Embracing the MCP Suck: Taming the Wild West of AI Protocols

Hey guys, thanks for the throw. io, and well, we're talking about MCP, otherwise known as the Model Context Protocol, and there's been a lot of, well, debate about where this thing is headed in the future and how relevant it may or may not be. Joey, welcome to the show.

Thanks for having me. Yeah. I recently actually just went to MCP, was it MCP Dev Summit in New York, and I'm really excited to talk to you about my experiences there and what all the conversation was about.

Well, we were both there, so maybe that's a good thing we can talk about it at the same time and have the same references. " But what's your assessment on what's going on here? Yeah, I think because I definitely heard that in a lot of the conversation at the summit, and I think, one, the fact that it's even coming up in the summit is a signal in and of itself.

I don't think personally that MCP is going anywhere anytime in the future. I think it's more of a question of what's the right tool for the right job. You have skills, you have MCP servers.

You don't need to reach for both of them 100% of the time. It's what do you need at that right moment. To your point about that, how do I figure out what is the right moment for the right tool?

Because I think a lot of folks, maybe they just have a tendency to go with one or the other as a default option, but maybe we need to think that through a little bit. Yeah. Honestly, I enjoy using them together a lot, and I'll write skills that you can call and utilize MCP servers.

I think you were right when you said that grabbing data and getting information out of a system is probably going to be the dominant use case moving forward for MCP, but there's capabilities that you can't wrap in a tool completely, or in a skill, I'm sorry. There was a lot of previews of what's upcoming in the next release of MCP, and also a lot of discussion about security, but we'll come back to that in a bit. But is there anything in the next rev that you're looking forward to, or is there something on your wishlist that isn't on that to-do list just yet?

Honestly, I learned a lot at the conference, and I wasn't aware of MCP apps at all, and mostly that was my fault because I utilize the CLI more than anything. So that was like a new world completely opened up to me, and that was the number one thing that I'm really excited to start actually diving into and start using. Elicitation was brought up, sampling was brought up a lot during that as well.

Those were two things that were decently new to me. I knew about elicitation, but sampling was... It's crazy that MCPs can now use the actual LLM of the client, so those two specific things I'm super excited to actually jump in and try.

With great power comes great responsibility, and a lot of folks are kind of worried that we're maybe over our skis a little bit when it comes to MCP because we're providing access to all kinds of data without necessarily the right security protocols and compliance protocols in place. And at least my take on the conference was fixing that or addressing those issue was not on the table for 2026. So what's your assessment?

Yeah. Honestly, even at Spacelift, we have Spacelift Intent, and we struggle with this internally on how to give people access to do specific things within specific guardrails. And at Spacelift, what we do is we allow people to write policies based off of what their interactions with that MCP server are actually going to create at the end of the day.

So I think that there's definitely room in the protocol to allow those type of guardrails to be directly built into something like MCP. But right now, like you said, the community just isn't even there at the moment. I think part of the issue too is when we have guardrails or policies that we hand off to humans, there's a reasonable expectation that they might adhere to those.

But with AI agents, it feels like they are just programmed to be as aggressive as possible, so- Yeah ... unless you say, "Thou shalt," and slap its hands a couple of times, there's going to be trouble. Exactly.

Yep. And that's what we're doing at Spacelift too, is you definitely can trust the end user to make... It's to some extent, but you're going to always have malicious actors in some form or another, and with intent, us being able to monitor what the actual output of those calls to the MCP server are and apply policies against them, I think is something that the whole community could benefit from looking at.

As I'm speaking, and I'm saying this out loud, does it belong in the MCP protocol? I don't know that actual answer, because the implementation of each different thing that an MCP server could do could be vastly different from one to the next. One of the things, and this is more of a feeling, but I wonder if you shared it, but it seemed to me a lot of the folks who were driving MCP were kind of surprised about how rapidly it's been adopted and the use cases that are expanding.

It seems to have gone well beyond their initial vision for it. Yeah. I remember I was actually sitting in one of the talks, and I think it was one of the Anthropic guys that I was sitting in on, and he showed a graph of the adoption of MCP, and it was like nothing, nothing, nothing, and then out of nowhere, it was just this giant line to the ceiling.

As we look at the impact of this, and maybe at the risk of a gross generalization, but let's say that I have agents everywhere, and they're accessing MCP protocol to pull data from any number of applications. Eventually, do all our applications essentially turn into headless services that are basically going to be invoked by these AI agents? And do we need GUIs anymore, or how does this all play out?

Yeah. This is actually something that I've thought about a lot. A coworker, his name's Jake Nestler, him and I actually put a project together on an AI agent type internet where there is no GUI, and it's all just text-based that the AI agent could interact with.

I think we're still really early in the AI age. It's been around for what? We've been talking about AI now for like ten-ish years at this point, starting from...

I remember, I'm going to go off on a little bit of a tangent. Do you remember the Daydream Google? The very first implementation that was commercially available of AI in the commercial space.

Do you remember that? I vaguely remember that, but if you needed me to cite chapter and verse, I'm not your guy. Well, the reason that I'm saying it is I remember 10 years ago, we had Google, and Google was daydreaming, and there would be all these swirls and these pictures of AI generating images.

And then now here we are, 10 years later-ish, and AI is literally writing code for us that's pretty good code. It's not great. You're always going to get AI slop, but it's pretty good.

So I think we're still very new to AI to really get a grasp of where it's actually headed. It's literally creating new industries as we speak. So I don't have a great answer for your question.

My solid answer is that it's just so new. There's just so many things that could possibly change. Maybe the future is GUI, maybe it's not.

Maybe it's some new thing that we don't even know about right now. So is there anything... I know normally I ask people this question in the context of years, but given how fast MCP is moving, is there anything you know now that you wish you knew four months ago?

Yeah. So, just the features of MCP, there's a lot that the community hasn't even started implementing, and I did my talk at MCP Summit on the actual pieces of the protocol, like tool calls and sampling and all of those features that come out of the box. When I very first started putting the idea for that talk together, things like resource calls and the MCP protocol weren't implemented in something like Cloud Code or inside of OpenCode, or I don't even think it was inside of Antigravity at the time.

And three months later, everything seemed to be implemented, and I had to kind of rethink the entire talk in and of itself. So just how quickly that everything is being created and then needing to be implemented inside of the MCP leaves a lot to ask of the community, I guess you could say. In that, I think that there is no...

What's a good way to say this? I'm trying to think of a way to say this without being a jerk in this call, but it feels like the protocol moves faster than the clients that actually implement the protocol itself. No, I think you're spot on.

I've talked to some folks who are like MCP as they see it today barely resembles what they first saw, and they're wondering six months from now if it's going to resemble anything like it looks like today. Exactly. Yep.

So if we have those issues, then how do I manage that? Because a lot of IT organizations, they live for the word stability, right? When things are rapidly changing, it makes them uncomfortable.

So, how do you manage or deal with that when the thing is changing so rapidly, and is this going to wind up with Kubernetes kind of set up where I have so many versions of Kubernetes running, and they're all vaguely incompatible? Listen, I was never in the military, but I had a military friend one time tell me that you should always embrace the suck. And that just means like, hey, it's going to be hard.

It's not going to be easy. But if you embrace it and you go along with it, the great things will come out of it. So, like I said too, we're really early still in the whole AI era of the human race.

And things are going to change, and they're going to change rapidly, especially because AI is enabling us to change things super rapidly. Really embrace what's changing. People aren't doing these things to be negative.

They're not doing it because they want to make your lives harder. Everybody wants the ecosystem to grow. So to your point about that, do you think that maybe there are a lot more rogue MCP servers that have been deployed out there than the average CIO knows about, and maybe we need to have a day where we declare amnesty, and to your point, everybody embraces the suck?

Yeah. Yeah, there definitely is. I went to a security talk at the summit, and one thing that came up during that security talk was that folks were spinning up MCP servers or writing skills that they weren't really aligned with the C-suites wants out of the company, I guess you could say.

So I don't have a great answer, man. It's going to happen. There needs to definitely be a focus on security around MCP and AI in general.

It kind of feels like the Wild West. We're kind of shooting from the hip right now. And again, it just boils down to AI is just so new for everything that we don't have all this figured out right yet.

I think we're going to get there. We're going to figure it all out, but right now there's just so much that's happening. It's hard to keep up at the moment, so you have to brace the suck and move forward, and hopefully either you or me or somebody in those space is going to come up with this brilliant idea that enables us to have all of the things that everybody wants.

All right, so last question. Beyond just embracing the suck, what's your best advice to your fellow developers about how to get started with MCP? Best advice is to just do it.

The thing that you need to look out for is what kind of MCP tooling that you reach for. One thing that really tripped me up when I started actually looking at doing MCP stuff is in Go, for example, you have defaults for everything inside of Go. A string is just an empty string.

A number is just zero. And if you enable a feature on the MCP protocol in Go, but you don't actually implement it, so if you enable it but don't implement it, you're going to get weird things that you see inside of your MCP implementation. You might ask for a prompt to get created, and instead the MCP server sends an empty string.

So my biggest piece of advice is jump into it, start using it, but be careful and understand the nuances of what it is that you're actually implementing. All right, folks, you heard it here. There's no substitute for hands-on experience when it comes to MCP servers and the protocol because, well, it's all that brand new to begin with.

Hey, Joey, thanks for being on the show. Thanks for having me. All right, and back to you guys in the studio.