Latest / State of the AI Union / MCP is the AI MVP
Transcript
- 0:10Hello, everybody, and welcome to this week's episode of the State
- 0:13of the AI Union. I'm your host, Laura, and join
- 0:16with me today is Ribu Chawa. Ribu, do you want to introduce
- 0:20yourself? Hello my name is Ribu.
- 0:22I am a product manager at Devref.
- 0:25I'm working on CX agents and a lot of interesting stuff around
- 0:31AI agents offering. Awesome.
- 0:33And Ribu, you're a return guest today, so we're very excited to
- 0:36have you. Thank you for having me.
- 0:39Of course, And today, our episode we wanted to talk about
- 0:43the MCPS. So first of all, what does MCP
- 0:46stand for? It's a model context protocol.
- 0:49It's an open source protocol. This was found by Anthropic last
- 0:54year, but really picked up this year and a lot of different
- 0:58players have picked it up, including open AI.
- 1:00So, you know, in today's landscape, if open AI and
- 1:05entropic, both of them are supporting something, you know,
- 1:08it has weight. And we can go deeper into what
- 1:10it what that really means for end users, for businesses, for
- 1:13companies like Devrev, for companies who are looking to use
- 1:17products like Devrev. Yeah.
- 1:19But this is very impactful. This is one of the major
- 1:21milestones in the AI landscape, AI journey until now.
- 1:24Awesome. So model context protocol.
- 1:27OK, so first what does it do? What does the MCP do?
- 1:29So let me put it much simply for anyone to understand.
- 1:33So if the LM is like a hardware where everyone has been excited
- 1:37about the hardware itself until now, you're going to have more
- 1:41layers on top of that, right? So we see Agents as more like an
- 1:45operating system where MCP is like USBC.
- 1:48So now you can imagine how valuable USBC is to all of your
- 1:53devices that you use could be different hardwares, but that
- 1:56port enables a lot of integration, a lot of
- 1:59reusability and interoperability.
- 2:02So MCP is that in the landscape of softwares until now, one of
- 2:07the studies recently suggested that failure rate in AI
- 2:11applications is around 80%, which is double the non AI
- 2:15applications from past Wow. So that's huge, yes.
- 2:20And the reason being one of the reasons is that the whole AI
- 2:23world is probabilistic. It's not deterministic, unlike
- 2:26the previous world. And in this you're only as good
- 2:29as your quality of data. So that is very, very important.
- 2:33The way early users have been doing is that they would try to
- 2:38integrate one of application very tightly because there was
- 2:41no such protocol existing in past.
- 2:43So if I have to integrate, let's say Salesforce, I would just do
- 2:46that API integration. But I don't know about the other
- 2:49APIs. Or it can only work in one use
- 2:51case and not the other use case. And that data coming from the
- 2:53API is not really originally designed for this use case.
- 2:57So some of the use case or one or two test cases really worked
- 3:00well. But when you go to production,
- 3:02the user don't care about what you're using underneath, right?
- 3:04Especially because of ChatGPT, everyone expects magic.
- 3:07It's like like type something and everything that I ever
- 3:10wanted would just happen, but it doesn't happen out-of-the-box.
- 3:12So a lot of people have to work really hard.
- 3:15So the whole generation text generation process, it took us
- 3:20many, many years to reach to that part where we could
- 3:23understand what you're saying and generate text out of that.
- 3:25But the next layer where we want to do things as well, not just
- 3:28generate text. Now doing things also in a loose
- 3:31way is connected to generating text.
- 3:33Because if you can generate text, you can generate
- 3:35instructions and you can give those instructions to those
- 3:38tools to perform those tasks for you, which is where MCP comes in
- 3:41place, where your LLM can now easily talk to other tools using
- 3:46this protocol. So if you have a dev Rev and you
- 3:48have let's say Slack or you have your Gmail and you have other
- 3:52tools, how do you like talk to these tools?
- 3:55And they would also be evolving. Their API's will be evolving.
- 3:57How does your LLM go and do things for you and get those
- 4:00answers for you from those tools?
- 4:02That's where MCP comes in place. So it's the API version of
- 4:06Agentic World. So, so Ribu, you know, you're
- 4:10comparing it to, you're comparing it to APIs.
- 4:12And you know, you mentioned earlier that in the past, right,
- 4:17we would be using APIs to tightly integrate and transfer
- 4:20data from one system to another. And that's how the LLMS would
- 4:24work. They would just work on the flat
- 4:27data transfer. Let's just say right now, are we
- 4:30saying that with MCPS and now that we are going into this, you
- 4:34know truly I think agentic AI type of world, does that mean
- 4:38that API's become obsolete? No, not at all.
- 4:41Absolutely not. So MCP is built on top of API's
- 4:45and if it doesn't, if it didn't, it would not have worked because
- 4:48that supports the backward compatibility and it ensures
- 4:51that all the millions of the users don't have to rework their
- 4:53systems. So let me put it this way.
- 4:55Let's say let's figure out like how does the agent actually
- 4:58work? And then we can easily relate.
- 5:00How does MCP help here? Let's say you want to you want
- 5:03to buy a new pair of glasses, a new pair of shoes on Amazon.
- 5:07So what would you do? You would go to Amazon, search
- 5:11something on the search box, then go through the list.
- 5:14You already know your requirements.
- 5:16So you'll apply a filter for size filter for like colors and
- 5:20maybe OK, you like some of them and you know that this one looks
- 5:23good. This is what I find.
- 5:24This is what I want. Now, if you were to do all of
- 5:26this agentically, if you want AI to do this for you and maybe
- 5:30like give you those like shortlisted 5 things, which it
- 5:33knows your taste, then it would have to call one API, then
- 5:37separate API, then identify this API needs to be giving me this
- 5:41much result. And if I don't get results from
- 5:43this API, let me try a different API.
- 5:45So to take all those decisions on the fly for agent to like
- 5:50which should I call? Which endpoint should I reach?
- 5:53What is the input to that endpoint?
- 5:55And after getting response, am I satisfied or not?
- 5:57And if I'm not, then let me like this is my backup API.
- 6:01This is what I should call. All of that information, all of
- 6:03that, all of that work has to happen by agents And the way you
- 6:07do all of that, they you are able to understand what API is
- 6:10to hit and what sort of how to like filter requests on the API,
- 6:15how to expect response. All of that is part of the MCP
- 6:18protocol. So without MCP, your agent would
- 6:20not would have no idea how to talk to Amazon.
- 6:23It would have no idea like what all Amazon can provide.
- 6:25So are we so like, could we analogously say that like MCPS
- 6:29are like the orchestrator of the API's?
- 6:32Yes, OK, Yes, OK. In a Lewis way it helps
- 6:36orchestration here. So it gives all the context to
- 6:39the agent so that it can further orchestrate.
- 6:42That's why Agent and MCP works really well.
- 6:44So I'll give you a contrast today, like where we are in the
- 6:48whole MCP journey in Postman around open APIs, there are like
- 6:52100,000 plus open APIs available, which means both
- 6:56100,000 products like Google, like Dev Rev, like Notion, all
- 7:00of those tools have their APIs registered that someone can use
- 7:04their APIs. And in contrast, there are only
- 7:06like few thousands of MCP servers available right now.
- 7:10So the MCP marketplace right now is in single digits compared to
- 7:14the API marketplace. That's how early we are in the
- 7:17game today. And just like someone could hook
- 7:19up Devrev's MCP server and use any MCP client to operate
- 7:23entirely on Devrev. So someone could use, for
- 7:25example, Entropic and say like, just create this ticket for me
- 7:28and Topic will go and create a ticket in Dev Ref.
- 7:30So that capability is limited to very few companies because a lot
- 7:35of people are not as early in this game.
- 7:37So that's what MCP unlocks for companies.
- 7:40So Dev Rev is one of those companies which is opening its
- 7:43own APIs and opening its own MCP server.
- 7:47Yeah. So every company out there has
- 7:50like a set of APIs. They've been building APIs for
- 7:55the last 15 years, right? So now they're going to have to
- 7:58start thinking about MCPS, right, and building their MCP
- 8:02server. Is that correct?
- 8:04Everyone will have to think about MCP at some point, whether
- 8:07they're going to be consumers or they're going to be servers.
- 8:11That's the different approach. But everyone will play some
- 8:15role. If they want to stay alive,
- 8:17everyone will have to play some role in this ecosystem.
- 8:20So a lot of companies, for example, and people think of MCP
- 8:23in context of let's say Notion. That's like all of these apps.
- 8:28But even for internal servers, internal applications, you have
- 8:31API site, yeah, everyone works on APIs.
- 8:34So if you just think about internal world, which is, which
- 8:37is like all the external APIs are tip of icebergs, you know
- 8:41the internal stuff, internal APIs are like multi fold more.
- 8:44So if you have those exposed correctly, that's when you would
- 8:48see internal optimization, internal process optimization.
- 8:51That's when you see your AI agents working to multiply
- 8:56productivity of their employees. So they would have to adapt MCP,
- 8:59all the internal tools which could talk to each other
- 9:01internally, but just talk to each other.
- 9:03We'll have to adapt those. And right now, none of those
- 9:06tools talk to each other. Humans talk to each other.
- 9:08So one person goes to 1 tool, get the data goes to the other
- 9:11person who talks to the other tool.
- 9:12Our tools are not talking to each other, so our AIS can't
- 9:15talk to each other. That's why this is such a
- 9:17critical piece so. The first step I'm hearing, so
- 9:20now that we're going through this AI transformation where
- 9:23everybody now has to be in some, it has the initiative.
- 9:26So as part of the AI transformation, what I'm hearing
- 9:29is first step. First step is start to build
- 9:32your MCPS. I would say first step is start
- 9:34using MCP. Not using.
- 9:37Because then yeah, because then you'll realize the potential and
- 9:39you say, hey, why can't I have, I give this to my users, what I
- 9:43can do with maybe Gmail or what I can do with Notion or Dev Rev
- 9:48and all of these apps which are giving me MCP servers.
- 9:51Why can't my app also become like this that my customers can
- 9:55access my data and my customers can access our software through
- 9:59MCP, Right, right, right. You can start with internal
- 10:02users. You can start with their
- 10:03customers. Everyone has is in their own
- 10:04journey to find the right place. For all you know, your customer
- 10:08support chat, which is like your customers are talking to you via
- 10:11that support chat and you want to plug in MCP because you want
- 10:15to directly update their profiles or you want to like
- 10:18hook your different tools. You want to automatically create
- 10:21tickets if the customer is raising complaint or even
- 10:25generate a PRD or a requirement dog based on that.
- 10:27So depending on your use cases you will find different places
- 10:30where you could use. When you think of I think MCP is
- 10:33a very technical word. I don't think entropic has done
- 10:35a great job. It feels like an internal term
- 10:37which became so big. But just think of it as like a
- 10:40multi plug. Sure, what all can I plug in and
- 10:44like what tool can plug in to what tool to add value?
- 10:47And when you start thinking like that, MCP is just the layer.
- 10:51Got it, got it. And what's the level of effort
- 10:53for an organization to stand it up?
- 10:55On their own to start using MCP. It's pretty simple.
- 10:59To stand it up for their own organization so that they can
- 11:01be. Their own, yeah.
- 11:02Their own server, Yeah, it's not as hard as it may appear because
- 11:08the documentation is public. It's very, very easier.
- 11:10The tools like Zapier, which helps you build your MCP server
- 11:14on top of your APIs, if you have good documentation, good API
- 11:16documentation, it's not as hard to start setting it up.
- 11:19You would still need to figure out like everything else,
- 11:23direction is equally important as the the intensity.
- 11:26You need to figure out what all tools you want to expose.
- 11:28It all starts with what are the use case you want to solve.
- 11:30If someone comes and say like I want MCP and they have no
- 11:33clarity of why they want MCP then of course you will be in
- 11:37the 80%. So very important for you to 1st
- 11:41envision the end customer use case you want to solve because
- 11:43that'll tell you what all should I expose in my MCP and then you
- 11:47start trying it out. It's same as API's in terms of
- 11:50like how do you expose something.
- 11:52If you have API, you can easily have an MCP on top of that and
- 11:55host an MCP server. But just identify what tools you
- 11:58want to expose because you don't want to do too many.
- 12:00If you expose all your 300 APIs, that means the agent will always
- 12:04have to identify. And if two API documentation are
- 12:06very similar, agent will take more time to like identify right
- 12:10API and hit it and then realize oh this does not give that
- 12:13information that can do this one.
- 12:14So that might cause some problems.
- 12:17So you want to find the right number of tools to expose?
- 12:20So if you have good API architecture, then standing up
- 12:24in MCP after that should be fairly straightforward, not so
- 12:27complex. Great.
- 12:28Now, OK, so so we talk about, we talk about organizations using
- 12:32MCPS to be able to become a genetic in nature right from
- 12:38just let's say, you know, reading text or whatever.
- 12:41What next? What else do organizations have
- 12:43to do to become like? Already.
- 12:46Yeah. So MCP, as I said, it's one
- 12:48layer, it's a connector. It it doesn't solve for
- 12:51everything, right. Just just because your computer
- 12:54has a USB C port doesn't mean that like you have everything in
- 12:57place. You still need all those
- 12:59peripheral devices to connect to those ports and you still need
- 13:04the operating system to understand in the old days,
- 13:06there used to be drivers. Now all of those are straight
- 13:09into your OS or automatically downloaded, but you still need
- 13:12all those peripherals to work and to put into MCP.
- 13:17So MCP is 1 layer for enterprises.
- 13:19You also want to be permission aware for enterprises.
- 13:21You also want to ensure that the data is protected like whatever
- 13:25data comes in. If you only rely on MCP, then
- 13:30you would start hitting ceiling very, very soon.
- 13:33And this is 11 place that we are experimenting in develop and we
- 13:36have a very good visibility now of when and where to use MCP
- 13:40versus when and where to use other methodologies to use data.
- 13:45Because if you want to search, for example, search information
- 13:48and you use MCP, you would be relying on the that specific
- 13:52tools own search mechanism, which may or may not be as good
- 13:56compared to like whatever environment where we bring in
- 13:59data and we take bring all the governance in as well.
- 14:02And we build a knowledge graph, we enrich the data, the data, if
- 14:06there is some support data coming from one tool and then
- 14:09some customer data coming from another tool and then some
- 14:12project management and development data coming from the
- 14:15third tool, we make sense of all those data together.
- 14:18So MCP would not be able to do that.
- 14:20And then when you search on MCP, it will easily take you minutes
- 14:23if you have more complicated search, right?
- 14:26But same search when you do on a tool like Dev Rev, then you're
- 14:29talking about a few seconds. So that's only possible when you
- 14:32identify when to use which tool. So that's where tools like us
- 14:37come in place, where we use MCP, but we also use a lot of other
- 14:41things. So for example, you can't have
- 14:45complicated workflows because MCP is just one layer of
- 14:47communication. So when you realize it's a layer
- 14:49of communication, then you build on top of that.
- 14:52This is you have all these agentic workflows and all the
- 14:56CRM and all the goodness that come on top of it.
- 14:59So permission aware is important, MCP, these are also
- 15:03important. And then what are some other
- 15:05things that organizations will have to think about?
- 15:08Like what about how they store and manage their data for
- 15:11example? Does that change?
- 15:12Yeah, yeah. So that depends on the
- 15:15implementation style. With MCP, that may not change a
- 15:17lot or the data could keep residing in the tools that they
- 15:20are using. But then that comes with the
- 15:21downside, some of the downsides as I mentioned earlier.
- 15:24So I think MCP is a very good starting point for someone who
- 15:27is not at all into agentic experiences yet so that they can
- 15:32and start experiencing and start feeling those pains for
- 15:35individual users. Like my wife use MCPA lot to to
- 15:39like look for Google Docs and look for Gmail and look for
- 15:42other things in one place. So for her that makes sense.
- 15:44But when you want to do that at the org level, you want to make
- 15:48sure like I'm not just looking at my inbox.
- 15:49I want to look at like everyone's inbox, like all the
- 15:52support inboxes and I want to correlate that information.
- 15:55So when you want to do it at the enterprise level, then you think
- 15:58about like the workflows that I want to run automatically.
- 16:00So workflow should be running MCP for me.
- 16:01I don't want to be thinking about doing this myself.
- 16:04So And how do you like build analytics on top of that?
- 16:07So when you start thinking about the enterprise offerings, then
- 16:11you realize, OK, this is just one of the layers, but I need
- 16:14like more offerings on top of this.
- 16:16I want the whole AI stack. I just don't want like one piece
- 16:19of it. Got it.
- 16:20So if I'm hearing you correctly, what we're seeing here is that
- 16:24the MCP will help us kind of make those decisions on like
- 16:28which APIs to call, right? And so they could do it at an
- 16:31individual level and be as fast as the individual basically.
- 16:36So, so just, you know, they would have to make a decision,
- 16:38OK, it's this API or this API that be whatever.
- 16:41And that would work. But if we're moving to
- 16:43organizations, having the MCP do all those decision making at a
- 16:48scaled level is going to have the same effect as it's going to
- 16:54have the same effect on on the lack of scalability as like an
- 16:57individual human would have like it would have to call individual
- 17:00things. So there has to be a way for the
- 17:03MCP to kind of leverage all of the data together, which is a
- 17:06separate strategy from yes, from their communication strategy.
- 17:11Is that what? I'm hearing yes, yes, because
- 17:14MCP can bring data or MCP can send an instruction.
- 17:17OK. That's what it is doing.
- 17:19But what do you do with that data?
- 17:21When do you send that instruction?
- 17:22What instruction do you send? Is it always human generated?
- 17:25Is it automated? Is it like filling your pipeline
- 17:28in with some more context or is it just like one shot?
- 17:31I just ask something, get answer, finish.
- 17:33That's a very simplistic use case.
- 17:35With Enterprise that's there are much more complicated use cases.
- 17:38So here's a plug for Dev Revs Knowledge Graph.
- 17:41Tell us a little bit about how Knowledge Graph solves for that.
- 17:43Yeah. So when we import data, we
- 17:46ensure that we are bringing all the.
- 17:49So let's talk about the problems that anyone who wants to bring
- 17:51in data, like anyone could do this, right?
- 17:54So when anyone wants to bring data, they'll start figuring out
- 17:57that this does not have permissions.
- 17:59First of all, my customers like all my tickets, which were which
- 18:02had very well like walls protecting walls there and
- 18:05firewalls here are like out and open here because anyone can see
- 18:08anything now. So you need to bring all of
- 18:10those policies, all of those rules first, then bring in data
- 18:14and then make sure that like everything is mapped correctly
- 18:16in this new, new residency as well.
- 18:19So it has to be permission aware.
- 18:21And you want to make sure that in the agentic search that
- 18:24permission resides. So if I search for a document, I
- 18:27should not be able to search like Laura's document, whereas
- 18:30AI has access to everything when responding to remove it never
- 18:33ever can. My policy can never look at
- 18:36Laura's doc. It can't even look at the doc.
- 18:38So we want to make sure that that happens.
- 18:40Then when bringing data, you want to keep the data clean and
- 18:43only enrich it, not make it worse.
- 18:45So when you bring the data, you need to have all that amount of
- 18:49customization available in the platform that I can mirror the
- 18:53data state there while also enriching it further and say,
- 18:55OK, these are all the fields which are searchable.
- 18:58These are the fields which relate to this other data form
- 19:00in this way, like this is the customer.
- 19:02The customer we're talking about in the finance tool is the same
- 19:05customer we're talking about in the support tool and the same
- 19:08customer that team is building a product for.
- 19:10So it's the same customer which relates everywhere.
- 19:12So that's where the knowledge graphs comes in.
- 19:14So the data that coming in comes into the dev RIP is becoming
- 19:17only richer without losing any of its original qualities that
- 19:20it had in the earlier tool, which makes AI on top of it much
- 19:24more efficient. So that's the 10X value we had
- 19:27where all of the data is super clean and now you can
- 19:31identically ask questions, you can do a lot of automatic
- 19:34workflows, workflows and operations on top of it.
- 19:37So one of the things that we keep in our mind is every time
- 19:41we are thinking of these new features or new workflows,
- 19:44workflows is it, how does a human do it today?
- 19:47How can we help that human in this format?
- 19:49So agent can take those decision.
- 19:51And for that we need very, very rich data.
- 19:53So that's what the knowledge graph enables.
- 19:55That's what make develop unique. And then on top of that, the
- 19:59agent offering we have, we smartly decide when to use MCP,
- 20:03when to import data and store it ourselves.
- 20:06There are there are places where MCP like really, really helps
- 20:08us. So it's not an enemy.
- 20:10It's really a big leverage that we have today that.
- 20:13Be able to smartly decide when to use MCP, how and give fast
- 20:19responses fast resolutions to the end customer with same
- 20:23accuracy. Got it.
- 20:24Got it. Tell us a little bit more about
- 20:27these efficiency improvements that you've seen with MCPS on
- 20:31top of the dev Rep data. Yeah.
- 20:33So a lot of those things are still yet to be announced.
- 20:37So there are a lot of projects happening internally.
- 20:39But I can tell you that with MCPA, lot of things that would
- 20:43have earlier taken us weeks or months to develop.
- 20:46We are able to develop those things in days while still
- 20:49keeping like there is offering unique while still making sure
- 20:54it plays well with the rest of our product.
- 20:56And that's for the knowledge graph that we have built.
- 20:58OK, wow. That's a big improvement.
- 21:00Yeah, absolutely. And we have a lot of interesting
- 21:05announcements coming next week. So a lot of these offerings will
- 21:09become more public soon, so. And beyond like let's say like
- 21:16on the development side, what do you think on the user side, do
- 21:19we really get to see the benefit of the MCP?
- 21:22I think the best technology and best designs are transferring.
- 21:26So as the end user, what you should and what you would really
- 21:30care about is does the software solve my problem?
- 21:32Sure. So if we could integrate with
- 21:36one tool or if we would integrate with two tools, you
- 21:39don't care if like behind the scenes is MCP or TCP or
- 21:43something else that would have created TCP.
- 21:46What end user really care about is like, does this really work
- 21:49for me or not? That's what our focus is.
- 21:52Our focus is not to make sure we use MCP or make sure we don't
- 21:56use MCP, which is why I said there are places where MCP
- 21:59really improves the performance for us.
- 22:02And one of those places is fast actions, for example.
- 22:05So if you want to take an action in some other tool, you can take
- 22:09it really, really fast. And then we can make sure that
- 22:12we can do those actions while ensuring that knowledge graph is
- 22:17as rich as it was earlier. So MCP really complements our
- 22:21knowledge graph that way. So as an end user, maybe like
- 22:25it's speed like we would, we would experience speed and
- 22:28efficiency and the, and the way things are done, we like, we
- 22:31don't have to wait so long, you know, and it's, it's kind of
- 22:34taking the right actions. Yeah, it will both quantity and
- 22:37quality together. So as as we vet things
- 22:41ourselves, you will slowly and gradually have more things you
- 22:44could do, and the things you can do become really fast and
- 22:47efficient as well. Awesome.
- 22:49All right, so to sum up our conversation, MCPS are the
- 22:53orchestrator of the APIs. If you haven't used 1 yet, or
- 22:57your organization doesn't have an MPCP, you might want to start
- 23:00thinking about it. The best way to start is to just
- 23:03start using it and that MCPS alone are not enough to become
- 23:09AI ready. You ought to also think about
- 23:10your data, what to do with it, how your data is permission
- 23:13aware etc. And reboot.
- 23:15Is there anything else that you wanted to tell us about the MCPS
- 23:18that we haven't covered yet? Yeah.
- 23:19So you don't need to think a lot about like don't need to be
- 23:23worried about this, this very, very interesting and powerful
- 23:27thing in the world. I don't want to call it feature
- 23:29because it's more than a feature.
- 23:30And tools like Dev Rev work on the both spectrum of MCP.
- 23:33So we are both an MCP client and an MCP server, which means that
- 23:37we can talk to other tools who support MCP as well as we also
- 23:40support MCP. Someone who is using let's say
- 23:42cloud or even ChatGPT plus recently launched.
- 23:45So someone who's using those tools and want to use Dev Rev,
- 23:50they could YMCP, they could do that as well.
- 23:52And a lot of of us internally as well use this very frequently.
- 23:57So. Yeah, and if you don't have one,
- 23:59use this tool like Devref that supports it so that you can
- 24:02experience the benefits of it. Absolutely.
- 24:04Reboot, thanks for joining us and to our audience, thank you
- 24:07for listening. If you enjoyed this content,
- 24:10feel free to leave a comment or give us a like so that we know
- 24:12to give you more. All right.
- 24:14Thank you and see you next time. Thank you.
- 26:52Music.