building good MCP servers, looking at like how to use them, um, I think that's super important. The second aspect to that is like we're a fully open community, and we're running it as a, as a traditional open source project that is based purely on like what people are able to put in in terms of, uh, in terms of effort and time.

And so just like being an active part, like either giving us feedback, being in the Discord channel, talking with us, giving us ideas, while also just helping us implementing like the TypeScript SDKs, the Python SDKs. We're always looking for new SDKs, right?

Like we have active Go SDK development, but like we don't have a Haskell SDK. I don't know, if you're a Haskell developer-

Mm.

... maybe you wanna write that.

Right.

Right? Yeah, there you go. And so I think there's a bunch of stuff we can do and, and be part of it. And I think, um, yeah, don't underestimate how much you can just be part of the community, but also just like, uh, go and build.

Um, and I think there's so much opportunity now particularly to build like amazing clients now that we have understood progressive discovery better-

Mm

... now that we have understood code mode better. There's just this next iteration of clients to build and the next iteration of servers to build that I'm just looking forward for people to do. Yeah.

Um, my, my last, uh, question or call-out is, uh, I wanted people to hear directly from you. I sense the energy. I'm very, uh, excited by everything that you're doing. Um, but a lot of people are, are anxious about joining-- MCP joining the Linux Foundation.

Yeah.

They're like, "Oh, is this Anthropic taking its eye off the ball?"

Yeah.

Can you address those concerns?

Yeah. I love that you asked me that. Like, I think, yeah, yeah, I can totally see why people think that, but, like, it's actually quite the opposite.

Yeah.

Like, the commitment of Anthropic is the same, right? I'm still-- We still have the same people. I'm helping with the SDKs. We're still super committed in our products to MCP. I'm still the lead core maintainer. Nothing has actually changed.

What really is the main part of the foundation is, like, two things. The number one is, like, making sure that the whole industry knows that this will stay forever open, that this cannot be taken away. And there have been, there have been-- Like, Anthropic would never do this, I think, but there have been histories of, like, companies going, like, taking an open source project and suddenly making it proprietary again.

Mm-hmm.

We have protocols that are proprietary. Look at HDMI. Look at, like, what's the problems of HDMI in Linux.

Wait, what, what's wrong, what's wrong with HDMI?

Uh, H-HDMI 2.1. The HDMI Forum does not wanna allow the AMD to develop open source Linux drivers for HDMI 2.1. Really. There, there's some... Look it up.

Wow.

Um-

I love it

... so, you know, there's people that, like, keep a very close tab on it, and what this does is, like, no, this is now owned by a neutral entity. It will always stay open. You can use the, the, the word MCP.

Nobody's gonna sue you over it. So there's a bunch of, like, just giving the ecosystem and the industry that confidence that this stays neutral. I think that's important. The second part to that is that I think o-one thing I'm-- If anything, I'm the most proud of is that I think we have set the, the tone for open standards in the, in the industry, and being able to now use that momentum to build, like, community and a space where people can come and bring really well-done, well-supported, m-well-maintained, uh, projects and have, and have them part of this foundation.

I think that's the, that's the other part to that. But the funny part is, like, our bar for the, for the foundation is gonna be like, it needs to be, like, really well-maintained.

Mm-hmm.

It's not like you getting-- you taking the ball off. It's actually exactly that what we don't want, and so we will not do that. For us, MCP is still core to the product and still super important to Anthropic, and so, um, we're still just as much as committed as we've ever been.

Amazing.

Awesome. Thanks for joining, David.

And we're here in the studio with core team members of AAIF. Um, it's, it's, uh, it's the biggest panel we've ever had on, on the podcast, so welcome, guys. Uh, maybe we'll go left to right and introduce everyone, and also identify the voices for people listening on audio.

Uh, I'll start. I'm Jim Zemlin. I'm the, uh, CEO of the Linux Foundation. I've been working there twenty-two years, and, um, I was the person who helped facilitate the launch of the, uh, foundation, uh, but take no credit for any of the technology work that ha- That's to my, my left.

I'm Nick Cooper from OpenAI. I've been there just over two years now, I think. Uh, I'm generally OpenAI's, like, head of a lot of protocol things and very interested in the open ecosystem and our representative for AAIF, as well as a core contributor to MCP.

Got it. What, what's another protocol that might fall under that umbrella?

Uh, Agents and D. Just in general, like, not just the protocols, but also the product experiences of where OpenAI products intersect with other SaaS provider things and other systems.

Uh, I'm David Soria Parra. I am the-- working at Anthropic, member of technical staff there. I'm the co-creator of MCP and, um, yeah, at Anthropic, I mostly lead all the MCP efforts.

Great. And I'm Brad. Uh, I'm the principal engineer at Block. Uh, so by day, I build AI products, and by night, I work on open source like Goose and am the original author of Goose.

It's great to see everybody come together. I think when I heard about the, the news, I didn't really expect it. It wasn't on, on my bingo card, so maybe let's have a little bit of inside baseball. So you obviously have OpenAI and Anthropic, and yesterday at the launch event, you were joking on how, uh, you didn't know that the two companies even talk to each other.

Uh, and then, yeah, how did the conversation start?

The conversation started out of two things. The first one is that on the MCP side, we always knew that we wanted to find a neutral home for MCP to make sure that the, uh, that the industry understands that, uh, this is, uh, stays open, that this is, uh, something safe to adopt.

Um, and then, um, very early in the process, as we were looking around, like, what to do about this, should this be a project i-in a foundation? Should this be inside its own foundation? Which is, like, these co- the common patterns you see for this kind of work.

And we got approached by, by the, our, our friends at Block to, um, discuss, uh, because they were looking into, like, donating, uh, Goose, I think, at the time. And so th-there was a question around doing something together.

And then, um, we approached, um, uh, OpenAI, and, and they were very, very, uh, welcoming and, like, very open to the idea as well. And, and it slowly, like, formed. And I think, you know, the timeframe of this is, like, a few months.

These things are not happening out of thin air in, like, a week or so. And so just a lot of conversation, like, what do we wanna do? What are the kinda, like, constraints we wanna have, and what is the thing we wanna build?

And of course, we were looking for where to put this kind of stuff and, and that's where the Linux Foundation comes in as, as, um, I think the biggest, uh, foundation of its kind and certainly, uh, has, like, the, like, decades of experience helping companies through a process like this, uh, and building what is a tech- what's technically called a directed fund within the Linux Foundation to build these kind of things out.

I think David said basically all the story from my side as well, which is, um, so we saw this, like, need to connect systems, and then MCP gained such very large developer traction, and we at OpenAI were very excited to, like- use and then contribute and actively participate in this.

And from my point of view, it was always very natural that this would grow into something bigger and move to a neutral place. And like MCP has always been like a foundation for like communication between agents and contexts.

In a similar way, the Agentic Foundation is, well, it's a foundation, but also it's like the starting point where we really look forward to other contributions, like starting with Goose, our own agents MD, where we're really open for like a lot of technical contributions to build out like a full agentic ecosystem.

I'm curious, Jim, uh, you know, I've been to Linux Foundation events before. I've spoken at them. Is it... It's almost like, isn't MCP so early that like how do you even structure it in a way? I'm curious because so many of the technologies that the foundation supports are kind of like core pillars of-

Yeah

... infrastructure and the internet. This is probably like the youngest technology-

Yeah

... that you brought in as a foundation. What are the goals of it?

Yeah, I mean, I, I think what's interesting here is even though it's young, I think if you-- I think AI years are kind of like dog years.

Absolutely.

It, it ju- Do, do you use this metaphor, man, but -

Yeah, totally. This is why I run three conferences a year.

Yeah, exactly.

You can't do annual.

I think last night someone was asking: What do you, what do you see a year from now? And I'm like: Well, if I dial the clock back a year, would have i-I anticipated where we're at right now? There's no way.

And so I think part of the thing with MCP is that we're just living in this kind of dog years velocity. Uh, in the past, I think things took a lot more time to coalesce, and what is clear is that a lot of people are adopting MCP.

You see it in commercial products that companies are rolling out. You see a lot of usage in the enterprise already, and there still is a ways to go in terms of the technology becoming mature. But I think the, the same thing held in internet protocols, you know, that took a little bit longer to mature, and the internet matured over time.

But I think the thing I'm most excited about, it's becoming clear that MCP will be a key protocol f- this technology movement. And I think, you know, David and these folks were all pretty wise to realize that, you know, if internet protocols had been owned by a single entity, we'd still be calling it American Onli- America Online.

Mm.

It like, you know, it would be... It wouldn't work. And, uh, I think that this has got all of the underpinnings to be a huge movement. At, at the Linux Foundation, we ask three questions for every project. Will this be meaningful and impactful for industry and society?

Uh, the second question is, do you need more than one organization to collaborate to do it, otherwise you don't need us?

Mm.

This case, clearly we've got that. Then three, can we get the resources and build an ecosystem around it? And fifty companies on day one, you know, a, a huge set of folks in line to participate and join my email inbox, I'm sure your guys' are too, is like full in twenty-four hours.

"How do I participate? We want to contribute. How do I get in there?" I've never seen that kind of inbound interest starting any project at the Linux Foundation in twenty-two years.

How do you pick? So you got all these people reaching out, you know, there's Good and Bad.

It's a really good question. I think it's like how to pick, uh, how we expand the foundation itself from like a governance standpoint, but also like technical contributions and how can the foundation best support them as well. That's like really top of my mind.

It's like the first thing we need to like define some structure and work out what we... Like how to bring it all together. But I think even before like those details, there's such value in establishing this one forum that people can come to.

Like even having a list of eager technical participants and potential opportunities, that's a huge opportunity in front of us to like distill what's truly meaningful to developers, users and everyone, and very appreciative the Linux Foundation for acting as like sort of a galvanizing rod for this like attention only.

Brad, on, on the, uh, sort of Block and Goose side, it's an interesting... The involvement that you guys have had and the sort of engagement that you guys have had, uh, what was your calculus in joining the AAIF?

So for us, I think it's, uh, like in developing something like Goose, I think it, uh, the thing that I see it as being part of this umbrella is it's the most con-concrete piece. So you can actually like download Goose and use it in a way that you can-

You can download an agents MD?

Yeah. Like, right. Like what, what do those parts do together without having something that actually like connects it to client-

Yeah

... like a real client. And there's I think a lot of value in that because when you, you get into the co-protocol space, you wanna add things to it, but you have to actually show like what is enabling, like why are you making the protocol wider?

And putting it into like a reference implementation shows you like: Oh, it's giving this value to people, like very concretely. And I think there's some like, for example, there's like a, a spec, uh, for MCP, MCP apps that is brand new, and we've been working on MCP UI.

For Goose?

Yeah. So, so Goose has been, uh, like kind of a day one partner with the MCP UI team.

Oh, I didn't know that.

And so now we've got MCP apps. We'll go... We, we have opened an issue today about how we're gonna go get that into Goose. And so that's something where I'm... People, I think, you hear something abstract like that, like what is send-- what is the server sending an iframe to the client like do?

And, and I think Goose is a place where you can see it. Like, okay, you're gonna build a dashboard, or you're gonna have this enhanced chat experience. And so this is something where I think we collaborate more and more to say like, "This is what it looks like and how you look at some of these abstract things and make them real."

Yeah, I think the other tidbit here, maybe like back to the history of, of both MCP and Goose is like, um, Goose was the first open source agent interface or agent that reached out to us and worked with us to integrate MCP, and I think Brad is actually like technically the first non-Anthropic contributor to MCP ever, uh, on like day two or something like that, like very, very early.

So this goes all the way back to like November last year to like- the partnership of having MCP inside Goose

Yeah, we had, uh, we had a, a version of, of Goose, um, that was still... You can go check the GitHub history. It was there a little bit before MCP came out, and we were sitting there with, like, a plugin ecosystem and we were like, "This is awful."

Like, what, like, why would you, why would anyone-

Homegrown.

Right. Yeah. Why would anyone come develop a plugin just for Goose? But we saw all these opportunities, and so we started talking to Anthropic, and we were like, "I think that there's a space here for a protocol." And they're like, "Well, let me tell you about, uh..."

And it was really cool to see, you know, what the Zet site... No, no. No, we didn't. We reached out before we heard the Zet thing, and so we were like, "Okay," like, "yes, we just want to, like, we want to pile onto something that has a chance of succeeding," because as, like, a client, it's like an ecosystem, right?

Like, the more people who are using it, you get more value as a client than as a server because you know your servers are gonna work with any client, and then as a client, you know, you have this li- giant library of servers.

And so that's been, like, a big part of what Goose does, is that it's not really... Like, it is a coding tool. People use it as a coding tool. But you can turn off the code part, and you can just connect to any MCP server.

And so it can be, like, operating, like, a, you know, a science experiment, I've seen that, or, like, just, like, Google Docs or whatever, and I think that kind of shows you how MCP goes beyond just, like, the back coding space.

Yeah. I think as well it's also the fact that it's concrete is so important. Like, for all these standards, like, there's a long history of standards throughout computing that, like, people, like, proactively write a standard, and then it, when, you know, the, it's actually tried out, it has problems.

Yeah.

But, like, for MCP and, like, all these new agentic standards we're coming with, we really want demonstrated utility. Like, w- the most common thing on the core committee is, like, there's a proposal and we come back to people saying, like, "Have you tried it out?

Does it work?" But the protocol's about communication, so if you're trying something out, you need collaborators, and you need, like, concrete open source projects like Goose, and you need clients, you need a variety of servers, 'cause it's only with that sort of open ecosystem they can meaningfully understand if this is actually gonna work.

I totally agree with that. I mean, I think the world of standards and open source development are just merging, right? You sort of co-develop these things together. Um, I think I was trying to figure out whether David is Vint Cerf or br- uh, Linus Torvalds for agents.

And I think maybe it's a little more, it leans a little more Vint Cerf, and then maybe Ghost is a little more Apache web server, and my whole Linus part kind of falls apart at that point. Uh, but you do need something substantive to try the protocols out in order to make sure you know how to improve them.

It's that feedback loop that's so critical.

OpenAI also has a coding agent that is open source. I, I think what's the thinking there apart, apart from, like, well, would Codex ever be donated to AAIF, or we just don't know yet?

I think the short answer is we don't know yet. But, like, it's sort of like a feedback loop, which is, um, in this open ecosystem, like, we don't wanna have too much alignment all on, like, one implementation, one thing.

There's, like, real value to users and developers or whoever's the participant to, like, active competition in some parts. So there's this balance of, like, we need openness to foster, like, collaboration and experimentation, but, like, I would like to see a variety of coding agents and each one might deliver unique value, different value, and be free to explore independently.

So it is sort of like a careful balance here. There's a bit of, like, a taste-making approach to, like, contribute things that benefit from being open. Like, Agents.md is an example, which is open up any GitHub repository, it has this file, it works the same way.

Like, if everyone sort of did their own thing there, that's very low value and potentially damaging, by the way. But, um, so there's commonality value. Whereas for actual concrete implementations and projects, it's great to have reference implementations in many ways or experimental grounds like Goose, but I really favor a huge variety of them, 'cause that way we'll see what comes to the be the best.

Is there a roadmap for what you want to add? For example, the Agentic Commerce protocol, like, ChatGPT already uses, but that's not a part of it. Um, there's no model as a part of the foundation. Um, do you already have a roadmap, or like you said, you're just kinda, like, going month by month, seeing what are people using and what should be in there?

I think we don't have a roadmap in the sense of, like, well, projects lined up. But I think what we have is principles by which we will select products, um, to some degree. And w- I think the effort here is mostly around sitting together after the, now that the, the foundation is created, and then evolving these principles as we're seeing people going to ask us about the projects they would wanna put in, um, and then ev- develop the, the foundation, um, further as time goes.

But at the moment, I think the most important part is that we have the principles in sp- uh, in place, and then, um, you go and having the conversations with people who, who want to, uh, be part of this foundation.

One principle that, like, really comes to mind to me is, um, composability. Like, I often use the analogy of, like, Lego blocks sort of thing, which is agentic systems are a sum of many, many parts. And so, like, something that I hope the foundation can evolve to do is have these, like, interoperable composable bits that all work together seamlessly.

And so we don't have a roadmap of, like, future contributions, but, like, I welcome all contributions that play nice with other contributions and, like, really create this potentially, like, a future flexible open agentic stack to be, like, not a universal agent, but an agent that suits everyone's purpose or need.

Yeah, it, it's tricky. You got... It, it, it's a hard... And, and these guys have the harder job of early in innovation cycle, you don't wanna restrict innovation by saying, like, "Oh, well, this is the one," you know, versus "that one."

But you also don't wanna let every single random thing, you know, into an organization like this. And so I do think you need this, uh, taste-making is a, a good way to describe it, where a group of, you know, elite architects and developers, you know, folks like the three people sitting next to me are, are more curating.

And, you know, some, some things can work, some things might not, but there needs to be a process which I think we'll define, uh, to do that, that curation, um, that happens via tastemakers. And, and it's more of a, not just more, but essentially a technical effort, not something where a committee of, you know, folks from vendors get together and say, "Well, my product should be in this roadmap, and that guy's product should be in this roadmap."

That tends to not be very successful.

Yeah, which leads to this other principle that we really want projects that are-- have a lot of traction, that are well-maintained, that, that have, um, that are very, very healthy in the foundation. I think that's super important to us.

Right. Yeah, I think, like, you're, you're looking for something to have already found a niche and to be established because you don't really want to be pushing like a speculative architecture. You really wanna be embracing something that already works.

And so I think a lot of the stuff that we're talking about, like payments or like kind of a-- I, I, I really enjoy the, like, interface to model architectures. Uh, I think those are really interesting, but it's not yet obvious that that pattern needs to exist.

Uh, and so that's something where we can go see it and, like, try to make it work, um, in some projects and then come bring that back if it really has a, a role.

And on, on the opposite, what's my incentive to bring my project to you? I have a, a s- project with adoption. It's well-maintained, it's healthy. What's the benefit that I get from, um, donating to the foundation?

I mean, I can start that, but I'd, I'd love to hear from these guys as well. I, I think what you-- All technology is an implicit futures contract, right? And so, you know, if there's technology that has traction and, uh, that traction sort of wants to be built upon, having that technology at a neutral place like the Agentic AI Foundation, where the whole industry is making decisions about how to invest.

And, and when I mean-- when I say investment, I don't mean like becoming a member-

Yeah

... of the foundation because you don't need to become a member to participate on the technical side. It's decisions about, "Hey, I'm gonna, you know, assign 10 of my company's engineers to co-develop this with your organization, the contributing organization."

And, you know, the-- that's a, a way that we can all essentially co-develop together, and that will provide better support, more development ve-velocity, higher code quality because more people are participating in it. A-and that's a massive incentive if, you know, you want your technology to actually be used and adopted in industry and get more feedback and kind of a, a positive feedback loop of great project begets great products in the market.

That market feedback then, you know, allows companies to make money off of them. They then pay engineers to improve the project, better products, more profits, better project, and, you know, that's the incentive, uh, which is a pretty high one.

Could add a technical sort of spin, which is none of these things are built in a vacuum. Like, all these projects build on lessons and learnings or practical code from other projects, and that's a big opportunity. Like, any technical, technical contribution will bring its own unique value to the foundation.

At the same time, it then gets to learn the lessons that all the other participants in the foundation do. And, like, I found it really valuable over this past year working with David and others on the MCP committee about, like, it's actually that communication thing that makes our ideas more robust, makes the implementation better.

We can be sure it's secure and safe and actually works, and this requires communication, and that's sort of the foundation's the natural, like, town square for this in a way.

And so one last angle on this. I think if, if you're working on a standard or a protocol, this is such an obvious decision, right? Like, the value in the protocol is about how many people are adopting it, so being a part of this, like, gets you that reach.

But I will say as someone who's working on a client and not a protocol, I think there's value there too, right? Like, you-- we want this to be part of the foundation because we, like, develop these ideas together, to your point, and so it makes it better.

Like, we're, we're donating Goose because we think it's going to make it a higher quality tool.

I, I actually have a follow-up question on just the LF side. LF has many other funds and, and, and organizations, including Data, Data and AI F- uh, Foundation, as well as like dedicated, like PyTorch and all the other-

Yeah

... uh, ones.

Yeah.

Um, I guess why a new foundation?

Well, because everyone's special.

No, I, I think that the way we look at, uh, in this space, and, you know, I'll, I'll put aside the projects in semiconductor tech and operating system and stuff, but in AI, we think of it sort of like how the market has evolved.

You know, it started with tools like PyTorch and, you know, that, the transformer tech that is used to create LLMs. The, the Linux Foundation kinda took a pass on, uh, like frontier model world because in the open source space, having connection to an inter- the internet and some intelligence and a computer, that's sort of entry.

In the world of frontier LLMs, it's a computer connection to the internet, some intelligence, two billion dollars worth of GPUs and a ton of data. Harder for consortiums to do that kind of work, so pass. Uh, then you look at how, you know, reasoning models have come out, need to be, you know, so in, in inference world, things need to be scalable.

Okay, now you've got interesting technology, vLLM, Ray, things like that. They have to be deployed on something. Kubernetes is sort of that. Uh, these are all distinct components- Agents are a distinct enough set o-of technology that it merits its own community.

Ah, separate from data and AI.

Yeah, because like a PyTorch dev isn't really doing a ton of stuff in agent land.

Yeah.

Right? You know, somebody working on Dockling maybe is a little more adjacent, right? But not quite the same as somebody who's, uh, like working on transformer tech or VLLM. And so they are logical categories. I mean, sometimes stuff comes in over time and, you know, we sort things out later.

We had early on in the telecommunications sector, you know, a software-defined networking effort, a network function virtualization organizat- uh, orchestration effort-

Wow

... a whole bunch of stuff, uh, all separate entities. And, uh, they-- I was like, "Let's just bring all these things together because the technology's now mature. We're taking all this money in, but we don't really need the resources anymore because the market's already mature."

And so it took me a year to get all these companies to decide to bring all these things together and not pay all these separate fees and have all these separate orgs. I have a little folder in my inbox that says, you know, "Convincing people not to give me money."

But in this world, I think it's a different kind of audience. I think it's narrow enough. I think it's specific enough to agents where it merits its own entity.

I think as well, it, um, dovetails somewhat with the earlier thought, like with the taste-making aspect to this.

Yeah.

For these organizations to be effective, they really need a focus, like something that brings them together.

Mhm.

And ultimately, like you can imagine an alternative where we snowball and there's only the Linux Foundation. There's this uber do everything remotely connected to a computer, and that wouldn't be that effective. So there's a taste-making here as well, which is we want to be focused on agentic systems and how they connect together, hence Agentic AI Foundation.

But like everything's about growth and evolution, so there's a possibility that later down the line, we recognize some natural affinity.

Yeah.

There's something new, something old-

Yeah

... and then they can be brought together. But the focus helps at the beginning, certainly.

What's gonna be the actionable outcomes? So obviously you have the funds to direct. Uh, I know a lot of the Linux Foundation does events. There's also like eventually like certification, things like that.

Yeah.

What's the split of the foundation investments? Is a lot of it going back to, uh, different projects individually? Is it about the community building? And then from people that have not been involved from the outside, it's like this just seems like a nice blog post and a bunch of logos, but like in reality, how are things gonna be actioned?

Yeah, I mean, 50 companies coming in to fund a bunch of blog posts seems like overkill.

Right. Exactly.

Um, so I think there's a couple of things. One, the intellectual property assets now are owned by this entity. That entity is responsible for making sure that, you know, that IP is managed effectively, that licenses are complied with, that intellectual property problems are dealt with.

Some funding goes to that. There's a leadership function where, you know, to help bring consensus a-across the industry and within developer communities, you have to have a special kind of someone to do that, and I think they need to be technically knowledgeable but humble enough to know that they're-- that the community is the one who makes the technical decisions.

So kinda just, you know, sort of lead through influence to kinda help people organize things effectively. So you, you hire some people to do that. You hire people to do like developer outreach, community engagement, because you want more developers coming into the community, so funding to go to that.

And then there's a huge convening function. The Linux Foundation hosts 50,000 plus, uh, virtual meetings a year. So we have this like... I think we're probably like one of the largest users of Zoom. I know for sure we're the largest- ...

Slack user in the world. And so that convening function is, is critically important. Make it as seamless and easy as possible to convene. And then, yeah, we, we hold events because I think, you know, to your point, developer engagement, face-to-face, being the p- the town square where you physically get together means something.

So I think you guys have been to KubeCon. You know, we have on, uh, uh, easily 10,000 people that come to that conference twice a year. Uh, in Europe this summer, there were 13,000 folks there who, you know, come in and they exchange ideas.

The core maintainers get together and make real decisions. Uh, and then the last thing we spend resources on, and, and you can go even just check these out for some of our other projects, is we have a whole platform that enables, you know, maintainers to look at their community and, and understand, you know, like, what's our velocity?

How many developers are we adding? Like, what's the social media scuttlebutt around this project? Uh, what are leading indicators of adoption? How's our security, uh, doing? Like, you know, do we have good practices about application security? And those are all things that we, you know, invest in to help make these communities, you know, better commercially adopted so that we get that positive feedback loop of like adoption begets more investment in the form of developers providing input, and that, that virtuous cycle kicks off.

That's where the funding goes for these kind of things.

I put in that question into our doc because it says it's a directed fund, so, uh, my cheeky question was, well, what are you directing them to?

So, so yeah, I mean, that's-- it's kind of... So directed fund gets into like the, the-

They are directing it

... the nerdiness of this. Yeah, so but it's, uh, the, the reason we structure it that way is somebody has to own everything.

Yeah.

The Linux Foundation is actually the ownership vehicle. And remember, we separate Technical governance from the governance of actually how money gets spent, because we don't want this sort of pay-to-play aspect of technology that tends to screw everything up.

And so the directed fund is really like, you know, real stakeholders who really care about this tech, put money in, and use it in a way to help build the market and the community and all the things I just talked about, and just let developers do what they're super good at: get together, solve tough problems, be, you know, tastemakers.

That, that's something that we separate.

Yeah. I think there's a great essay by Rich Hickey, who created Clojure, about, um, open source is not about you. Uh, just because something in open source, I don't owe to respond to your issue and to progress. And I think some of the worry sometimes that people have about the group is like, well, you know, if not, you're a part of this thing, uh, am I supposed to also listen to your thing and implement the thing that you said?

So, uh, I, I think that's gonna be a super interesting thing in a technology that is so new. You know, I feel like everybody-- Because there's so much venture money in, like, early stage companies and, like, uh, obviously the foundation model labs have raised so much money that they need to be on top of it.

There's a lot more pressure, I think, from the community to try and be a part of it and, like, put their stake and be like, "Yeah, we contributed that," or whatnot. So I just think it's like a unique...

Compared to like the CNCF, for example, where the hyperscalers are kinda like around the clouds, and we all know what those workloads look like and, like, nobody's really trying to influence. There's not like a OpenAI preferred thing versus like a Anthropic preferred thing.

But it wasn't, it wasn't always so. So when we started CNCF, I got a call from, I think it was Urs Hölzle and Brian Stevens, who were over at Google. It was 2014, I wanna say.

Mm-hmm.

And, uh, you know, they're competing... Well, they had-- They weren't even competing. They weren't in the cloud business, and they wanted to be in that business. Amazon was, you know, hosting virtual machines on EC2, and they were the de facto leader.

They said, "We will give away Kubernetes," which was kind of the Borg, and they renamed it Kubernetes, uh, "to the Linux Foundation. And we, you know, because we've never run a virtual machine, we think containers are a better way to scale cloud applications.

We'll give you this tech and, and it'll be helpful to us if the entire industry adopts containers and Kubernetes as the way to build and deploy applications." So that was the strategy out of Google, uh, and they contributed some serious IP that we all know today is awesome.

Yeah.

But at the time, remember, Mesos was still a thing.

Mm-hmm.

Like, PaaS was still-

Right

... a thing, right? Like, you know, Heroku, Cloud Foundry-

Mm-hmm

... even OpenStack, you know, like virtual machines were still kind of a thing. So it wasn't clear what the abstraction layer for cloud computing was. But once the market started sort of piling on to Kubernetes, y-you know, like, oh, now Microsoft joined Cloud Native Computing Foundation.

They're investing in Kubernetes and creating Kubernetes services. Oh, wait, Amazon's now investing in this? Then the consensus was really coming and built up here. I think there's a somewhat similar situation here with the caveat of saying like 10 times faster.

Right.

Like, just day one, so much momentum around MCP, so much interest in this. And then also 10 years of CNCF to sort of teach the developer community and the vendor community how to do this well, where, you know, investment is not mutually exclusive to great technical outcomes, I think has been super positive.

So I think this thing's gonna move super fast.

Awesome. Uh, we don't wanna keep you guys too long. I'm sure you've been on a media tour this week. What's maybe from each of you, like, one thing you look forward in the new year from the foundation?

So I don't really know what it's gonna be look like, but I really look forward to, like, the next step. Like, as David mentioned, like it's been months of development and discussion or whatever to bring us to this, and there's this sense of, I guess, relief, achievement.

There's like, you know, you made a foundation. We're collaborating. We created this open space. It's great. But the what next, I'm super excited for the next technical contribution for the first AAIF event or night or conference or whatever form it ends up taking.

'Cause there's another world where like bodies and foundations are created and then, like, eventually they get forgotten. And this is not that. This is really a beginning, and so I wanna see it be healthy and grow, and I just don't know what comes next.

So I'm most excited to see that in the new year.

Yeah, I think I'm most excited, like i-if I really take a neutral look of like what just happened in the industry with creating this, it's like you have Google, Microsoft, Amazon, um, uh, Block, Bloomberg, Cloudflare, OpenAI, Anthropic, just a platinum member, create a foundation.

I think it's just like this is actually quite, quite cool and quite substantial. And now it's just like, now we are at this like starting point of like, what can we do with this? And, and to, to Nick's point, we don't-- I think there's a lot of like things we don't know yet and like things we need to figure out.

Like for Anthropic, this is the first big foundation, um, we're creating, and we have to learn a lot here. Um, but I think it's just a, such an interesting like starting point, and I'm just super excited for these like new, uh, when you, when you start something new of like what you can build with it, and it's, it's in a way of building something that I'm not familiar with, so I'm super excited to learn about this and seeing what we can do with, with this like, I feel like quite unique vehicle now and like really driving the agentic, like AI, um, open source community forward and focusing on what, where some of these companies who are very competitive with each other have common ground and where we can build things together that is just benefiting

and uplifting every user in the market and every developer in the market and every builder in the market significantly. That's what I'm really excited about, um, to see.

I definitely agree with both. I think there's a lot of like opportunity to figure out what the structure does. But let me give you something more specific that I think is like already in, in coming up, which is, uh, I, I wanna see how agents become asynchronous.

And I'm really tired of like reading through chat sessions, and I, and I want this to be a thing where I can go have like 20 agents working for me and actually see that come together. So I think that MCP is starting to like approach that answer, and then we wanna like figure out how to make those reference implementations and show people how they can actually get like another order of magnitude out of what AI can do for them.

You don't enjoy pressing yes every five-

The approve every, every three seconds? Ugh.

Bypass. Bypass.

Dangerously skip permissions.

Yeah, turn it all-

I, I'm with you on that one. I think what I look forward to is, you know, the success stories of, you know, the, the organization that's implemented agentic technology i- in that way and hearing how it really impacted their business.

I'm looking forward to stories about MCP startups that made a ton of money. I'm looking forward to stories, uh, like, like in CNCF this year, uh, CVS Pharmacy joined the Cloud Native Computing Foundation, right? Like a, you know, a, a pharmacy company that's really a user and adopter of technology, sort of, you know, the late majority.

I, I think we're gonna start seeing organizations really, really use this tech impactfully, provide feedback back to the community, and like it, it just... The potential of the technology, I don't need to tell this crowd how huge it is, but we'll start to see that truly manifest.

That, that is gonna be cool.

Well, thank you all so much for joining, and congrats on the launch.

Thank you.

Thank you.

Thanks for having us.