Show notes
We are happy to announce that there will be a dedicated MCP track at the 2025 AI Engineer World's Fair, taking place Jun 3rd to 5th in San Francisco, where the MCP core team and major contributors and builders will be meeting. Join us and apply to speak or sponsor!When we first wrote Why MCP Won, we had no idea how quickly it was about to win.In the past 4 weeks, OpenAI and now Google have now announced the MCP support, effectively confirming our prediction that MCP was the presumptive winner of the agent standard wars. MCP has now overtaken OpenAPI, the incumbent option and most direct alternative, in GitHub stars (3 months ahead of conservative trendline):We have explored the state of MCP at AIE (now the first ever >100k views workshop):And since then, we’ve added a 7th reason why MCP won - this team acts very quickly on feedback, with the 2025-03-26 spec update adding support for stateless/resumable/streamable HTTP transports, and comprehensive authz capabilities based on OAuth 2.1.This bodes very well for the future of the community and project. For protocol and history nerds, we also asked David and Justin to tell the origin story of MCP, which we leave to the reader to enjoy (you can also skim the transcripts, or, the changelogs of a certain favored IDE). It’s incredible the impact that individual engineers solving their own problems can have on an entire industry.Full video episodeLike and subscribe on YouTube!Show Links* David* Justin* MCP* Why MCP WonTimestamps* 00:00 Introduction and Guest Welcome* 00:37 What is MCP?* 02:00 The Origin Story of MCP* 05:18 Development Challenges and Solutions* 08:06 Technical Details and Inspirations* 29:45 MCP vs Open API* 32:48 Building MCP Servers* 40:39 Exploring Model Independence in LLMs* 41:36 Building Richer Systems with MCP* 43:13 Understanding Agents in MCP* 45:45 Nesting and Tool Confusion in MCP* 49:11 Client Control and Tool Invocation* 52:08 Authorization and Trust in MCP Servers* 01:01:34 Future Roadmap and Stateless Servers* 01:10:07 Open Source Governance and Community Involvement* 01:18:12 Wishlist and Closing RemarksTranscriptAlessio [00:00:02]: Hey, everyone. Welcome back to Latent Space. This is Alessio, partner and CTO at Decibel, and I'm joined by my co-host Swyx, founder of Small AI.swyx [00:00:10]: Hey, morning. And today we have a remote recording, I guess, with David and Justin from Anthropic over in London. Welcome. Hey, good You guys have created a storm of hype because of MCP, and I'm really glad to have you on. Thanks for making the time. What is MCP? Let's start with a crisp what definition from the horse's mouth, and then we'll go into the origin story. But let's start off right off the bat. What is MCP?Justin/David [00:00:43]: Yeah, sure. So Model Context Protocol, or MCP for short, is basically something we've designed to help AI applications extend themselves or integrate with an ecosystem of plugins, basically. The terminology is a bit different. We use this client-server terminology, and we can talk about why that is and where that came from. But at the end of the day, it really is that. It's like extending and enhancing the functionality of AI application.swyx [00:01:05]: David, would you add anything?Justin/David [00:01:07]: Yeah, I think that's actually a good description. I think there's like a lot of different ways for how people are trying to explain it. But at the core, I think what Justin said is like extending AI applications is really what this is about. And I think the interesting bit here that I want to highlight, it's AI applications and not models themselves that this is focused on. That's a common misconception that we can talk about a bit later. But yeah. Another version that we've used and gotten to like is like MCP is kind of like the USB-C port of AI applications and that it's meant to be this universal connector to a whole ecosystem of things.swyx [00:01:44]: Yeah. Specifically, an interesting feature is, like you said, the client and server. And it's a sort of two-way, right? Like in the same way that said a USB-C is two-way, which could be super interesting. Yeah, let's go into a little bit of the origin story. There's many people who've tried to make statistics. There's many people who've tried to build open source. I think there's an overall, also, my sense is that Anthropic is going hard after developers in the way that other labs are not. And so I'm also curious if there was any external influence or was it just you two guys just in a room somewhere riffing?Justin/David [00:02:18]: It is actually mostly like us two guys in a room riffing. So this is not part of a big strategy. You know, if you roll back time a little bit and go into like July 2024. I was like, started. I started at Anthropic like three months earlier or two months earlier. And I was mostly working on internal developer tooling, which is what I've been doing for like years and years before. And as part of that, I think there was an effort of like, how do I empower more like employees at Anthropic to use, you know, to integrate really deeply with the models we have? Because we've seen these, like, how good it is, how amazing it will become even in the future. And of course, you know, just dogfoot your own model as much as you can. And as part of that. From my development tooling background, I quickly got frustrated by the idea that, you know, on one hand side, I have Cloud Desktop, which is this amazing tool with artifacts, which I really enjoyed. But it was very limited to exactly that feature set. And it was there was no way to extend it. And on the other hand side, I like work in IDEs, which could greatly like act on like the file system and a bunch of other things. But then they don't have artifacts or something like that. And so what I constantly did was just copy. Things back and forth on between Cloud Desktop and the IDE, and that quickly got me, honestly, just very frustrated. And part of that frustration wasn't like, how do I go and fix this? What, what do we need? And back to like this development developer, like focus that I have, I really thought about like, well, I know how to build all these integrations, but what do I need to do to let these applications let me do this? And so it's very quickly that you see that this is clearly like an M times N problem. Like you have multiple like applications. And multiple integrations you want to build and like, what that is better there to fix this than using a protocol. And at the same time, I was actually working on an LSP related thing internally that didn't go anywhere. But you put these things together in someone's brain and let them wait for like a few weeks. And out of that comes like the idea of like, let's build some, some protocol. And so back to like this little room, like it was literally just me going to a room with Justin and go like, I think we should build something like this. Uh, this is a good idea. And Justin. Lucky for me, just really took an interest in the idea, um, and, and took it from there to like, to, to build something, together with me, that's really the inception story is like, it's us to, from then on, just going and building it over, over the course of like, like a month and a half of like building the protocol, building the first integration, like Justin did a lot of the, like the heavy lifting of the first integrations in cloud desktop. I did a lot of the first, um, proof of concept of how this can look like in an IDE. And if you, we could talk about like some of. All the tidbits you can find way before the inception of like before the official release, if you were looking at the right repositories at the right time, but there you go. That's like some of the, the rough story.Alessio [00:05:12]: Uh, what was the timeline when, I know November 25th was like the official announcement date. When did you guys start working on it?Justin/David [00:05:19]: Justin, when did we start working on that? I think it, I think it was around July. I think, yeah, I, as soon as David pitched this initial idea, I got excited pretty quickly and we started working on it, I think. I think almost immediately after that conversation and then, I don't know, it was a couple, maybe a few months of, uh, building the really unrewarding bits, if we're being honest, because for, for establishing something that's like this communication protocol has clients and servers and like SDKs everywhere, there's just like a lot of like laying the groundwork that you have to do. So it was a pretty, uh, that was a pretty slow couple of months. But then afterward, once you get some things talking over that wire, it really starts to get exciting and you can start building. All sorts of crazy things. And I think this really came to a head. And I don't remember exactly when it was, maybe like approximately a month before release, there was an internal hackathon where some folks really got excited about MCP and started building all sorts of crazy applications. I think the coolest one of which was like an MCP server that can control a 3d printer or something. And so like, suddenly people are feeling this power of like cloud connecting to the outside world in a really tangible way. And that, that really added some, uh, some juice to us and to the release.Alessio [00:06:32]: Yeah. And we'll go into the technical details, but I just want to wrap up here. You mentioned you could have seen some things coming if you were looking in the right places. We always want to know what are the places to get alpha, how, how, how to find MCP early.Justin/David [00:06:44]: I'm a big Zed user. I liked the Zed editor. The first MCP implementation on an IDE was in Zed. It was written by me and it was there like a month and a half before the official release. Just because we needed to do it in the open because it's an open source project. Um, and so it was, it was not, it was named slightly differently because we. We were not set on the name yet, but it was there.swyx [00:07:05]: I'm happy to go a little bit. Anthropic also had some preview of a model with Zed, right? Some kind of fast editing, uh, model. Um, uh, I, I'm con I confess, you know, I'm a cursor windsurf user. Haven't tried Zed. Uh, what's, what's your, you know, unrelated or, you know, unsolicited two second pitch for, for Zed. That's a good question.Justin/David [00:07:28]: I, it really depends what you value in editors. For me. I, I wouldn't even say I like, I love Zed more than others. I like them all like complimentary in, in a way or another, like I do use windsurf. I do use Zed. Um, but I think my, my main pitch for Zed is low latency, super smooth experience editor with a decent enough AI integration.swyx [00:07:51]: I mean, and maybe, you know, I think that's, that's all it is for a lot of people. Uh, I think a lot of people obviously very tied to the VS code paradigm and the extensions that come along with it. Okay. So I wanted to go back a little bit. You know, on, on, on some of the things that you mentioned, Justin, uh, which was building MCP on paper, you know, obviously we only see the end result. It just seems inspired by LSP. And I, I think both of you have acknowledged that. So how much is there to build? And when you say build, is it a lot of code or a lot of design? Cause I felt like it's a lot of design, right? Like you're picking JSON RPC, like how much did you base off of LSP and, and, you know, what, what, what was the sort of hard, hard parts?Justin/David [00:08:29]: Yeah, absolutely. I mean, uh, we, we definitely did take heavy inspiration from LSP. David had much more prior experience with it than I did working on developer tools. So, you know, I've mostly worked on products or, or sort of infrastructural things. LSP was new to me. But as a, as a, like, or from design principles, it really makes a ton of sense because it does solve this M times N problem that David referred to where, you know, in the world before LSP, you had all these different IDEs and editors, and then all these different languages that each wants to support or that their users want them to support. And then everyone's just building like one. And so, like, you use Vim and you might have really great support for, like, honestly, I don't know, C or something, and then, like, you switch over to JetBrains and you have the Java support, but then, like, you don't get to use the great JetBrains Java support in Vim and you don't get to use the great C support in JetBrains or something like that. So LSP largely, I think, solved this problem by creating this common language that they could all speak and that, you know, you can have some people focus on really robust language server implementations, and then the IDE developers can really focus on that side. And they both benefit. So that was, like, our key takeaway for MCP is, like, that same principle and that same problem in the space of AI applications and extensions to AI applications. But in terms of, like, concrete particulars, I mean, we did take JSON RPC and we took this idea of bidirectionality, but I think we quickly took it down a different route after that. I guess there is one other principle from LSP that we try to stick to today, which is, like, this focus on how features manifest. More than. The semantics of things, if that makes sense. David refers to it as being presentation focused, where, like, basically thinking and, like, offering different primitives, not because necessarily the semantics of them are very different, but because you want them to show up in the application differently. Like, that was a key sort of insight about how LSP was developed. And that's also something we try to apply to MCP. But like I said, then from there, like, yeah, we spent a lot of time, really a lot of time, and we could go into this more separately, like, thinking about each of the primitives that we want to offer in MCP. And why they should be different, like, why we want to have all these different concepts. That was a significant amount of work. That was the design work, as you allude to. But then also already out of the gate, we had three different languages that we wanted to at least support to some degree. That was TypeScript, Python, and then for the Z integration, it was Rust. So there was some SDK building work in those languages, a mixture of clients and servers to build out to try to create this, like, internal ecosystem that we could start playing with. And then, yeah, I guess just trying to make everything, like, robust over, like, I don't know, this whole, like, concept that we have for local MCP, where you, like, launch subprocesses and stuff and making that robust took some time as well. Yeah, maybe adding to that, I think the LSP inference goes even a little bit further. Like, we did take actually quite a look at criticisms on LSP, like, things that LSP didn't do right and things that people felt they would love to have different and really took that to heart to, like, see, you know, what are some of the things. that we wish, you know, we should do better. We took a, you know, like, a lengthy, like, look at, like, their very unique approach to JSON RPC, I may say, and then we decided that this is not what we do. And so there's, like, these differences, but it's clearly very, very inspired. Because I think when you're trying to build and focus, if you're trying to build something like MCP, you kind of want to pick the areas you want to innovate in, but you kind of want to be boring about the other parts in pattern matching LSP. So the problem allows you to be boring i…
Full show notes at the publisher 
Our TOPPODCAST Picks 
Stay Connected
