Twilio: Demystifying Model Context Protocol (MCP) And Real-World AI Deployment

[00:00:00] This podcast only exists because of partners like NordLayer. I'm incredibly grateful for the support that helps me deliver these conversations every single day. And one of the things I've learned in those many conversations is that hybrid work has changed everything about how teams operate. People are now working from different locations on different devices and often outside traditional networks. But the one constant here, no matter where people are working, is the browser.

[00:00:27] This is where collaboration happens. That's where the data moves and increasingly that's where the risk lives. Now, NordLayer's business browser is designed for this new reality. It allows teams to work securely from anywhere, even on personal devices, all while giving organizations better control over access, data movement and browser activity. And it even supports modern ways of working without forcing old security models onto them.

[00:00:57] And it is that balance between flexibility and control that is where a lot of companies are still struggling. So if anything that I mentioned there resonates with you, if it sounds familiar, please, I encourage you to take a closer look at NordLayer.com slash browser. But now, on with today's show.

[00:01:20] What if the biggest thing holding AI back is not the model itself, but the way that it gets access to the right information at the right moment. Because when dealing with AI, I think one thing that has been missed or neglected over the last few years is context. And today I want to discuss a topic that is suddenly everywhere in AI circles.

[00:01:44] Yet, to outsiders, it can feel confusing, especially if you're a business leader, marketer or anyone just trying to separate the hype from the reality. And what I want to talk about today is Model Context Protocol or MCP, as you've probably seen it mentioned in numerous articles online. And yet, tech does love acronyms. And to many of you listening, it might sound a little dry.

[00:02:12] But the idea behind it could have a very real impact on how AI systems move from clever demos to tools that are actually useful in the real world and solving very real world problems. Because here's the issue. Large language models, yep, they can sound smart, but without access to live business data, systems and tools, they are often working with an incomplete picture. They're lacking context.

[00:02:41] And they may know the theory, but they don't always know what is happening right now inside your calendar, CRM, your customer support platform or your communication stack. And here's the thing. When context is missing, the answers quickly fall apart. And this is why the conversation matters and probably why MCP is quickly emerging as a way to standardize how AI models connect with external tools and data sources,

[00:03:11] ultimately making it easier for them to pull the context that they need and take action in a more reliable way. Now, for some, this means fewer custom integrations. For others, it could mean faster experimentation, simpler maintenance and a much clearer path from pilot project to real tangible business value. So to help make sense of all this, though, I'm going to be joined by Peter Bell, VP of EMEA Marketing at Twilio.

[00:03:39] And Peter does a brilliant job of stripping away all the jargon and getting to the very real question that every organization should be asking before chasing any of the latest trends. And that is simply asking, what problem are we actually trying to solve here? So whether you have been hearing the term MCP everywhere lately or this is the first time it's crossed your radar, I'm hoping that today's conversation will finally make it all click.

[00:04:06] But enough for me. Let's get Peter back onto the podcast now. Welcome back to the show. Last time we spoke, we promised you would come back. You've stayed true to your word. But for anyone that missed that one, can you remind everyone listening with a little about who you are and what you do? Yes, of course. So my name is Peter Bell. I run marketing in EMEA for Twilio. If you're not familiar with Twilio, we are a cloud communications company. What does that mean in practical terms?

[00:04:37] It means we provide predominantly the developer community with programmatic access to voice messaging, email, video for that matter. So if you want to initiate a voice call, send a text to confirm an appointment as part of your application, then we provide that global infrastructure. Well, it's a pleasure to have you back with me today. For anyone that missed our last conversation, we did talk briefly about model context protocol or MCP.

[00:05:06] And we said at the time it's almost an episode on its own. So for everyone listening, that's what we're going to do today. And for anyone that's heard the term model context protocol or MCP for the first time, in simple terms, what is MCP? And what problems was it designed to solve in this world of AI that we find ourselves? So as a marketer, I will say it could have a better name. MCP or model context protocol doesn't leap off the page unless you're a software developer.

[00:05:36] But in itself, it solves a very simple problem. I'm just going to start with a straight consumer example. Yeah. You could go to your favorite unique, you could call, chat GPT, take your pick of conversational agents. And you could ask it, hey, what's the journey time between London and Manchester? And it will come back with probably a pretty reasonable answer. But what it's not going to have are the current traffic conditions.

[00:06:05] It's not going to know that the M40 is closed and will be closed for the next six hours. So you'll set off hoping that it will take about four hours. And it turns out it's a bit more of a mission because it didn't have the context of road conditions right now. And that's what we mean by context. You could say it data. It didn't have the data available to it to know that.

[00:06:26] But context is a better representation in terms of, in this case, it's dynamic data, which if you had access to when doing that, you'd get a far better answer. And you'd know to either delay your journey or reroute or indeed delay the journey entirely. It's such a great example. And I think it really hits, hammers home the point there of exactly what it is.

[00:06:49] And I think one of the biggest challenges with AI systems today is that data lives across dozens of tools from CRMs to databases to internal documents. And that list goes on and on. So how does MCP help an AI model access and work with information across all those different systems? Yeah, and it comes back to a problem that, frankly, is as old as software itself.

[00:07:13] As you said, there's always been from the get go, you'll have some data in one system and some data in another system. And, you know, we talked for decades about data silos and how to overcome that. And the standard kind of enterprise answer to this in the past has been the enterprise message bus or something along those lines, which is you provide a layer of abstraction.

[00:07:35] So rather than integrating every system with every other system, so you gradually gain it gets gradually more and more complex to both build and even worse maintain. Because if you've got five systems, they're all and you've directly integrated them with each other, then you've got five times five integration points.

[00:07:59] So you've got 25 integration points to maintain as opposed to in the enterprise where you typically use some sort of enterprise message bus where you each system is integrated with a message bus once. And therefore, maintenance and things become much easier. Now, so you can integrate an LLM via API. Just you can take a REST API. You could look up the documentation around the API and you could hook up the you could.

[00:08:28] And to stay with my previous example, you could hook up the LLM to Google Maps. Yeah. And that's actually would do the same job as were described without the need for an MCP server. The problem becomes more complex, though, when you look at a lot of the use cases for these conversational queries. I don't just want to know how long it's going to take to Manchester. I want to know if I can get public transport. I want to know what the weather is going to be like so I can plan for my journey.

[00:08:58] And suddenly the question you're asking is more complex and it demands more data from multiple from ever more systems. And you get into this integration complexity where, one, how do you anticipate all the systems you need to integrate with? And secondly, just the complexity of doing that work, because although REST APIs are standard in themselves, they all operate differently and the developer has to look up the specific switch API.

[00:09:26] And what happens if part of my question isn't covered by the existing integrations? Let's say I want to know what the weather is going to be like. It would affect road conditions and it would also affect, might affect public transport if I chose to fly. Or when I get there, maybe I pack the wrong things to wear. One of the great things about MCP servers are they do dynamic discovery. So it can ask MCP. MCP, I'll take a diversion.

[00:09:55] MCP is split into clients and server. And one of the great things about MCP, which an API does not do, is allows dynamic discovery. You can literally ask an MCP client, what do you do? What functions do you have? What gator do you have? And at runtime, rather than development time, hey, you've got the answer to a question I need to answer. Right?

[00:10:20] Because, of course, the random questions and queries we type in are dynamic, ever-changing, more demanding. Whereas in the enterprise, if I go back to our more kind of rigid enterprise, you'd have a systems analyst work with end users. And they would describe what they need to do for their line of it to do their job. The analyst would write that down. They'd go away. They'd go to development.

[00:10:44] They'd work out which systems have that data and then build this pretty rigid solution, which would serve that person's needs. Particularly as consumers, we don't respect those boundaries or that rigidity. We just want to know how long is it going to take to get to Manchester today and will it be raining when I get there or will it be sunny?

[00:11:01] So that dynamic runtime lookup and integration plus standardizing the way the LLM interacts with all of the clients or the systems which have the contextual data or functions that I need to look up things. It just makes it so much simpler and much more maintainable. You can integrate any piece of software with any piece of software, but if it's all proprietary and it's all system to system, then, as I said at the beginning, the complexity goes up.

[00:11:31] And then maintaining it just becomes a real headache. And I suspect we will have a lot of people listening who've been in organizations that have tried to build AI tools, but have they quickly run into problem of connecting those models to real business systems and the maintenance of that as well. So why has that integration challenge been such a barrier? And how do you see MCP changing that equation?

[00:11:57] I think MCP has answered this problem first and foremost. Of course, it's an open protocol. So in itself, the software has to go implement it and build it.

[00:12:09] So you've got companies like Zapier who've gone out there and seen an opportunity, along with numerous other third parties, who are building the MCP servers and then building out MCP clients for your standard business software or, for that matter, consumer-facing software like Google Maps where there's an API.

[00:12:30] And in effect, MCP is providing that layer of abstraction between the systems with the data and the LLM. And it just makes it much, as I say, simpler to build, simpler to maintain, especially. And with the dynamic lookup, you don't have to imagine every scenario and every question which might get asked.

[00:12:54] And it's just a very timely answer to a very pertinent problem because LLMs only know what they've been trained on without that integration to say, I'll stay with mapping with Google Maps. Without that integration into real-time data source of traffic conditions, the answer is incomplete. So the end-user experience just isn't that great.

[00:13:22] And, you know, it's one thing to tell me theoretically how long it takes to get to Manchester. It's a lot more useful to tell me it's going to take a long time today. And I think traditionally businesses and people listening would have relied on application programming interfaces or APIs. And many people compare MCP to the role of APIs in connecting software services. And it's easy to make that comparison and see why people would.

[00:13:49] So how do you see MCPs, though, reshaping the way AI agents interact with business tools and platforms? Because that feels like the next evolution right there, doesn't it? It does. And there are instances where there will be a good case to do native integration. That might be more robust. There may be more functionality available. You know, there was an answer, I think it was last week, Lovable.

[00:14:16] You can now directly call the Twilio APIs through, we've got about 1,500 APIs. You can call those directly from within Lovable. Yeah. That's a direct integration to not using MCP. It could have used MCP. And there's probably a good, you know, there is a, you know, there'll be a software team somewhere that needs to decide, actually, do we need to build that native integration?

[00:14:42] And are we prepared to maintain and pay for the maintenance of that native integration over time? And in that case, the answer is probably yes between Lovable and the Twilio APIs. But there could be other examples where you don't yet know you want to integrate with those systems. It might only get used occasionally. And the MCP wrapper, because the MCP conceptually, you could think of as a wrapper around the API.

[00:15:10] It hides, because the API is specific to the underlying system itself. I've used Google Maps, but it could equally be Gmail. It could be Salesforce. It could be SAP. There are MCP clients for all of these. And it may be, one, I just don't want to maintain this ever-extending list of software that needs to be available to the LLM. Or it could just be occasional. You know, it just doesn't crop up that much.

[00:15:40] It's perhaps not so mission critical. So there's a design decision to be made there. I think they're complementary. They will have different strengths. You know, the API is going to absolutely expose the fastest native integration you're going to get. But then again, you have to do that when you're building your software as opposed to runtime integration, which you can do with an MCP server. Hey, what services do you offer? Oh, you do that.

[00:16:10] I need that right now for the query I'm trying to answer. And just bringing back what we're talking here, if we bring all that back to Twilio for a moment, before you join me today, I was reading how you've introduced your own MCP server to essentially make it easier for AI agents to interact with Twilio services. So just to bring that to life, can you walk me through maybe a practical example of what that enables? Perhaps something simple like, I don't know, an AI agent sending an SMS or provisioning a phone number or anything at all.

[00:16:39] Can you bring that to life with an example maybe? Yeah, it could be, I think, very simple examples. So if I take a step back, let's say you're – I'm going to stay with travel today because we'll stick to one example and we've got some continuity. So let's say you're building a travel app and part of that travel app is – it'll enable you to book flights for wherever you're going, perhaps further afield than Manchester to London.

[00:17:07] So you've – and as part of that, you want – as part of that experience, you want to have confirmation texts sent through and that's where you'd use the MCP server. Maybe it's not – you know, you're not building – it's not a core part of the functionality, but maybe you just don't want to deal with the Twilio APIs, even though they are excellent. You just decide, look, we want to – we're just going to use MCP for what we're building. It could be as simple as text confirmations coming through.

[00:17:36] I mean, that's the service part of the base, a very basic service we provide. But it's probably not at the time of booking because you can see in your app. But what you might want back are delays or cancellations to your travel or interruptions. It could be very occasional people using your app want to speak to the call center of the airline. And the easiest way to do that is to invoke that through MCP.

[00:18:02] So it's – it will be a design decision as to whether you want that native integration through the API or whether you're going to use it through an LLM. I'd say the difference is the way you've got a conversational client and fundamentally it's prompt-based.

[00:18:26] Then I think that's where MCP comes into its own because I know on the basis of the prompt that goes in, I know I've got what I've been trained on as the LLM and then I've got the contextual data and I can ask those systems if they can help augment this query, which you can't do with an API so dynamically.

[00:18:51] With an API, you've got to know, hey, as part of my application design, I know I want to send confirmation or travel disruption notifications through. You know, so you'd probably do native integration there. But if it's a different scenario and it's more casual, maybe it's occasional is a better word than casual. Maybe it's an occasional need. Then I don't want to be bothered with the maintenance of it. I'll just use the MCP server.

[00:19:21] But it's a good debate. Where do you use which? It's a great debate because they serve a very similar function. They make two pieces of software talk to each other. But I think particularly for the LLM, standardising how they interact with a potentially endless list of MCP clients,

[00:19:42] as we referred to, justifies itself as an approach because of the lifetime maintenance of maintaining that. It really is a great debate. And I would encourage anyone listening to get in touch, share your side of the story, which use cases work for which in your organisation. I'd love to get a bit of feedback from anybody listening.

[00:20:05] And anyone that spends any amount of time online will know that there is clearly a lot of excitement around MCP right now. But we are still very early in that journey. So just to level our expectations somewhat, what are the limitations or challenges that still need to be solved before MCP becomes as widely adopted across the enterprise? I think, first of all, there's just that developer experience. You know, people building things.

[00:20:33] I mean, there's a good range of MCP servers available and MCP clients at this point. There's quite a lot of people chase this commercially. And like always, though, it's like, what's the problem you're solving? I would always start with what problem am I solving? I've built a conversational app which enables people to manage their journeys, manage and plan their journeys.

[00:20:56] And obviously, that is only as good as the LLM will have a good amount of information from its broad training. But it lacks that specific context, as I've talked about. But it may also want to interact with my calendar. You know, when can I make this trip based on my schedule? Maybe I want to, I need to email in. You know, maybe it's to rent a car. Maybe that has to be done via email.

[00:21:25] Unlikely, bad example. But maybe there's a reason to actually have to contact someone. You know, if I want, if I believe in kind of the vision of agentic agents doing work on my behalf, then you need MCP because you rapidly need to talk to a growing list of third parties for that contextual data. Or indeed, specific functions like what's the weather going to be like?

[00:21:51] Or there might be some level of calculation because MCP service clients can also perform certain types of calculation. And if I was to broaden this out, if I'm planning a journey, there's actually probably more goes into it than we've talked about today. Okay. First off, I want to travel from here to Manchester. This is my budget. I need to travel within that budget.

[00:22:21] I needed to work with my calendar so it doesn't just, you know, so that I've, I want to spend three days there, find a slot on my calendar. I'm going to need a car when I get there. So if I do, if I don't choose to drive, then what are my car rental options? These are all things when we're planning a journey, we, we sort of go through in our heads and we perform those tasks. If you want your agenda gauge into that, you've got to tell it.

[00:22:46] And in telling it that you've actually introduced quite a lot of complexity and the need for a lot of contextual information. What's the pricing of the train? Would it be cheaper to fly? But what, you know, what about car parking costs? If I fly, I've got a park there. I've got to get a taxi. Does it actually save any time? And all these kind of, we actually put in this very simple example, there are a lot of kind of conditions we put in when thinking about, well, am I going to drive? Am I going to fly?

[00:23:16] Am I going to take the train? Am I going to take the bus? And MCP really takes the sting out of building a solution for that. And as more vendors and platforms, as they all begin building MCP servers, it almost feels like an ecosystem that's forming around a new standard at the moment. So what does that growing ecosystem mean for developers and businesses that are all building AI-powered services? Where's all this taking us?

[00:23:43] So like all networks, the value is exponential in terms of the number of endpoints on the network. The value is exponential. And I think we're in that phase right now of the vast majority. I had a quick look before talking to you today about, you know, specifically line of business software. I couldn't find anything where there wasn't an MCP server.

[00:24:05] I haven't gone out and QA'd them and tried them out, but, you know, for your typical enterprise scenario, it looks like there's an MCP server for all of your existing, pre-existing line of business systems and software. You may have some internal proprietary software you've built that you could run up an MCP client. And then you can, frankly, you've unlocked a lot of value because you've broken down your data silos in principle.

[00:24:36] Now you've got things like, should I really give everyone access to procurement system unfettered? Should I really, should everyone be able to see the full company accounts and things? So actually the MCP solves the integration problem. What it doesn't solve is the oversight, the observability, the security controls, those checks and balances.

[00:24:59] I think they're a bigger challenge because there will be, there are lots of things I may like to query as an employee, but maybe I shouldn't be able to for statutory and legal reasons. And for any business leaders that are listening to our conversation today who are not developers, but just want to understand the bigger picture. Like you said a few moments ago, what problem are we solving here? Why does MCP matter?

[00:25:24] What opportunities do you see it unlocking for organizations that are trying to turn AI from just a demo or a pilot into something that delivers real operational value? Because it feels like there's a real appetite for this right now. There is. And I think first off, I would start with a defined and small scope. And this is not specific to MCP. This is specific to AI. The risk is you try, you know, take my simple example today.

[00:25:51] My first telling of my journey from to London to Manchester was very, very simple. I just want Google Maps to basically tell me if there are any traffic delays using their real time information. In my latter example, I threw in a whole lot of other things about schedule, cost, parking, train, you know, a very multimodal scenario. I would absolutely start with the former. Just do something quite simple.

[00:26:22] Prove it out. Prove it works. Understand the issues it uncovers. Those issues could be around security and observability. And then start to then start to think about the design. We have a almost successful internal AI implementation is actually email based. It's not a chat bot because when you say AI, everyone thinks chat bot.

[00:26:46] And what it's doing is if you submit a form on Twilio dot com, it will be an agentic agent that emails you back. We're transparent about this. We're not hiding it. And we've gradually won. It now answers about 95% of those form fills. We didn't start at 95%, to be clear. We started at about 10%, if not lower. It has boundaries on the type of questions it can answer.

[00:27:15] It's about eight different buckets of questions it can answer. That removes the risk of hallucination because it is only working on specific data it's been trained on internally. Otherwise, it escalates to a human. And we can always go back and audit as to why, what it did and why it made that decision. Now, we didn't set out knowing those things. We learned them through starting small with caution, accepting a degree of risk.

[00:27:45] There was a risk that someone could get a reply that they then forward to our CEO because they weren't happy with it. We accepted the risk. And then we, as crawl, walk, run, for want of a better phrase, we crawled, we walked, and then probably now towards the running phase. But we had a very defined scope and a very specific business problem.

[00:28:09] We did not have enough customer service agents and couldn't afford to hire enough to answer all of these form fields that were coming through our website. So we were solving a real business problem. We weren't inventing one. We weren't fitting one to make the technology look good. We took a business problem that was detrimental to our reputation because we weren't getting back to people.

[00:28:33] And that email agent, which amusingly is called Isabel, no relation to me, and I had nothing to do with the naming, has an NPS score of 9 out of 10. Wow. Works seven days a week, speaks 12 different languages across 86 different countries. And it's a business success story for us. And I would frame it as that rather than technology success. Yes, it's an agentic AI solution.

[00:29:01] And MCP would have probably accelerated, in fact, would have accelerated our internal deployment because we have done native integration to the systems that the agent is talking to. So MCP, had it been around at the time, would have accelerated that because we could have simply built MCP clients for all of the internal systems. MCP is everywhere right now.

[00:29:26] And I think you've done an absolutely phenomenal job of demystifying it, put in a language everyone can understand. And it's so refreshing to hear you coming at it from a what problem are we solving mindset rather than where can we fit this technology into. And I thought of you recently when I was at the Qualtrics X4 event in Seattle, because the big message there was much the same as what you're saying today. Context is so important when using AI.

[00:29:52] It's something that's been neglected over the last few years, but there's a big push towards getting that context right. And MCP plays a huge part in that. But he was almost joking and frustrated that online there are already articles over the last few weeks declaring that MCP is dead, which is ludicrous. Right. Any views on these stories that you're seeing out there? Look, people are always looking for the next thing, you know, always looking to be ahead of the curve. I get that.

[00:30:20] I don't see anything else out there which is solving the context problem. All that is making integration easier. And runtime integration for the world we're going to is, I think, going to find far more favor in many cases than development time integration.

[00:30:40] So, you know, you've got to anticipate what you integrate with because the inventive ways with these agents which have been built are being built. They're going to, you know, they can be moving so fast in their needs and what they need access to. Having native integration to everything, I just think, is going to be beyond all of us. There'll be times when you do need it, as I said.

[00:31:06] There'll be times when you just can't anticipate it and you don't have it and maybe you only need it occasionally. So you use MCP. And Anthropic published this protocol a couple of years ago, I think it was, late 2024. I haven't seen anyone else publish a competing protocol, but I could be in the dark there. Maybe I'm behind the curve.

[00:31:27] Well, I think for the most part, many of people saying this are YouTubers, content creators, all saying controversial things for extra shares, extra visibility. That was my understanding. But, hey, maybe people can point me in the right direction too. But for anyone listening wanting to find out more about all you're doing at Twilio, how you're using MCP, how the Twilio customers can use MCP, where would you like to point everyone? Look, particularly our developer blog.

[00:31:54] This is where you'd find, you can find more information about the Twilio MCP server and its current status. And particularly for these topics, yeah, the developer blog, just search for the developer blog for Twilio and that'll be the best place. Awesome. Well, I'll add a link to everything you mentioned there. And I'd encourage everyone listening, if you are hearing about it for the first time, go out and have a play. Set up a local Twilio Alpha MCP server. Use it to complete a basic task.

[00:32:24] Test how context and tool filtering affect performance and agent reliability. And please report back. I'd love to hear from you. But more than anything, a big thank you, as always, for coming on and explaining this in a language everyone can understand. Really appreciate you, as always. Thank you. For me, one of the biggest takeaways from our conversation today is that MCP matters because context matters. And AI has promised a lot over the last couple of years.

[00:32:51] But too often the conversation is centered on what the model can say rather than what the system can actually do. And as Peter made clear today, if an AI system cannot access the right tools, the right data and the right live context, it will always struggle to deliver consistent value. Value that goes far beyond just another polished demo. And what I also appreciated in our discussion today was the balance.

[00:33:18] Yep, there's real excitement around MCP right now and rightly so. But there are also practical questions around governance, security, permissions, observability. And sometimes an API integration might still make more sense. And I think it's this kind of honest conversation that we need more of. Less noise, less hype, more focus on the problem in front of us. And that may be the lesson worth holding on here.

[00:33:46] Start small, solve something real, learn from it, then build from there. So whether you are a developer or a business leader or someone just trying to understand where AI is heading, I think you'll agree that MCP feels like one of those ideas that could quietly shape the next phase of this market in a very big way. And if this episode helped demystify MCP for you, let me know.

[00:34:10] And if your organization is already experimenting with MCP servers, AI agents or new ways of connecting models to use very real business systems and solve real problems, what are you seeing so far? I want to hear from you. Techtalksnetwork.com. There's lots of different ways you can get hold of me there. But that is it for today. So a huge thank you to my guests for demystifying this space today.

[00:34:37] An even bigger thank you to each and every one of you for listening to the end. If you enjoyed yourself, why not come back tomorrow? We'll do it all again. Same time, same place. Speak with you then. Bye for now.