Download Otter for your meeting notes

20250919 MCP Exploration | Otter.ai

20250919 MCP Exploration

Rob Hirschfeld52min

Rob Hirschfeld

00:06

Hello, I'm Rob Hirschfeld, CEO and co founder of RackN and your host for the cloud 2030 podcast. In this episode, we continue to explore vibe coding and specifically MCP or the model context protocol, which is really a interface between a chat bot and a back end system. We talk about how they're built, what they're doing, how they work, and even if you don't have access to the video here, I think the discussion itself will provide you with a lot of interesting insight in what MCPS are, how to build them, why they're important, what challenges you might face. So a really interesting discussion as part of our broader exploration of vibe coding over the next several weeks.

Jon Aykroyd

00:56

So MCP is a model context protocol. It's basically a well protocol for AI agents to communicate with various different tools. So, like, the the demo that they start you off with is like, you know, getting the weather right, making a call to, like, weather.com to get the weather, and then, oh, now your agent, can, you know, get the weather, which is cool. I mean, I guess, right, every there's always a Hello World, right? But, like, that's not really useful, you know? And so the idea is that, you know, we want to, we want to provide tooling. We want to make these agents able to do things that they're not able to do out of the box, necessarily, right? We want to teach it, you know, how do we interact with DRP? And we already have, you know, DRP, CLI, right? Or, or we have the, the derby client for, you know, Python, communicating with API. I mean, at the end of the day, right? It's just API calls and so where we're basically, you know, teaching, teaching the agent how to make AI

Rob Hirschfeld

02:19

calls. Oh, cool. This is exactly what I was interested in

Jon Aykroyd

02:23

talking more about the the idea is, our user makes a query, right? Show me all of the machines in my infrastructure, or show me all clusters, or, you know, show me specific information about a machine that's forward to the MCP client, which is then, then talks the MCP connector, which we're using, where we I'm using, the fast MCP server, Python library that basically provide, provide some functions for interfacing with the various different agents. And then I've got a bit of some natural language processing to say, hey, you know this means this. This means that, you know, if I say, you know, we're creating a cluster, then you know, we're looking for specific key phrases, right? You know, we want to when we create the cluster, we need to know, like what the context is, or when we're looking at machines, we might, you know, want to look at the nick or we might want to look at disks, or we might want to do, do this, that or the other, right? And so we need to kind of figure out what those different pieces are. Yeah, I don't know that that's strictly necessary, but it seems to help in guiding the agent to the right

Rob Hirschfeld

03:59

options, right? Makes sense. This is simpler, similar to what I was thinking, where I was thinking, my next steps would go it in, like having prompts that sort of cover the basics for

Jon Aykroyd

04:13

Yeah, well, so I've been kind of doing it. What I would say is the hard way in, kind of manually, manually drawing out functions, right? So here's, you know, for instance, here's a cluster definition, right? And so it's got, you know, various different components, right? The the UID, or, you know, ID, or these different, these different pieces. So we need to explain, you know, what are the objects, right? And then we say, okay, hey, let's go talk to well. So this is. Is actually, so this is the actual tool that we exposed to

Rob Hirschfeld

05:10

the the the chat clients itself, okay,

Jon Aykroyd

05:14

but then this is the client that actually talks to DRP. This is what actually makes the various different API calls to actually affect change or interrogate the cluster. And so like the various different functions, like list clusters or create cluster is then basically called from the MPC server function over here, right? So why

Rob Hirschfeld

05:53

do you need this in front of the just the regular API,

Jon Aykroyd

05:56

well, so that is where I think actually and then. So that's where I said we're kind of doing, I'm kind of doing this the hard way by, by making all of these functions because they already exist in the derby client. So I think we could probably, I think there's probably a way to generate all of these functions. Instead of, like, you know, creating a list machines, creating a create machine, creating a because that's all already in the in the DRP client, we already generate that with swagger. So I think there's a way to, like, introspect the client, or maybe even just leverage swagger to generate whatever code we need to interface, because

Rob Hirschfeld

06:55

working on a prompt set that would because, you know, as I was playing with this, there were things that I had to be like, No, you need to think about it this way. And so I was working on prompts that would help nudge the behavior. But I you're trying to do something different, and I'm interested

Jon Aykroyd

07:12

in, yeah, so with, so with the various different tools, right? We basically have to tell it, you know how to connect, right? So if we do broad MCP list, we'll see that there are no MCP servers configured, right? And so we want to allow mutations. And I'm going to go ahead and just set the password here. I set it to the default because we were recording so I didn't want to expose any, you know, actual passwords. But you know, in the real world, you would, you would, you know, use a real password here. But then we can run this command here. We're going to say scope user, so that any invocation of cloud will have access to this tooling here, we set our various different authentication variables right where, where is DRP we happen to be running on the same server, So it's local host, but you could set any endpoint. At the moment, it only handles one DRP server. But in the future, I'd like to be able to, you know, either through, you know, multi site manager, or just being able to connect directly to different, different endpoints and make API queries. I don't, you know, I don't know, but for now, it just connects to the one server and knows how to deal with that. And then the allow mutations flag here, which says, hey, you know you're you're actually allowed to make changes on the cluster. And do do destructive and constructive operations. You're do, you're allowed more than just read only, um, and then the the actual entry point, which is this main.pi which really, or here, which really just kind of imports everything and starts everything up. You know, it's the, you know, the library of files that actually, you know, is doing the work. That's just kind of the the entry point, right?

Rob Hirschfeld

09:36

So this is because this is where I was, where I was thinking it would, it might go if you said, Hey, I need to know which servers, you know, tell me a list of all the Dell servers that have, you know, InfiniBand Nix in them. That's what. That's where you're going with this, right?

Jon Aykroyd

09:56

Yeah, exactly, yep. And so I've got, I've got a list. List of example queries here. So I've got just for stuff to look at. I've got a few machines, and then these are containers here. I just have the one, one cluster to kind of start with. And we can say, you know, show me all machines. And it knows, okay, I'm gonna go use this query infrastructure, and it, it doesn't show us the full result, but if we do control our we can see the full result and see, okay, this is, this is the list of items that it got back. Inter, interestingly, earlier, it showed me, oh, here's the machines that are are in clusters, and here's the machines that are not in clusters. But, you know, we could go, Okay, show me all containers. Oh, yeah, okay, so that's so that's wrong.

Rob Hirschfeld

11:04

Well, this is where you you'd end up refining right exactly the MCP code we could do, would you? Would you then think to do it like, oh, the type of server is probably a top level thing, and I need to help. I need to help find the in the MCP code servers by type like, would that be part of what you would end up doing? Um, this, because understanding how to what type of server is is, you know, knowledge that you would want to embed in the MCP I'm assuming.

Jon Aykroyd

11:40

Yeah, right. How would you express that? So I think, I think the problem is probably it's not, it's not filtering correctly, so it doesn't know what filters to apply. So when we say, Okay, show me all all all machines that are containers. It doesn't know. It doesn't know what filters to apply to say, Okay, what is a container versus what isn't a container? When I say, Okay, show me all all interfaces. It's able to filter through the parameters and go, Oh, I can look at, you know, it's

Rob Hirschfeld

12:24

pretty impressive. Can you, could you ask it, like, look at a machine and tell me what type it is from its data?

Jon Aykroyd

12:34

Um, so

Rob Hirschfeld

12:35

let's see, and then, and then feed it back to, I'm assuming you quad code the MCP, right? One of the things that's interesting to me is, what is, what is the development cycle? Look, you know, as a general thing, I'm trying to figure out, like, what's the development cycle? Are you when you're doing this work? Do you get to say, hey, I want you to improve the MCP by, you know, enhancing your understanding of type of machine and then encoding

Jon Aykroyd

13:04

it somehow. So unfortunately, it doesn't seem like so yes and no. So, because cloud is running in this directory, right, it knows about this project so it can work on this project, but you have to kind of reload after you make changes, you have to reload bot and say, Hey, I've made changes now. Now restart and reload the the new changes so that you can interact. It's not make live changes. Yeah, okay, but, but it is, it is, you know, you're able to go, Oh, I got this response. I want it to be more like this. So I think it just needs, it just needs to know how to filter the the machine object, basically, right? Because it's getting, it's getting the whole object back, but it doesn't know, it doesn't know what to look for, basically. So I need to, I need to look at that. But I can do kind of, you know, you do, like, show all clusters. So I definitely think, like, some, some generalization would be nice. So now I can say, Okay, show me so on the back end, it's, it's querying and going, Oh, hey, you know, show me machines that have, you know, the the profile manual assigned to it, they don't have an IP, I guess, because. There are containers that makes sense, but then we can also tell it like, give me a cluster, right? And you know if we want,

Rob Hirschfeld

15:13

okay, yes, that's pretty specific. Does it go all the way down to setting the parameter foo? Yep, and, but you coded behind that to have it understand what those mean, or does it able to interpret that?

Jon Aykroyd

15:29

Yes, yeah. So here there's, there's resource keywords, right? So if we see machine, machine servers. We know that we're talking about the machines, resource C, cluster, clusters. We're talking about cluster. And then here's basically the query where we're talking about the various different components, right? And I feel like this has gotten kind of out of hand. Well, it it started. It started off as just a couple of fields, but then, you know, as we start, as I start adding stuff to it, it starts getting more and more complex, and it's like, wow, okay, you know when, when I started on this, it was only a couple of lines, and now it's what you know, you know, 20 lines, 10 lines, 15 lines. And so where we're in here, yeah, so and I had some help with all this, with all the matching. But you know where we need to, you know, convert numbers to actual numbers. When we make our make our query, right? You know, here's, here's our list of operations, right? Where, you know, you might say any of these, but it's still the it's still the delete operation that we're going to run.

Rob Hirschfeld

17:30

And so it's, it's surprising to me that this is like code and not just a prompt. But maybe I'm, maybe I'm under thinking it, or overthinking it. I Well, because you're showing your what's up. I mean, you're, I'm trying to get a sense of this feels very it's like a, it's a, it's a proxy, it's a shim, which makes sense. Yeah.

Jon Aykroyd

18:02

I mean, you're, you're essentially teaching the agent, you know, how to make these calls. I'd be curious to see

Rob Hirschfeld

18:18

how to from an open ended English system into a, you know, into a more constrained this, you're basically teaching it the lexicon.

Jon Aykroyd

18:29

Yeah, yeah. I mean, and that's where the and that's where the natural language processing comes in, right? That's where this NLP comes in. Is the how do we, how do we get it from English into something meaningful, right? I There is. It's funny, I have another attempt at this.

Rob Hirschfeld

18:54

You're answering the question I ultimately was hoping to spend the next couple of weeks answering, or you're starting to address, which is, how are our customers going to use AI to improve their Digital Rebar experience? And you know, the simplest answer is they're going to use a chat bot to say, you know, hey, tell me all the servers where the BIOS is out of date that you know. You would say, Well, what does out of date BIOS mean? And and you would say, Well, it's, you know, not the current, not the current Rev. Or, you know, the bias check hasn't been run in the last 30 days. Or, you know,

Jon Aykroyd

19:33

yeah, well, so there's two, there's two potential use cases that I can see one is exactly what you're talking about. Is just, you know, let let the agent go to town, let it do work. I'll just, you know, kind of be project manager and tell you what I want done. You go figure out how to do it. The other is as kind of. A education tool, where you go, Okay, how would I list all clusters? And then it can respond with, oh, you could run DRP, CLI, clusters list. I have not addressed that second use case at all, right, that's not at all what, what? What I've been doing here. I've been focused on that, that first use case like you described, where, you know, we just let the agent kind of run wild and and have a good time, right?

Rob Hirschfeld

20:33

But, and that's, I think that's the more interesting. We tried the second one with the rag in front, and we had a lot of trouble. I mean, the systems were not nearly as powerful as they are today, but it didn't generate the caliber results that we were looking for, and it might now, if we revisit it. Well, so

Jon Aykroyd

20:58

Red Hat, actually, of all people, has their, what they call their learning lab, but they're essentially the mechanism, the mechanism I'm not entirely clear on, but you basically are able to to inject domain specific knowledge into various different models. And it's not quite a rag, it's, it's, well, it's not a rag, it's, it's a it's a whole, I don't know it's a whole different thing. I looked at it kind of briefly, but, you know, only so many hours in a day. But that might be, that might be worth in investigating for that kind of, that second use case we were talking about.

Rob Hirschfeld

21:52

But I'm at the moment more interested in the first use case, which is, you know, I want to build a get better intelligence, or build additional stuff in front of Digital Rebar, like, oh, I want to, you know, move all the Digital Rebar data into parquet files and then search the parquet files. Or parquet. A parquet is a, it's a unstructured. It's a semi structured data format that's used in scan learn, like machine learning and scanning a, r, Q, U, e, t, I think

Jon Aykroyd

22:40

French spelling, right? Yeah,

Rob Hirschfeld

22:45

like a, I think the parquet floor stuff, but in for s3 if you store data in parquet, it can it will scan it and index it, and then you can do searches off of those indexes to find data. Pretty cool.

Jon Aykroyd

23:04

Oh, interesting. Okay,

Rob Hirschfeld

23:07

so it's, it's structured, semi structured, so into an individual file is in a structured format that can be indexed. But the, you know, the individual files can just be dropped into a directory, but it's a Hadoop. It came out of the Hadoop.

Jon Aykroyd

23:27

Of course, of course, Hadoop. It's,

Rob Hirschfeld

23:31

but, yeah, it's not Jason. It's, that's semi it's, it's intense, structured, or more, it's closer to yaml,

Jon Aykroyd

23:40

interesting, okay, it probably is a very I

Rob Hirschfeld

23:44

haven't met, you know, but it like the stuff I've been building so far is you I want to write something that uses the Digital Rebar API in a in a new way, or in an aggregated way. One of the examples I thought of was like, if I'm Orion using digital rebar, and I'm like, oh, I want to make sure, I want to write code that will, you know, pull machines out of a pool if the or add machines to a pool, if my pool count is low. And so, you know, I was the MCP doesn't quite help you with that. There's a different thing to code for that.

Jon Aykroyd

24:30

Yeah, you would need, you would need some kind of, like, long running agent you.

Rob Hirschfeld

24:36

But I was just thinking about how you build the code for it, so I could put something on a timer that would would check, but I want to be able to, you know, like I was doing last week or two weeks ago, where I would say, hey, I want to write, you know, write a, you know, piece of Python that can scan Digital Rebar and, you know, rebalance my pools for me. You build a, you know, a cloud code file describing your MRD, and then you start putting in information. But there's a degree of like knowledge about the RP that has to be captured somewhere, and that that's what I'm like, Okay, how do we build up a prompt that somebody who wants to do Digital Rebar programming can start with to get a lot of the knowledge that you have captured in the MCP available to write fresh code. As you see, you see what I'm, you see what I'm. A lot of the stuff you have in that MCP is, really is, institutional knowledge about Digital Rebar. You're just adding a layer of English translation to to help la chat

Jon Aykroyd

25:49

behind.

Rob Hirschfeld

25:50

Or maybe I'm not understanding it.

Jon Aykroyd

25:53

I mean, it's, it's, it's basically teaching the agent how to use the the the derby client, and that's where I said, like I'm, I'm currently doing it the hard way, because I don't, I don't know a better way, but I think there's probably a way to like more, more generalize the code, so that either, either using like swagger to auto generate code, or just kind of, you know, Intro introspect, the derpy client, so that you know all of these functions aren't necessarily, aren't necessarily hard coded, you know, list cluster, right where you know that that already exists in the API so, you know, I, you know, we don't necessarily need to write another, you know, list client API call, right? I, like I said, I think, I think there's a way to kind of introspect our existing API and go, Okay, here's the functions that are available. And then from there, be able to say, you know, do these different things right? Like, right now, it didn't seem to, you know, be able to parse out, you know, BIOS information. I wonder, like we could try that, you know,

Rob Hirschfeld

27:26

no, I'm curious to see how we make the the MCP smarter, and then the stuff that makes it smarter, I want to be able to feed into the prompt for, you know, building general, other generalized DRP stuff, which is sort of that, that's where I keep, I keep coming back to, is like, Okay, if I am you a RackN customer, I my assumption is that we're going to start Having people use AIS to build RackN interface code, and I want to make that easy, because the likelihood of them doing it without our institutional knowledge makes it strikes me as painful, potentially risky, even

Jon Aykroyd

28:23

if bias info is absolutely available. So don't, don't lie to me. Yeah, there you go. But it's not in, it's probably not in some, some query.

Rob Hirschfeld

28:35

And you might, might be, you might be dumbing it down, right, yeah, right. Because you you might choose to say, you know, hey, I want list. I want my machines, but I'm not going to retrieve all the parameters from a performance like if I was building the MCP, you would want to say, hey, when you get the machines, don't ask for go high information, because it's really slow. So you can make a the request to not include go high but you can say, but if they're asking these types of questions, you you're going to need go I to get any machine data.

Rob Hirschfeld

29:15

Yeah, to say, you know, hey, don't, don't include go high data in the the go high parameter in the in the in the in the query. But if you're going to get back all the inventory data, you might have it in the inventory. Yeah.

Jon Aykroyd

29:36

Well, so when I, when I was originally doing it, it was just doing like a full list and getting all of the information. And it was, it was freaking out because it was, like, too many tokens. It was over. It was overflowing and, like, it just couldn't handle it. It was like, so I think you're absolutely right. I think that's probably what happened was, you know, I. I just never went back and went, Oh, give me a, give me a detailed. I mean, there, there should be.

Rob Hirschfeld

30:09

How does, how does the system, know, look at your MCP, what? What actions it can take. Is there a index for that, or is it looking at the code? Yeah. So, um,

Jon Aykroyd

30:23

right now I've just got the one query infra tool, which is kind of handling everything. And this is where I think maybe kind of rethinking this would would maybe help, because instead of just having one tool, you know, it could, instead of having just the query info like that, could be where the you know list clusters is, or the you know list machines, or all of that, because it's kind of abstracted now behind this query in for tool, um,

Jon Aykroyd

31:07

um, so that's because I'm handling the the natural language processing.

Rob Hirschfeld

31:13

So inside the MCP, that's That's partly like, like, it seems like the MCP should be able to say, here's things that I know how to do,

Jon Aykroyd

31:26

yeah, well, and so that's where, that's where this version here was kind of a different, a different take on that hadn't quite gotten it to work. So I wasn't really ready to show this one off yet, but this one actually exposes the tools, kind of, instead of you just see the one tool here, like this, it actually exposes list of tools. And so I'm not doing the the NLP here, you can see it's still, still defined the various different resources, but it's a little it's a little bit simpler. I think it might be easier to extend, but it wasn't. It wasn't quite working. So I wasn't really ready to talk about it.

Rob Hirschfeld

32:20

But your goal here is that you have a higher, higher level construct and can say, hey, tell me the machines. Tell me the whatever. Because this is the thing I'm not, I'm not seeing is, how does the you know agent, or how does the chat know what it can ask the agent to do, yes,

Jon Aykroyd

32:48

magic, no. I mean, pretty much, pretty much, right now it's, it's more or less just passing in the query directly to to to the agent,

Rob Hirschfeld

33:01

the MC MCP. MCP has to figure out how to understand, there's a degree of like, oh, I go ask Digital Rebar this question, and the MCPS answering that question, right? Exactly itself, right?

Jon Aykroyd

33:18

So, yeah, yeah, exactly. So, so it's doing, so I'm doing a lot of the work in

Rob Hirschfeld

33:29

also the person doing the chats, doing a little bit of the work by directing, knowing to direct the question.

Jon Aykroyd

33:37

Well, right, yeah. But I'm doing, I'm doing a lot of the work here, in this, you know, in in kind of breaking this out into the various different components. And that's where, you know, I think, I think by exposing, instead of, instead of having, you know, just one kind of one interface here, with the tool, um, having the various different tools, I think um is, is kind of the kind of the path forward for what you're talking about, as far As you know, making it smarter, so to speak. I think kind of by over engineering it, I inadvertently have maybe restricted some of the capabilities.

Rob Hirschfeld

34:31

That's the, I mean, that's the learning cycle. Is you're gonna have to go and play, and then you're gonna one have set, you know, just like you have sample queries or things that people would ask so that you can then figure out how to make it smarter and smarter.

Jon Aykroyd

34:46

Yeah, no, like the BIOS. The BIOS is a great, like, that's a great

Jon Aykroyd

34:56

version. That's a, yeah, I mean, that's. Super useful, like,

Rob Hirschfeld

35:02

the last time, last time updated. Yeah, so you would have something like, show me the versions, you know, show me the versions of operating system, the types and versions of operating systems that we have.

Jon Aykroyd

35:18

Yes, I think it can do that interesting.

Rob Hirschfeld

35:26

If you have tell me all the servers that are attached to this switch, or we have the LDP information to take, take that type of answer

Jon Aykroyd

35:38

attached to what

Rob Hirschfeld

35:39

no it was to a switch. You ran LDP on a system that should capture the LDP port and switch and information

Jon Aykroyd

35:50

is that something that is that, something that DRP is already exposing,

Rob Hirschfeld

35:59

I believe it has it if the if you're if it did an LDP sweep, yeah. So challenges when we when we run the systems, we rarely run with switches that can do an LDP scan. But when systems do LDP scans, we populate a parameter that has a whole bunch of data in it. I have the right system, right?

Jon Aykroyd

36:30

Yeah. So I think, I think kind of getting back to your point by, you know, making it making it smarter in in its ability to kind of interrogate and craft, you know, queries, right? Because, you know, if we want to look at the bios, we don't necessarily need the whole, you know, go high information, right? We just need the the BIOS info. Um, so figuring out how to, how to, you know, teach it to craft those kind of queries kind of more intelligently, I think, is, is probably my next step

Rob Hirschfeld

37:20

that makes sense. Well, and also getting, you know, the more specific examples that you have, the easier get the to test it and to figure out what you need to pass through. I'm really curious to see what happens on that. It'd be fun to have that type of back and forth where we were throwing sample questions at it. You could live, live, recode the MCP to get smarter about answering the question.

Jon Aykroyd

37:50

Yeah, I'm sure that's coming. That doesn't quite work right now, but there might be a way to set up so the way, the way I'm currently running it, as you saw when I, when I added it, here's read me when I added it. This basically creates a JSON config file. There's one that's actually not the right one. I actually set it up so VS code knows about it too. If we go to tools, we can see that it's it's here, so we can actually query it from here and go,

Rob Hirschfeld

38:44

you know, postman, yeah. So when I was, oh, wait, no, I'm looking at postman, here's

Jon Aykroyd

38:53

the way. The way it's currently running is it's just running against, you know, a static Python file, and it's just doing, you know, standard in out queries, okay, what? And so where, the reason it's called a server, you can actually run fast MCP, as you know, it's basically an HTTP server that you can, you can send JSON payloads to, and it, it returns JSON payloads. That's basically what we're doing here with the standard amount. But it's not, it's not going over a server, right? So, in theory, what we could do is actually set it up to run the server in one process and then connect and so instead of here connecting with Python, you know, give it the you know, 192168,

Rob Hirschfeld

39:58

you know, whatever. Right. No. It would rebuild every time it would,

Jon Aykroyd

40:01

yeah, and anytime you made a change, it would just rebuild and then go to town. So it's it's possible actually to do that makes sense. Uh huh. I Yeah,

Rob Hirschfeld

40:23

that's fascinating. I mean, I could see a time when we actually incorporated an MCP server into Digital Rebar itself so that it was helping do that, although you're there's an LLM, you have some language translation in there. Yeah.

Jon Aykroyd

40:39

No. Well, so really, all it would take is running a container. The container could have credentials or or whatever, so that it knows how to talk to DRP. And then you could just, you know, I could, you could then come in here and add the, add the IP of DRP, but you know, it would not be 8092 would be whatever the port is. You know, 8096 is our MCP server port, and that just runs in a Docker container because it's not doing any of the the AI thinking, right? It's just doing, oh, I'm gonna go execute this function. I'm gonna go process this data, right? I'm making a, you know, I'm making a curl request to some API endpoint. It's, it's an interpreter, right? I mean, ultimately, ultimately, you kind of nailed it when you said it's, it's a, it's a wrapper around an API, because it's, that's really what it is,

Rob Hirschfeld

41:55

yeah, no, it's Yeah, or it's really it, because you're doing a can be doing a bit of work. So it's an intermediate, it's an intermediary. It's or more proxy,

Jon Aykroyd

42:07

yeah, okay, a proxy, yeah, yeah, right.

Rob Hirschfeld

42:10

I'm going to ask you, you know, tell me all the tell me the machines about this, and it's going to translate that into appropriate API calls, hopefully efficient ones, otherwise you're going to be rushing your system, right? Yeah, there's the interest, there's the interesting opportunity for you to be like, or to say, oh, okay, well, isn't know how to do this.

Jon Aykroyd

42:32

Yeah, two things, right? It's, it's got a query, it's got a query the API, but then it's also got to figure out how to interpret that and hand that data back to the user, excuse me, and that data back to the user in a in a, you know, consumable way,

Rob Hirschfeld

42:54

which is also part of the translation out that you'd want it to be able to explain servers to people in a useful way.

Jon Aykroyd

43:07

Yeah, right, yeah.

Rob Hirschfeld

43:11

And would that be part of our MCP definition? You would say, oh, you know when you when you explain a server. I want you to include this information if they ask for more than go, use the params to fill in the details and offer, and then offer, if they want even more information to go fetch additional information. I mean, it would, it would strike me that that you, we need to, you need to encode that behavior. Sometimes. Just, I just, you haven't shown me yet where, where the MCP is sort of able to absorb that instruction. The stuff we've been doing so far is, like, pretty specifically coded to like, you know, get this data, get that data, use this data, make this decision. And we're using, I've been using the LM to build that, like I started working the one we did together was building that redfish simulator, redfish proxy for DRP. I started working on one that was a mass translator, so that you could make a mass call and it would translate it into DRP behind the scenes.

Jon Aykroyd

44:33

Oh, that's really cool.

Rob Hirschfeld

44:36

It started. I made some I made a little bit of progress. It was actually sort of interesting, like you can do a mass get can do them as Get, get the machines, and, you know, there are starting to get pieces and parts going. Not sure that it's useful, because I don't think people consume mass via an API, but Right, if we had a customer who was had a whole bunch of mass integrations. Yes, and this is exactly where my thought goes. It's like, all right, I'm a mass user. I want to replace, I want to put in Digital Rebar, but I have automation that's hitting my mass server, and we replace that without having to rewrite the automation.

Jon Aykroyd

45:16

Yeah, the you just, you just point at a different API endpoint, right,

Rob Hirschfeld

45:21

right? And then you would say, all right, but now we might as well rewrite the automation the way I did the proxy. I'll talk about this next week, assuming we can do this next week. I'll show you what I did, but I actually said and if somebody makes a DRP call, just forward the DRP call, because that way you could rewrite the logic to mixed mass and DRP commands, which I thought was sort of fun. It would just forward the DRP ones and let it, let people do it on their own, but the mass would be translated into drpp. I mean, I have it. I had it working, at least in a very primitive sense, from a Hey, show me the machines. And I had it setting things, and I had to honor, like mass has the concept of a lock. And so there's, like, a translation, you know, this translation stuff. And we have, you know, when you do something in a mass server, then you have to do, we have a matching thing, I it was, it was a, it's a fun little project. I'm trying to think through what, you know, projects somebody would work on that we could then basically build a, you know, prompt system that gets them down the code working, and then it becomes useful for us too, as we're building, I like, I like, what you do with the MCP, thank you.

Jon Aykroyd

46:51

Yeah, well, and I think, I think that kind of goes back to that second use case we were talking about with the the kind of domain specific knowledge and teaching, teaching the model you know, specifics about how do we code because, like, when I, you know, when I write a when I write a content bundle, I'll take a content bundle that I wrote previously and go, Okay, here's an example. Now I want to write another one that's similar to this. And then Claude goes, Oh, okay, I know what to do. If I just go, Hey, Claude, write me a content bundle. It, you know, hallucinates and just, you know, makes up shit. And it's like, well, that was not very helpful.

Rob Hirschfeld

47:51

This is, this is really good. John, thank you for sharing. Appreciate Yeah, I hadn't. I'll make one final comment, and we should close the bridge because we're over. But I've been seeing ads and like booths advertising MCP without any explanation of what an MCP was or why it was valuable. Like I heard an ad for postman, and they literally were just postman. It's great. Is the easiest way to build an MCP for your project. And I'm like, if you don't know exactly what they're talking about, they have no you know you are literally, you know the person they're advertising to is super narrow. This man's in a generic API reader, which we use, but, yeah, you can literally use postman to scrape our API, because it'll read swagger and then it'll generate an MCP server for Digital Rebar. Interesting, yeah,

Rob Hirschfeld

49:03

and then, yeah, I guess this is a place where we would post a DR a, DRP, MCP,

Jon Aykroyd

49:14

yeah. Well, there's a few, yeah, there's a few different places like here, there's a whole where is it?

Rob Hirschfeld

49:26

Like a marketplace?

Jon Aykroyd

49:28

Yeah, there's a whole marketplace that's easy to get a hold of. Um, here you go, browse MPC for servers, um, so on, on, VS code, like there's a whole marketplace that you can list. I don't know how you get listed here, but, oh, suggested here. So here's for VS code. But then you know, uh. I know Claude has one somewhere, and now Chad GPT has its agent mode. So you can configure, you can configure connectors for that too. So they all the, they seem to all use the the MCP protocol. So that's redundant.

Rob Hirschfeld

50:27

That makes sense. No, it's, yeah, that that this is, I guess, however they, however they got put together. It was smart enough to make it makes sense, to repeat it, reuse it. Cool. This is amazing. Thank you. Very, very, very helpful.

Rob Hirschfeld

50:56

Well, what we're doing here is really important in trying to understand AI, especially from an operations perspective, using digital rebar and RackN as the base and thinking through what our customers, which are data center operators, want to accomplish, is really a foundation for everything we do. And in this case, I hope you will find our vibe coding exploration helpful in that context. Thank you for listening to the cloud 2030 podcast. It is sponsored by RackN, where we are really working to build a community of people who are using and thinking about infrastructure differently, because that's what RackN does. We write software that helps put operators back in control of distributed infrastructure, really thinking about how things should be run, and building software that makes that possible. If this is interesting to you, please try out the software. We would love to get your opinion and hear how you think this could transform infrastructure more broadly, or just keep enjoying the podcast and coming to the discussions and laying out your thoughts and how you see the future unfolding. It's all part of building a better infrastructure operations community. Thank you. You.

00:0000:00