MCP, David Cramer from Sentry, A2A, Dave Yen from Orange Collective

So cringe.

Yeah, I've been full devrel, which is not my job to be clear. I run a company sort of or I help run a company, but it's been fun. Maybe just quick background, maybe that would be helpful. A lot of people probably know who you are but maybe not everybody. If you don't know me I'm the founder of Sentry. Hopefully you know what Sentry is. If not, you just go check yourself. It's not a good thing. But yeah, I run... I don't actually know what I do. I do a lot of random stuff but the last couple weeks I've been building the MCP server for us which is not usually what I do to be fair. I haven't written code for the company in 8 years until this. So, but it's been fun.

No, I think it's everybody wants to... it's like FOMO. Everybody wants to be like, pick me, pick me, pick me, I'm here too. They want to be relevant, you know? And I do think MCP is not a bad idea, mind you. It's a good architecture and it makes a lot of sense. There's still tons of unanswered problems, but it is literally... when I was writing this post about it the other day, I Googled MCP and I bet if you Google MCP right now, you're going to get a sponsored listing. I got one for Neo4j. It's like some old school graph database, but there's a sponsored listing for the term MCP. And I'm like, what is going on here? You can clearly see it's very much latching on to whatever's happening right now, which is to be fair, it's good from a go to market point of view, I guess. But a lot of the stuff doesn't make sense. Like, you don't need to stick all these random things in a chatbot. That's not how you consume it, you know?

So, yeah. So we were talking to some YC founders yesterday that they were MCP adjacent before MCP. So then unfortunately the hype kind of dissolves other protocols. Now they're trying to maybe get into the MCP space. They're talking to big companies. They talked to Asana. Let's just put... they talk to companies like that who have MCP servers with zero usage because for Asana or Linear, the UI has already been optimized for so many years for a user, so the MCP is not there. At least not yet. I think the promise is there, right? That's what everyone buys into this hype that this promise could exist but it doesn't exist not today.

And so like ours is, oh, Sentry, you probably want to do some stuff with the error data and all this other stuff that's in Sentry inside of something like Cursor. Totally makes sense. So, of course, we want to expose that there if we can find a way to do it. Well, I just think there's a... and that's fine. We at least have a thesis behind it and I think it's okay if you don't quite know how it's going to be used, but I just think too many people are like, we have no clues. We'll just here's our API. Good luck. Have fun. That's not good enough.

Exactly. And so like for example, we have our MCP doc server and we put all of our docs in there. We pre-process them in there and people can use that to essentially have an assistant while running with Cursor. Outside of that use case I've not used an MCP server for anything other than things like that that help me in IDEs, right? We've played around with them with Composio and all these different ones that help you. But yeah, it's just kind of a mess, I guess.

Yeah. I have no idea how this thing we're on works, but I can give it a shot.

Let's see. Oh, yeah. Yeah, I see it. Okay. Oh, this is going to be... Oh, actually, we can allow... I'm like, is this one of those things where I got to reboot the browser to give it permission?

Okay, I think we're good. All right. We're going to do the... Yeah, let's do it live. the recursive screen.

I guess I don't actually need to set this up. I just need something that's functioning. And fingers crossed Cursor does not break here. Let me swap this.

Yeah, I'm a Cursor guy. Strong believer it is the best in the industry right now. OB's a Windsurfer.

Oh yeah, I guess I don't need to bring up that screen to talk. I'm not a streamer. Sorry. But I do need this screen. Okay, you can share if you want. We're just having fun anyways.

So where do I go? Okay. So, we got this MCP server. If you don't know what it is, it's just a wrapper around your API. And so, it's a wrapper that takes something you're doing, could be a public API, could be a database, something. Converts it to, in our case, we're using markdown. Converts it to markdown and hopefully you can do something with that. So, a really good example. Let's see if I can avoid PII. This is always like the risk. I think we're good, though. If not, I'm going to ask for forgiveness. So, let's take... which one least likely has PII? Let's go with this one. If you want me to like... we're probably good. It's good. Somebody will be like, oh, an IP address or something, but it's actually not a big deal. Okay. So, I'm cheating a little bit. Don't worry about what's happening right now. This is one of our agents that you're seeing on the screen. This is our... it's going to be renamed, but this is our issue fix thing. What this is doing is root causing a problem. It may or may not do something useful here. I just need an example to show you why MCPs are interesting. Okay, so we want this... actually we can just take the URL. Let's get rid of all that crud. So let's say this is the MCP server set up to debug the MCP server just so we're clear on what's going on.

And we are in Cursor and we've got this configured here. If you haven't configured these, they're kind of... they're a little brittle. But when you got a green light, it's working. You won't always get a green light. It's complicated. All these work a little bit differently. It will probably get better in a few weeks across the industry, but we'll see. Okay. So, we did this thing. I'm trying to fix a bug I found in Sentry. Can you root cause it and help me find a solution?

Fingers crossed. Live demo. I don't actually know if this is going to do what I wanted to do because we're sort of banking on the agent prompt signaling correctly. Yeah, but this prompt you can... this thought that I just put into chat is actually how you would think about or how I think about it, right? And I'm pretty sure this is not whatsoever doing what I wanted it to do. So, let's try again. Use Sentry's issue fix behavior. This is fine. I guess we should not expect... This is not what I wanted to do either. I didn't want it to begin. So, here's the problem. MCPs are super limited right now. And it takes sometimes a couple minutes to do this thing in Sentry, this thing that our agent does. And so, I try to signal to the agent that, hey, don't kick off an issue fix right away. Check to see if one's in progress first. Yeah, because then I can cheat and I can start it in the background or something like that. But yes, pull until it is ready. Let's see what happens here. Usually it gives up after a few tries, so maybe it'll actually work this time.

Yeah, but it's exposed in this complicated abstraction. So, I'm calling an API on Sentry, which is running an agent behind the scenes. It's like I'll continue pulling, but it's actually not continuing pulling. It just thinks it is. And so try again and wait a few seconds in between. I don't think, for example, though, it doesn't know how to wait. So, you'd have to expose like a wait tool Cursor would. Otherwise, it can't actually do any of this, right? Okay, cool. It finished. So, oh, it's not done yet. It's not done, but it has partial information, which is what it's giving here. And so, it's like, okay, the root cause is this blah blah blah. I didn't actually look at this error, so I have no idea what it is. Unfortunately, it's like, oh, I need to keep pulling. Pull again. But you can understand why this is not great today, though, just as an example of a really good use case for this, and it's still bad.

Yeah. People are too wrapped up in chat bots and stuff thinking that's the solution. Yeah, I mean I think... I mean I'm not the one who came up with this clearly but I've heard chat bots are like the CLI of AI. So it's fine, but in most cases it doesn't seem like it's the best use for... Yeah, because I mean like you look at this autofix thing in Sentry. This is an agent. It's not a chatbot. Like there's some text inputs in here, but like down here we've got... where is it? Hang on. I need to move this little screen. I think it's in here somewhere. So diagnosing this is... this is wrong. This is where LLMs mess up, by the way. Is like fail to parse token response does not mean the user is missing. I don't know where it came up with this idea. Usually it's pretty good, but sometimes there's just errors that make no sense. But yeah, so we've got an actual UI that tries to solve. So code it up will just give me a solution. Again, you got to think about how to use these user interfaces, but the concept of I want this context inside of Cursor is the important one. How we get it there, I don't know. The only way we're getting anything anywhere right now is MCP. That's exciting for folks.

Yeah. Yeah. And you'll see we actually came up with... I mean it's the same thing. And so but the idea is like I can create a PR out of this, right? But it's probably wrong because it's an LLM, right? I'm going to have to change it. But why would I create a PR? And I actually don't even know what checkout locally does. But really I just want it in my editor right now. I want to be like, give me the potential changes you're suggesting so I can review and tweak it, etc., etc., etc. Because it is totally wrong right now. But, just contrived example. Anyway, so I think that's an interesting concept. Obviously, there's other things you could do, but it's silly because like here's an example of how every single MCP works right now. And this is not real value add. It's like, show me my most recent issues in Sentry remote MCP. It's like why is this useful to me? Like I got a chatbot to do this when I can go click a UI that's way more consumable. It's not useful and this is every MCP that's been exposed recently, right? And so that's the problem. I don't know. Again, there's lots of useful use cases if you think of as tools and reusable tools like maybe somebody is... so here's a real good example. If you're like a Devin or V0 or pick one of these code generators and maybe you had a mode that's like I can deploy this to production and then you could monitor Sentry, look for errors, use the MCP to call our agent or something. So you could plug in these tools from Sentry to augment a bigger workflow, but you still have to design a system. You don't just plug in abstract tools and pretend they're useful, you know?

Yeah. I mean, yeah, the context... doesn't matter how big the context is, it's usually still limited, right, on what it can understand. So, yeah, 100%. And sure, maybe you could feed it all the context in the world, but are you going to pay $10,000 for this bug, whatever this is? It's probably not even a real bug. No, you're not. It doesn't make any sense. So, I could totally see though as you have agents that can, not just run when you tell it to on your input, but imagine you have alerts that come into Sentry and you can wire up agents to do the first triage of that and maybe propose a fix and things like that could become really useful. It's almost like they were a Sentry, you know what I mean?

So, this is what I think is wacky. I heard you all talking about agent to agent and stuff I think right before joining but like agents are a good idea like that's a good construct tools are a good idea MCP is a great way to make tools sort of composable reusable right but agents are really the value prop and everybody's over rotating on MCP because it's like oh we can integrate our just like here's our API we're in the conversation now but you're not and there's a lot of use cases for agents and I think MCP is going to be actually way more valuable in the sort of people are latching on to right now when we can expose agents through it. So like you saw what I was doing here and it was bad. It's like I had to tell it to check again but really it just communicated with an agent. It should be able to stream that back like any other thing that's happening in here. Just the protocol doesn't support that right now. And so I think once that happens you'll effectively unlock agent to agent via MCP which just means it's super pluggable and it's still high value at that point. And I actually think that opens the door to a lot more free form uses because I can just expose a Sentry agent that on our servers I can parse your query that so like effectively agent to agent in Cursor sends me and then I can do something much more useful with it than these abstract tools that we're just exposing to you.

Okay, so here's the thing like we are deep into A2A right now more so because we have to be and it's the same kind of growing pains are going to happen just as MCP is going through but what's your take so far or if at all I don't know if A2A is like... is this like you can't make fetch a thing or whatever. When I looked at when I saw the announcement, which was conveniently right after MCP was starting to get popular, and I looked at all the logos in the... like look at all these enterprise people supporting A2A, it's like a bunch of irrelevant companies. Like these are not the companies that are leading the way in anything. These are the companies trying to latch on to things that are pure enterprise, have not been super relevant for many years. And I'm like, that's signal to me. And I look for any signal I can get in anything. And so right now, I don't know, we'll see what happens. But I don't have any confidence in that.

Yeah, it is. One is also much more complex to implement by the way. Like this was my point because MCP is also, you know, I like the guys building it. I'm a fan, but it's also needlessly complicated. Like I cannot... like the way it works is I send in a request. It goes through one route and then the response comes through another system at the end of the... It's like a multi-channel thing. So it is like fire task over here and then receive a response over here. And that is so complicated to reason about when we have streaming, we have web... like we have HTTP, we have all these protocols that exist. And my biggest complaint is I'm like you really don't need to reinvent the core protocol for the 99% use case to support some 1% outlier over here. And yeah, sure, maybe not all... maybe a lot of stuff does have to fire off and then check back later. That could still be a stream. So there's some things there. But I just think it's needlessly complicated and I wish it was not. And this is why, for example, out of the box, it's not that easy to just support streaming responses for agent to agent in MCP. And I wish they would do something about it and just dial it back. But I know there's already been discussions and they're like, no, there's use cases for this. And this is how all these design by committee things go, these big academic concepts. And you know, it's not that it's invalid, it's just it's complex, and not everybody needs that complexity. And so hopefully the abstractions make it easier. But still, I don't know. We'll see what happens.

I don't know. Here's my kind of point of view on everything. First off, anything good usually comes from one or two people like their opinion. It's not designed by committee, right? It's like no, this is our view of the world. That's okay if you don't agree with it. And the best things come with our view of the world and that happens to be a really good view of the world that changes a lot of things, right? And so I generally am anti any version of design by committee. So that's one. I think two, one thing we're seeing right now already, LLMs are not intelligence. It's just pattern matching. You know, sure it's complicated pattern matching, but that's all it is. It should not be confused with any version of intelligence. And an analogy you can apply from humans to LLM is if it's hard for a human to reason about, it's also hard for this thing that's unintelligent to reason about because we pattern match, it pattern matches, the delta is we have a ton of context. And sure, in the vast infinity of compute, it can have more context, but that's just not the reality we live in. And so when you look at things, even something modern right now like Next.js, it can be quite hard to reason about what's going on in there. TypeScript errors quite hard to reason about what's going on in there to the point where I can't solve them. You think this machine is going to solve it? Like I'll spend hours on the same problem. And it's like we shouldn't be surprised that it's hard for the machines to solve some of these problems. I guess is my point. And so I do think if we want this technology to succeed, it might push the conversation of simplicity again, which doesn't mean under the hood things have to be simple, but the exposed experiences to say the framework or the programming languages have to actually be pretty simple. I don't know what that looks like honestly, but I'm trying to rewrite... I have a side product right now. I'm trying to rewrite it to be OpenAI API native. Which then will hopefully get me to OpenAI native, which I'll talk about. But I'm trying to rewrite it. I'm trying to focus on DX, like the ergonomics of code, and I'm trying to just remove every piece of boilerplate. The less code you have to write, the simpler it is, the less variable it is. Thus, the more likely I can code gen everything I want and not think too hard about things. And so, I'm spending a lot of time like how do I get this to the bare minimum of simplistic interfaces for the 99%. Again, you can have an outlier case where complexity enters the mix and maybe the human gets involved, but most of the time I'm like, copy paste this thing, change this behavior, add a test that's copy pasting this other test. It's like that is most of the crud we do, right? And so this is kind of my litmus test of like, well, if I spend a lot of time to intentionally make this really, really, really simple like that, you know, a brand new contributor could figure out what's going on just by looking at it. Pretty sure that any model will be able to code gen really effectively at it. And again, this has nothing to do with LLMs. This is how we should design things in the first place. We just don't, right? And so I don't know. It's like hopefully this is a forcing function again because the efficiency comes from simplicity. And I don't know. We'll see. We'll see.

And I also think like V0 is pretty effective for Next, but they spend a lot of time building... basically the strategy is summarized context over and over with RAGs and so they spend a lot of time building these tools and these contexts that allow it to be more effective at Next. Yeah. But these are generalized models. And so if you want it to be genuinely effective at technology, things need to kind of look similar. They need to look simple. They need to not be innovative. And to be fair, technology doesn't change. Like MCP is not a new idea. It's an RPC protocol. It doesn't have to be as complex as it is right? So it didn't have to reinvent a new thing. The exposed interface looks like RPC for the most part so it's easy to reason about fortunately. But yeah so I really hope we see more of this simplicity because if you can design simple tech hopefully you'll see literally financial advantages to doing that thus it'll motivate the right people to make those decisions.