navarre.ai

ClarisTalk AI · the FileMaker AI podcast

Agents and MCP in FileMaker with Michael Wallace and Ronnie Rios

June 14, 2025 · 87 min · Matt Navarre and Cris Ippolite

Jump to transcript ↓

Cris and Matt are joined by our friend Ronnie Rios from Claris, and rising star Michael Wallace from Empowered Data Solutions. Michael does an amazing demo of a solution he just released for free that allows you to interact with a FileMaker database entirely through natural speech. You have to see it to believe it. He then shows and explains how it works, and compares it to a solution that Ronnie

Apple PodcastsSpotifyYouTubeAll episodes

Transcript

Transcribed by machine (whisper, large model), 16,310 words. Names and product terms are sometimes misheard; the audio is the record.

Claris Talk AI. It's Claris Talk AI. We have Christopher Light. Hey, Matt. Ronnie Rios, our regular AI expert, and a special guest today. We have Michael Wallace. Welcome, sir. Oh, so excited to have Michael here. Rising star in the FileMaker and AI community. I'm really happy to have Michael here. Yes, and I just met you, Michael, and I'm excited. So, you have a free utility you can download from your website and i downloaded it and i've been playing with it and i'm like crazy impressed to the point where i just went in and spoke like i was still in the 80s um wicked oh my god um i think we should see it show your screen for those of you listening um you're going to want to really watch this one on youtube uh and we'll see what this does and then we can talk about how you how you made it because i'm so impressed all right uh this is empowered agents uh named after company empowered data solutions and uh so when you come in uh you'll be on this screen and you can just click this button to start the agent hello how are you doing today i'm functioning perfectly ready to tackle any tasks or questions you have all right what's on i would like to know

Can you give me the top five companies by the total dollar amount ordered for 2024? Transferring. Yep. There you go. Wow. Let me know if you need anything else analyzed. Yeah, actually, I'd like to see this as a bar chart, please. so for the audio listeners a window just popped up it analyzed something in the database by magic we'll learn about and uh shows the numbers on the screen and now presumably we're going to see a chart yeah generate a web viewer window yeah they all must be web viewers yeah it is web viewer and i think that we interrupted it whenever it's thinking about yeah settle down matt oh was it listening to me i thought it actually listens here's you too so let's see if we can can you Give me that data as a bar chart.

Thank you. Oh, wow. Whoa. Even animated. I can mute it. I put a little button. So if I click on this, then it on the agent, then it mutes it. So it doesn't talk and it doesn't listen to me. So that's what I do in between times. Just like my wife. My wife, well, see, my wife has that for me because all she has to do is look away because she's deaf. And so she likes to look away and I'm like, oh, come on, that's not fair. You know? Oh, my God. I'm just going to, I'm done. That's it for me. Sorry about that, everybody. But sincerely, what we're looking at here, we've got a couple of FileMaker windows that were dynamically generated based on the instructions that Michael was giving to the agent.

super impressive and then the visualizations are gorgeous and animations as well. This is an agent experience in FileMaker. This is some pretty amazing stuff that we're seeing here. But it's way more than just like questioning your data. You can ask it questions about the structure of the database. Show some other sweet things that it can do if you would, sir. Give me the table of contents for the help for this file. Here is the table of contents for the help file. Let me know if you want to dive into any specific section. You didn't show it. Go ahead and show me the table of contents. Wow. There we go. So the -- and then I can ask for, you know, about specific things.

So give me more information about extending this file. That should give you a good outstanding of how to extend the system. We're seeing a Yeah, we're seeing the window pop up with all the information, which now this is information, I'm assuming you're storing internally, as metadata, or there's table information or something like that so what what breakdown for people what happened from a file maker sense if a human were to do this let's let's talk about how many steps that you just eliminated uh from you having to do to be able to do this well um for the for example for the um when you're working with doing the the charts for example um it's doing multiple multiple steps because that's not just a single agent it's actually multiple agents so i have a the primary agent I call it Empowered Core, and it has a tool that's called Call Agent.

And so it will – I delegated other things like getting the data and charting to separate agents. And so when I asked for that information – and I can do it all in one thing. If I said up front, you know, give me a bar chart with such and such information, it would make a call. It would use the tool to call the agent to get the data. that agent. And it just, it makes up its own prompt for that second agent. So it has a prompt that it creates, and it prompts a second agent with what it wants. That second agent takes that prompt and processes it, makes a SQL, uses this tool. For the data retriever, I actually have, I have like find existing SQL and execute SQL, or what I have turned on right now.

And so find the existing SQL is a table of prompts and SQL queries that will, it does a semantic, ones that I've kind of proven. And it does semantic search for the prompt, compares it to what they asked, find something that's similar and use that as a ground, you know, as a grounding point. And then makes a SQL call with the execute SQL statement. I'm passing in the Get DDL information to give it all the information about the database that I want it to have to execute the SQL. So then it makes that call, pass that information back up to the empowered core that then calls the charting agent with information that it received from the data retriever.

And the data retriever creates the chart, shows it, and lets the calling agent know that it's done. So it bounces back and forth, back and forth with tool calls and with that to get everything done that it needs to do. And even though one quick thing about it is that even though the primary agent in Powered Core uses the real-time preview, which can be kind of expensive, most of my work is done with those sub-agents. And so it's not using the primary agent as much. And I'll also limit how much it speaks because speaking is cost more than just displaying the data. Yeah. So for those that the real-time API is what Michael is using to do that really, really cool voice interface.

I can interrupt it and everything. It's called the real-time API. And it is more expensive, but Michael did a fantastic job here with kind of splitting it up. let's just use it just for the, for the user experience, all the hard work of like generating all the, doing all the tool calls and SQL and stuff. Let's do it. Let's offload that to cheaper, you know? Yeah. Yeah. Yeah. I'm sure it wouldn't be a big deal to switch it to a text input if you didn't want the voice. Right. So the real time agent is just driving the voice anyway. So you really just need that at the one that's interfacing with the user. Yeah. It's all the layers. Let's, let's back up a second though.

We can't just assume that everybody listening even knows what we're talking about when we say agents like so let's go to the very top of this and talk about how what this is you know how it's happening it's incredible what what you guys are hearing and seeing this is a you know the interface of the future it is let's talk about uh determinism how about how about we throw that uh topic out there you're gonna go back to philosophy um oh that's right that that Maybe I'm extracting a little bit too much from philosophy there. But what all of us have done in our FileMaker careers for many, many, many years is take requirements, build out those requirements in order to meet the goal.

So what we do, we hear our customers or if it's ourselves that are building something for what we're doing, we think about what the outcome needs to be. And what's the first thing that we think about doing? We plan how we can actually use the tools that we have at our disposal to solve for that. Then we'll go in and we'll actually do, in our case, we don't really call it programming in FileMaker, but we'll go in and we'll do our development. And at some point when we test everything and it looks great, then users who told us what the problem was are able to go in and begin the process and get some feedback or an outcome from the FileMaker system that we create.

So, interestingly enough, over time, we have been the agents for our customers. Now, I'm kind of using that metaphor to describe what it's like to be the third party building something. So, if you really think about the user, they're telling us what problem they want to solve. We go build this thing for them, and then it eventually shows them the answer. So that is not unlike what is happening here. Ronnie, I'll let you talk a little bit about how this is the same process, but now the agent is actually doing that for us. Yeah. So agents, there are different definitions out there. The one that I really subscribe to is that agents are a, you know, it's a system where it's leveraging large language models to dynamically, that dynamically direct their own path, the process that they're doing, leveraging the tools and resources that are available to them, working off of a loosely defined objective.

And independently, they're able to operate, plan what they're going to do, and determine when they're done. So to me, that is the definition of an agent. And there are agents of all different sizes, right? There are very small agents with little tools, a little processing power. And there are others that are massive ones where you can actually have a network of agents working as a single one. But they all have the same structure, the same principles, right? Loosely defined input, dynamically used tools and resources available to them. I'll be able to inspect the environment and what has changed, determine what path they want to take, and they're able to determine when they are done independently.

Yeah, you and I subscribe to the same definition, and it's really – Oh, you don't? Okay, Matt, what's your definition of agency? I can't remember, but I heard a podcast about it. And then we'll tell you if you're right or not. If you summarized it, well, you wouldn't know, would you? If you summarized it as basically it can operate autonomously in a complex environment with complex data. That's not the definition of an agent. Because that's fine then. The key part of... I think that's a much looser, it's a very high look down. Because I think agents, like what you're just defining, Ronnie, is very much what we just saw. But we didn't see a couple of things in Michael's demo yet.

Like the ability to go to layouts, enter data, create records, delete records, stuff like that. Well, those are just different tools that are available. Yeah, it doesn't matter. It just so happened that what his problem statements were didn't involve those CRUD actions, right? But the key part about what Ronnie was talking about is how it bookends, how it can create the plan and then reflect upon the plan. And that actually is a really important part about this because what we're seeing here today was not possible in GPT 3.5. And the reason was because there's this concept of reflection that really came out in the maturity of the frontier models, specifically in the GPT-4, which was like, what, two years ago now?

Yeah. And so what it was able to do is – and actually, you've seen this. For those of you that have worked with GPT or ChatGPT or any of the other models before, a pro tip, by the way, is when you ask it to do something and it appears that it's done something, if you simply say, okay, audit yourself or validate that you did that. that task correctly. In models prior to GPT-4, it really wouldn't be able to reflect upon what it was doing, but then it just sort of got this like compressed intelligence at that level up where now, and even now with the more mature models in 4.0, you can say, did you do that thing? And then it goes, oh, you know what? Actually, I didn't. Now that I'm looking back on it, or I can validate that it did. It's that one reflection piece that literally was the epicenter of the explosion of agents within uh llm so that's why ronnie and i are right uh when we when we come up with our definition because agents the agents really emerged and they were birthed at that moment in time in the in the uh evolution of llms now that was a couple years ago now uh i think most people in the ai space uh came into the year 2025 saying this is the year of agents now like michael what What's your take on why 2025 was the year of the agents?

I mean, we've had GPT-4-0 since back in 2023, right? Well, I mean, I think that it just kind of dawned on people with the, you know, function calling has been around, you know, a little bit longer than that. That's a huge point. You know, it just kind of, people go, wow, you know, with the function calling, it's choosing what tool to use, which means we can just take this the next step further and let it continue to do that until it gets the, you know, the response. And that's like what Matt was talking about. Like a function could be one of the CRUD operations. It could be running in FileMaker, could be running a script, which, by the way, is an important part about why this is so cool in FileMaker.

And so all those things come together. So really, I think if we take an amalgamation of what we've been talking about here, the definition of agents is being able to plan for itself and reflect upon that plan and then to keep going. That's where the autonomy comes from, Matt. So like, you know, that was a big part of it there as well, too. But then having tools available, having tools dynamically available, too. So it's kind of like, hey, agent, here's all the stuff that you can use to solve this problem and and then have at it. Here's the problem. So you go figure it out. So that's really at its core what all these crazy things are doing. And I think from a consumer standpoint, even though we're going to dig back into what you guys are doing with this one at FileMaker, for those of you that have ChatGPT or any of the different IDEs, if you've ever done a deep research, first of all, that's my favorite new AI thing that has evolved from all these different model front ends.

But if you do a deep research and you hit the little button, for example, and you hit details, you'll see that it's sitting there talking to itself. DeepSeek actually really kind of came out and they did a really verbose job of the thinking part and showing the thinking. And if you see what those things are, what it's saying, it literally will go, okay, well, the user said that they wanted to do this. So maybe I should think about this, that, and the other thing. Or maybe I should do a web search or I should probably pull in. It's like going through all the tools it has, doing its plan, and then it executes on the plan. And then later you can see it reflects on it, kind of like what we're talking about.

It's reflecting all the whole way through. It's like, I think this is this. Let me see if that fits. No, okay, I'll try another thing. Yeah, yeah, yeah. I love watching it. By the way, Ronnie, you and I were just talking about this yesterday. Gemini just released like a random tool this past week called, oh God, what is it called? Like Gemini thoughts or something. Thought details or I can't remember. Summarize it. Summarize it. There we go. Thought summaries. Yeah, it's exact thinking summaries. It's actually, it's exactly what we're talking about. You can actually extract all the thought processes. And I was just so fascinated by that because if you could take what the problem statement was and then extract the whole thing, you could actually, you know, use that as a knowledge base or train on that even.

So it's kind of cool. And so now, Michael, what's your take as to why FileMaker is so perfectly suited for this? Oh, man, if you haven't done some of the training and the certification stuff for, what is it? Hack and fix for agents and stuff like that. And you get into that and work with Python and everything. And you have to do so much work to get what we already have in FileMaker. And with FileMaker, I mean... Well, talk about that, though. Dig in, because that's a really important point here. So security and you have, you know, the access of what you want to give it and the scripts. I mean, the biggest takeaway for me for all of this is that a tool for an agent is a script in FileMaker.

So anything a script can do, an agent can now do because you give it that ability to run that script. And now if it's going to check your email or any kind of API call or interact with Claris Connect or whatever it is, all of that can be done now by an agent autonomously to decide when to do that. And that's a big deal. I love that you mentioned Connect. So just to double click on a couple of things that you said there, Michael. So first off, a tool could simply be like go search the web or send an email or create a record or whatever. So not only in FileMaker do we have the ability to create a sequence of tool use with logic and we'll appease Matt here, complex logic as well too.

You can actually then have that entire script. So that's really a sequence of tools, but you can then have the whole sequence, a script, be a tool in and of itself. Now, then you extend that even further and you said something really interesting to Claris Connect, which is a series of tools between utilities and different connectors that you could sequence as well, too. But you could treat them independently as tools or the entire flow as a tool. And then you combine that together with what FileMaker is doing. And it really is like this tool amalgamation and sequencing platform, if you kind of look at it that way. So that really brings a huge advantage.

One would almost call it a platform, right? Yeah. That makes applying AI a lot easier, right? And it's funny because, Michael, one of the things that we were talking about at Engage was exactly that, right? Using these AI tools shouldn't be hard. It should be very easy. Creating these, you did a fantastic job, by the way, looking at this code, Michael is a freaking superstar. Creating the agents, I mean, creating the agents in and of itself isn't too hard. You can create small agents that are fairly easy. but have it working efficiently under security um having the the automation having all the structure you know all the data available to it like those things we believe are really hard right to to bring during one place and i i think michael absolutely we were talking about this uh last week during uh during uh office hours as well we were saying yeah we sometimes don't know how easy we have it here right uh yeah because we're building on top of this platform it's basically you're having a, it's almost acting like a human essentially, like an assistant, you know? So you're, you're, you're asking for something like, go do this thing. And then you sit back and put your feet up and it does it like a human would do it. And it's like an assistant. I used to,

the people I ask kind of what AI is and I tell them, it's kind of like having a thousand interns, you know, they're going to, they're going to do things really fast because there's a thousand of them, but they're not going to be as smart as you are. You have to give them really specific advice. And if they get it quite a little bit wrong, you have to ask again, be more clear. This is really like that. This is really like the whole experience of using like ChatGPT 4.1, except with your FileMaker database. And that's nuts. And this is the first, this is the beginning. It only gets better, right? Yeah. This is the worst it'll ever be. Yeah, exactly. This is the worst it'll ever be, jinx.

Yeah, that's my favorite new term, by the way. And I think Sam Altman just sort of came out and said that last week about the models. And to think about AI at any given time, it is the worst that it will ever be. There's something else really interesting. Matt, I also listen to podcasts. And there's this great one that I heard some VC founders actually talking about, the way to, you know, vibe coding. If we kind of take this out of sort of agents for just a second and think about another value for the platform. is that they were talking about, okay, yeah, you can go in for like whatever the cloud version was at the time and you can vibe code out an entire web application.

Well, you're done in that AI environment. You have to now do something with that code. You have to find a server environment. You're going to need database backend. You're going to need user management. You're going to need security, scalability, and all the management of the uptime and monitoring of all that. And what they were saying in this podcast, these are a bunch of tech VCs. They said that they feel that the vibe code part of that is really just the 30% of what it really takes to go from idea into deployment. So the real heavy lift is all that infrastructure that exists. And Ronnie, you can speak better than anybody on this. I mean, that is what the platform part of the Claris platform really is.

So you take all the different tools, the fact that we can sequence tools, the fact that everybody in this industry has been using tools and sequencing tools forever, all that together with the platform to be able to deploy these things, really all you're needing then is just a user with some input. And then the user becomes the person who's interacting with the agent. It's really just amazing stuff. And yeah, I think people are probably a lot closer to being able to do this than they think they can. and the good news is uh michael and ronnie well ronnie actually hold on a second let's let's let's not leave ronnie out so anybody that went to engage saw your version of this as well too what has become of that file oh that you're gonna have me go after michael um actually yeah that you saw from me i stole from ronnie so that's not true that's not true uh but yeah we uh i uh and of course you like you've seen uh seen me work over over um oh man yeah at least for over two years right we started very very small with just that's an important part of it ronnie talk about where it started first before it even became an agent because because you were kind of creating the need for it and then all of a sudden you know real-time api became available the sort of the agent it was your your story is really

interesting because of the evolution of it take us to back to the beginning yeah well um similar to you like i i gained access to uh open ai's api very early on i signed up i was like to before they um kind of kind of to put a pre-release or something like that and started playing around with it and and saw some you know said well how do i bring this into filemaker how do how do i leverage this to bring like really amazing user experiences and so you know the obvious thing is is talking with your data and things like that and generating sql executing sql and things like that. So it started just like, you know, let me see if I can answer questions, right? If a user asks a question, generate the SQL and answer the question there. And that, that became a thing. I saw like immediately there was, there was a need for kind of create, create some structure behind it and things like that. And I remember creating, this was over two years ago, creating like my own little, um, function calling type of type of, uh, structure behind it. And then I remember like two months later, they come out with function calling and the API and everything, but this has been my little experiment you know going forward just just so i can understand what's going what our users are going to be asking for as these tools we get get better and better what's going to be like because i'm a user too so how do i how do we leverage this these amazing technology to create some some really amazing user experiences so uh over you know over time has been adding more and

more functionality function calling became available better uh better models like like 4 and 4.0 and had much more thinking processing power and things like that the concept of agents you know being realized right in like literally within the last two years and so adding more and more functionality to it being able to create uh artifacts in real time right and we saw anthropic come with you know show uh the concept of artifacts and we saw open ai as well bring canvas and all those things and so how do you bring those things into you bring those elements into your everyday application. And so that's how this little, you know, this application was my little experiment, a little lab, as you will.

And you saw it evolve, Chris, you know, over the last few years. I mean, yeah, every step of it, too. Right? Well, and there's a couple of important points that both of you have mentioned. Ronnie, I'd like you to speak to it. But first of all, function calling. We're throwing this around like people know what we're talking about. And it is such an important part of tool use, actually. Well, actually, function calling versus tool. I mean, there's some nuance in between there. Yeah, go ahead. Yeah, address that. Yeah, well, they started, you know, I remember when OpenAI first launched that functionality in their API, it was called functions, right? It was function calling.

And I think they suddenly quickly realized it was more just than functions. It's not just about calling a quick function. It's really a way to expose all the type of tooling that you might be able to run on your end. And so I think changing the name also opened the eyes to a lot of us. Like, yeah, it's tools. It's much more than just being a recaller. Yeah, because you can build tools to be much more powerful. A function is something that just credits math or something, you know? Exactly, yeah. I think the connotation itself just kind of opened up a lot of our different ways of thinking about it. Tools is a very big superset of functions. yeah and then it was funny because um one and i remember after conference a couple of uh i think was like a year and a half ago um i was i went to the conference saw an ai conference i saw some really amazing things you know everybody who's all these different companies are building these amazing tools and everything i was really really psyched getting back to the hotel room it's like oh my god and i need to build something right i you get that that urge and i started thinking well how do I apply this into FileMaker and it dawned on me how really smart these models have gotten where they were able to start creating entire workflows where they can use not only determine which is the right tool to call but also understand what the input and output of each one

of them is and be able to stitch them together into custom workflows to address the particular ask of the user right and so and i remember at that time i created um a version of this that would talk to the filemaker uh admin api and be able to like work out your admins your admin assistant and and so i'd exposed um um claris connect flows to to help me do to help me do those things so those flows became tools as well expose those and again the concept starts starts building up and you start seeing what's possible and kind of where we might be going. And so, you know, this all ended up me showing these things off at Engage and a lot of people got excited and were asking me for the file. So, I'm announcing here that I'm going to be releasing that file today. Yes. Should be hopefully by the time people hear this recording, it will be up in Clara's community. Fully open. So, you'll be able to download it, look at it.

I'll post a link in the show notes. Yeah, please post a link in the show notes. Understand, I just told you, there's some code in there that's two years old. So a lot of people have helped out cleaning up a little bit. It was a much larger task than I thought it would be, so it took me a little bit longer. But I'm hoping that with this file and with what Michael is releasing as well, I think the communities can see some examples of how they might be able to apply this to their own applications this is the way i learned to code looking at other people's code and i think this is uh i'm hoping that this helps out the community so you guys have both ruined everybody's weekend um you can get michael's file so michael tell us uh where people can get your file as well too so that uh they can book in or they can do ronnie's and yours at the same time and compare and contrast mine's been submitted to the marketplace but i haven't heard back yet so should be getting ronnie is on it i know i know a guy i know exactly uh you can't go to my website uh empowered data solutions.com and uh there's the the files on on there for you to download um and uh in fact i may just put an access to it around the front page so it's easy to find but nice um but then uh so you

can download that there. And also, I do have GitHub repos that are, if you want to look at some of the code, like the real time API and things like that. I have that in the file as well that you can go and look at that stuff that's not FileMaker, that's JavaScript and that kind of stuff. If you want to get a little more techie on that kind of stuff, but you don't need to. I highly recommend it. By the way, I used exactly that. For a while, I tried to do the implementation of real-time API. Ronnie, I remember years before that. Let's talk about, so on your journey with that, Ronnie, what before real-time API and why. So we've said that a few times, so let's back up a second here.

Yeah, how to do it, right? Because that's a web viewer thing or web direct, no, web viewer connection that using a new feature a new hook that came in 20 which version was last year yeah before real time api you could use voice uh as a matter of fact uh opening i probably all started with whisper uh they still have a uh an open source they're not just foundation models they have open source models as well too one of which is whisper which if you want to have voice control uh both both the directions have it talk and have it be able to listen and process you can use the open source whisper so it began with whisper but ronnie what was the challenge with the whisper api and also by the way for those of you that talk to your chat gpt on your phone the early stages of voice there was something kind of interesting with that that needed to be solved what was that ronnie yeah so just like like um everybody else i started with just a regular chatting a text chat and everything um and like i said whisper comes comes along and just disrupts everything you have a very high quality uh speech-to-text uh uh telling you out there which they they replace as open source i mean it's really really good um and and so you're thinking well let's see how we can leverage this and i discovered very quickly that uh i prefer it i prefer prompting or interacting with

the ai through voice and there because there's a there's big differences between how you interact with with these with these tools um typing versus versus speech when you're typing i've realized i'm thinking too much about the structure of what i'm writing right what is i'm trying to say i'm thinking about the sentence structure and things like that i don't give as as many details as i do as when i'm when i'm thinking this is natural like this is this is what humans are built on both work and so i really wanted to be able to take that and build that in so i i um they also released the api to do you know text to speech and speech to text and it's like oh let's go ahead and do that um and it's just amazing the the the user experience is just so so much better um but at the time we were still relying basically basically to walkie-talkie so you like like listen to me now yeah okay i stopped talking yeah over i'll do that all right and then but it's basically just another layer on top of your of your text chat right it's just you have have another layer that's doing the speech to text and text to speak but underneath it you're still doing the same chat type of interface it's just another layer taking care of that that voice on it and what's the problem with walkie talkie though what's what's the problem just so that

people can kind of identify the evolution there well what is the flow right it we don't talk like that like we don't really you know if you're talking about walkie talkie it is a little bit artificial this is not the way we talk we sometimes interrupt each other we you know it's that there's different flow when we're having a conversation and so you lose it you lose that there and you can lose your train of thought sometimes you just sometimes just pause in the middle in the middle of something and so it's not as natural as as a regular conversation and for a while but for a while that was that was fine then uh open ai released their open their um real-time api which just allows for some really amazing experiences i mean we start we first start experiencing it through their advanced voice interface that they've released for ChatGPT, which is really amazing. You start using it. Yeah, there you go.

This is one of the things I throw out, by the way, just not to interrupt, but typically people speak much faster than they can type. Absolutely. Yep. So that's another really big factor, just thought I'd throw that out for a second. Yeah. And we give a lot more detail, which is the other thing. So I started using, I remember I was part of the preview of the advanced voice and started testing that out and it's like wow like this is it like this is really the interface that i want to be using and so how do i bring that in um originally the the the api was required a server and everything as you know web sockets and it was a bit complicated a little bit over my head um but then they created a web um a web rtc version of that and guess what filemaker web viewer supports web rtc and so right again talking about being having the platform having every all the tools that you need in order to take advantage of these technology um and then michael did an amazing job of creating like all the code required to run this within a web viewer and i totally stole his his code and it works i've coded it so you know you didn't feel you still chat about his code yeah yeah that is it i i strongly believe like at least for me and i think for for a lot of people like this is the way i want to interact with it right i want to Interacted in a very natural, very fluid manner the way I would interact with another person.

Let it figure things out for me. I don't want to have to pause. I don't want to have to think about structure or anything like that. Let me just talk fluidly and let's figure it out together. It's a more collaborative experience. I have a question for you. When you guys are talking to the agent and you interrupt it, do you apologize for interrupting? You know, there's a, what's his name? Sam Allman was talking about that it's costing millions of dollars for people being polite to AI. He said, stop saying thank you. Stop saying thank you. Because you're like, the amount of compute power it takes just to generate those bots is crazy. But that's actually Sam's fault because he could just put a little hook in that says if someone says thank you, don't go process it.

Ignore it. Yeah. What I've done lately is when I ask questions, I no longer bother with the question mark at the end. It's just a wasted character, and I have to hold the shift key down. That's profound. Well, I mean, I gave up exclamation points. Ronnie knows this from New Year's two years ago, so I'm just trying to save the planet. No one noticed, Chris. That's because it was so quiet because I wasn't exclaiming enough. Well, this is just a – you still use them when you speak, FYI. And by the way, real-time API, because of what we're just talking, I mean, we're kind of joking about compute costs, but that is really expensive for them to be able to do.

And therefore, they pass that cost on to the customer. So what Michael did where he obfuscated the use of real-time, well, just to be clear, the real-time API is the part that's doing the talking. So you can kind of see at the top of the funnel of all the, in his multi-agent system, that the one is doing the talking. And none of them, none of the other ones need to. So you can mitigate the cost that way. Interestingly enough, let's actually look at, let's explore a little bit of the multi-agent system that you talked about there too, Michael, because that on its own- And I want to ask about cost. Is, oh, go ahead. Yeah, dig in on cost. So like, not to interrupt, but that's my job.

I get to interrupt because I'm the stupid one here. Thank you. You guys, the demo that I'm going to be releasing is like, you know, way behind this, but the stuff I'm trying to do or things that mere mortals can implement, I looked at Michael's code. I'm like, yeah, I'm going to have to drink a lot of coffee to understand this. So my question to you, Michael, is kind of expensive is one thing. But like if a user uses it fairly heavily for an hour, what you've built, what would that cost in fees? The most expensive part of it is when it's speaking. And so I've built it in a way where it produces it as text. And actually, when it produces all that formatted stuff, it's using Markdown, which they're really good at.

And I have a library that takes the markdowns in the HTML. So in the web viewer, it just shows the HTML with the H1 tags and all that kind of stuff. If you know that, just formatting. So that saves part of the time and part of the money because all that stuff that you're typing out, you don't need to say. And the other part of that is that, you know, I prompt it in a way to tell it to be concise. Don't talk a lot. It comes out to about 24 cents per minute of it speaking. So if you had it running all the time where it was just talking and talking and talking like I do, like when I'm driving, a lot of times I'll have ChatGPT going and I'm interviewing it and having it tutor me about all different kinds of stuff and AI and everything else.

So I'm constantly having conversations with it while I'm driving. Okay, tell me about that. Okay, give me more information about this. Those don't actually cost extra, though. So, if you just have your regular old $20 a month account, it's, yeah. Right. And so, for this, it's, you know, it's, I mean, I'll use it, you know, heavily while I'm doing testing and all that kind of stuff and doing a bunch of different runs with it and it may cost me $7 for the day. Okay. That's kind of what I was looking for. A whole lot of testing and then you can take that mixing the two, you can actually, with the Whisper, you can actually reduce it even further because Whisper is really cheap.

So you can put this into a text-only mode where it will listen to you, but it will generate text. And then you have Whisper do the audio for the thing. That connects to that other little thing that I showed, right? You can read much faster than you can listen. Right. But you can speak faster than you can type. So that actually sort of really maximizes what humans are designed to do. And you can think faster than you can speak, which is like the Neuralink evolution. There's a really big issue with this for people who have ADHD, Chris. What's that? We can speak, those of us with ADHD, we can speak much faster than we can think, and it's a problem. Yeah, I know. It is.

All right, I cut you off. You were asking another great question. No, actually, that is thrilling. And in all seriousness, the next thing here is brain-computer interface, which sounds space-age and crazy, But like, literally, there's technologies out there allowing people to be able to interact with computers without with just their thought. I mean, it's literally and there was FDA approval, actually, just last month, which is totally insane. So we'll have that podcast next year. We'll start thinking about it. We'll be I think we'll be thinking. Yeah, we'll be thinking about it. We'll just think about it. But I want to I want to ask Michael a bit like Michael, if somebody out there downloads your file and wants to kind of plug it into their own data.

what does that look like right I haven't dug deeper into yours to that point in the documentation that you can ask about I do have a walkthrough that just puts the copy and paste in order kind of steps you know in there you know how you have to do that but I am planning on releasing this as an add-on that you can drop into your own file and then you and so it can figure out get the ddl from that file and know what it you know what execute sql it can run that kind of stuff so um but on top of that i want to emphasize the fact that you don't have to use the real-time api to use this i i took the same code dropped into a customer's file for and for a agentic process that i'm doing for them and um in fact i'm gonna if i can show my screen one more time real quick just uh just real quick um this um i have this button here that's not on the current version but i'll be releasing manual run manual run is and i call something else i don't know but so it's a very simple script that well actually first let me make one change before i do this because um under in my agents here um i have this red by the way because it just to say hey this is

expensive you can use the so that's the model that's the model number that's being used to do that's in the real-time preview is what you know is what we're doing with that so but if I go and well you know for one with just a rather for this one's fine for the 441 so I've changed it to a regular text model here instead of the real-time model and then on this file on this button if I turn on the script debugger um i'm running the the actual tool um right here call agent that's the tool that the main agent uses to call a sub agent so it's just a simple script that takes agent name and prompt and so that so i'm using that here so great to say to run say i pass the name of the agent which You can see down here is empowered core and I have shown me a bar chart for the top five companies But I told all around the same thing prompt I gave before so I'm calling the script the master agent From a script the same way that that it would call another agent so this is all you have to do in order to take this this functionality if you're copying the scripts and stuff that you're

needed as far as you're and I'll show you here, it's the call agent and the subscripts related to it, and then different tools you want to copy in, that kind of thing. But if I close this, all I did was, in that script, was send it a thing, and it does some of the stuff you could do on the server, so it doesn't flash around while it's thinking and that kind of stuff. But either way, it runs that code, it gives me this, and I just had it give me a dialogue at the end of that script with whatever text would have come back from the result basically from the results in other words it said don it said i created the the bar chart or whatever so so i just so i just had this process this you know thing that you could run from any you know if you want to have a chart you know some kind of a text field on your layout or whatever to run this process say hey i'll give give me such such information then it would do that without you having to worry about doing all the speech stuff or worry about the price of this the cost of using the speech stuff and everything like that so this is very this what you're saying is this is very modular like you can just take certain functionalities plug that or tap into just that one thing expose it into whichever way you

want it in your application right and it'll still work because they are they are built as as agents and it's easy to call upon just by natural language. Right. So I tried to build everything very modular as far as this, like the call tools and all that kind of stuff. That is fantastic. And then here's all of our tools down here as far as the selecting a window like you talked about or navigating. And that's a perfect example of what I was asking about before. And I think that's an important thing to call out, too, is the multi-agent systems. So what really probably has been the thing that's evolved most in 2025 in the AI community are multi-agent systems.

And Michael just showed an excellent example of it. It's essentially having an agent itself be a tool to other agents. So that, you know, things get really meta at that point. But then what you end up doing is, as Michael showed, really just designating certain agents to do certain tasks. So you can kind of think of it as a team and you have specialists on the team and that the project manager of the team says, well, this is the problem. I'm going to need this agent over here, this agent, and these other tools, whatever it is. And then you guys go do this and then come back to me and let me know when you're done. I'll determine whether or not the project is completed.

If not, you'll keep doing it. If so, I'll go let the client know, so to speak, that we're done and give them the results of the whole process. So this is like having like 100 project managers, each of which has 1,000 interns and also QA people and coders and other things, right? So you get a lot of specialties that they can call. And then the QA people can say, is this actually what you wanted or not? Yeah. And then another nice thing about the tool calling and stuff like that, and it's something else I stole from Ronnie, is the validation. So, in fact, I don't want to take up too much time on this, but I think this is really important. When you're setting up tools, you have JSON schema, which is just JSON that has a definition of other JSON.

So you have JSON that says here if I have script parameters, this is what I need for you to give me in order to call this tool. So for example, this is the one for execute SQL and I have different arguments. So this is an array of arguments like what you would pass to execute SQL as the other parameters like ID for the contact or whatever. And then here's the field separator, query the row separator all the things you need for for execute sql and so um i wrote this to make it easier to um to edit these things but it has all of these different possibilities that you can use for validation saying that it needs to be this type of thing or it needs to be one of these values or something like that and or what's the default value for us if it's optional so all of that is used to validate what the what the llm sends me so that um says hey i want to run this this filemaker script with this parameter first thing i do is i pass it against uh in a in an api against this stuff here and it comes back human readable saying you're missing this parameter or your or this this parameter can only be one of these five items or whatever it is and that api

is nice because it just gives me the text back and i just go okay you want to run this is that right okay, this is what it said. Oh, I'm missing this parameter. Okay, well, here's the end. It will just pass it back and forth until it gets the right parameters to call it. Not to mention the fact that it does a much better job of calling it the first time because you're telling it, this is what you need to call this tool. So that right there is one of the secret sauces for this kind of functionality because you're telling it up front, this is what I'm expecting, And I'm validating that saying, you know, this is one of the tools you can call. And yes, this is the correct parameters.

Now I'll pass it off to run that script with this parameters. And by the way, as you notice, I named my scripts by these tools by the name of the tool is the name of the script. So all I'm doing is perform script by name and passing the parameter through. Another native FileMaker tool that makes agents even easier. Oh, because we built it for it, right? It's like-- but I love the-- yeah, the validating the JSON, that is a very important component here. And I don't know, maybe that should be something native. I don't know. I think so, man. Taking bets. It's been done in other platforms. Why bother? I think you're giving hints, though, aren't you? What else do we have today?

Each of us are going to do demos that long, right? but i think chris you do have something to show i'm not sure about if you do runny oh no ronnie has something to show runny i get it backwards but yeah well i don't want to go after michael well i'll i'll even show just the my own version of this but again you know michael's gotten a lot more sophisticated um but i i can show it anyways it's people just have an idea of what what the um please do well yeah because you're giving it away you can kind of set some expectations with folks right yeah what they're going to say i love that i was shocked to hear michael that you can and i'll interrupt running where you're setting everything that this is actually something that you thought about really implementing as opposed to just like hey this is a proof of concept that shows what can be done as opposed to like no this is something you can actually really use it's almost all black box copied in it's just going to work um uh so you don't have to play with it i love that yeah i commend you for that i'm using it with a couple my customers where yeah it's just a it's a like the main difference between what we could do what we can do now is we used to have to stop when a human made had to make a decision now we don't have to now the the llm makes a decision we keep on moving so i'll be adding this to the curriculum

for the ai classes i've been teaching for sure nice all right let me uh all right i recorded mine because uh i i know how fickle the demo gods are but i can show the application itself but um I'll go ahead and run this real quick. Hey, how are you doing today? I'm doing great, thank you. How can I assist with the app or any FileMaker-related task today? Do you know who I am? You're Ronnie Rios. I've got your preferences in mind, so I'm here to help you with anything related to the app or the FileMaker platform. Let me know what you need. It's pronounced Rios, but I'll let that pass. Got it. Ronnie Rios. Thanks for letting me know. I'll make sure to remember that. How can I help you today?

Well, I like to ask some really intelligent questions about the solution, but I honestly don't know a lot about the schema So do me a favor bring up a diagram of all of the tables of the system with the relationships between them Here's a diagram showing all the tables and how they're related. Let me know if there's anything you'd like to dive deeper into Yeah, give me a list of all of the accounts with their revenue. Here's a list of all accounts along with their revenue. Let me know if you'd like to filter or analyze this data further. Yeah, it's probably too many. Why don't we do this? Let's focus on the top five accounts by revenue. Here are the top five accounts by revenue.

Let me know if you'd like any more details or further analysis on these. Yeah, I'd like to drill down into, I think, con code. Create a chart of all of the opportunities for that customer across the stages of the deal. Here's a chart showing the opportunities for can code across different deal stages. Let me know if there's anything specific you'd like to explore further. No, I think that'll do it for today. Thank you very much. You're welcome. Feel free to reach out anytime you need more help. Have a great day. Can't be short. That was welcome. Does it cost extra money if ChatGPT says thank you to you and have a nice day? Probably. Anyway, that's really sweet.

so it looks like you've got this is a web viewer chart but the other windows look like regular old file maker lists like it did it showed you everything you saw everything you saw here are actually web viewers there's not a single everything here including the that list is just a it's really just html yeah that would make it a lot easier frankly yeah because they can return really nicely formed uh html but giving it like a set of found records and file maker is hard but yeah with stuff on the stuff michael showed they could probably do that too you can get you know you could get sequel back that gives a list of the keys and you could go you know make a global-based relationship or something and find those records absolutely absolutely yeah yeah this is cool yeah just let me uh to me it was just to make it easier again i was kind of pushing what i what i could do um but i i would i would see this more as i wouldn't be i probably wouldn't use this with only you know only web viewers i probably have a a mixer between, you know, regular FileMaker layouts. When I say regular, I mean, you know, deterministic, you know, kind of predefined layouts. And then some things here that are completely dynamic that I'm going to use at the moment, bring it up. Okay, I answered the question.

I'll keep this floating here for a while. And in conjunction with the other tools and layouts that I have, it's probably the most efficient use that I would see it. I'm trying to think of like super everyday use cases of parts of this because i think i don't think you have a lot of users regularly running reports like you did they wouldn't use they wouldn't use much of their time on any given day probably coming up with a chart like that because they're using the same ones all the time now you you probably for sure are going to have some executives that want to like look at the data and they're going to want to go way beyond that they're going to want to do things like, you know, compare the trend of products in my top 20 customers for products released between this state and this state that are in this price range. Find a relationship by category of this blah, blah, blah. You know, you can get way more complex stuff.

Well, you're saying that they run the same charts all the time, but why do they? Because that's what we've provided them. That's what they have available. No, that's because that's what their KPIs are. There's two reasons for it. One of them is what we build for them, but we build for them whatever they want us to do. If you have a KPI, this is a chart you need, but this chart here is a really simple thing to do in FileMaker. To a large degree, though, we can't just dismiss that part. To a large degree, our users are running things because that's what we deterministically built into the system for them to run. If you really gave them actual free will, release the user, let them do what they want, then, in fact, they might needs, we might get into the nuances and the edges.

I think that's probably what the answers. But even KPIs, they are short lived. So your KPIs might be just for this quarter, even for this month or whatever, right? And so you might have charts that you're following this week because like, this is what's important. We've got to make sure that the inventory is at this level. We've got to make sure that utilization is at this level and we're keeping track of that. But next month, I really don't care. I'm going to go, that's going to go away. My focus is going to be something else. So the business changes really, really quickly, right? And the pace of change is just accelerating so it we so far we users are constrained to the things as developers that we build because it is deterministic uh that does no longer you know the shackles have been have been taken off it yeah that is not necessarily true anymore can you stop sharing your screen we can see each other's faces better um you put on something earlier that i really loved uh and it's something i've been playing with too and that is uh you've got you've saved a list of prompts, if you will, that are known to work. Right. And so this is, this is a use case that I think is more realistic, maybe where you can give the user like these reports that we know work, and then you can throw that report with any set of data that you want. And you know, the result is going to be right because some of the results, some of the queries that are going to give you are going to be wrong. You'll ask it for something and it won't give you the correct

data and you can't afford that yeah um and that happens much less in the 4.1 but it was a it was kind of an issue with previous models uh from what i from the tests that i've been doing um and that's that's a reason to not use this for sure and uh until it gets better to do that um but with the agent process you can have it double check you know um but and also by the way it's not all about just visualization this what this isn't a charting tool or a report tool there's all the crud operations in there too so you could literally just say you know go find these and update you know the statuses or something like that too just it's kind of like walking into a grocery store and seeing 400 types of toothpaste it's like i have no idea which one to pick i don't even i don't even see the ones that i've bought before still here you can't use that's not useful to users you want to go in and you only want to see like 50 yeah all the user needs to know is the outcome that they desire. And then they just express that to the agent. And if you really think about this, I mean, it's the same thing with like providing, you know, an agent interface for any experience. And that is that we don't have to go learn an application, learn the steps, learn the menu items and all the things. You just say what you're trying to do. And then the agent goes and figures all that stuff out. The point I'm trying to make though, is that you can't give the user

in a lot of these situations, unlimited choice. You have to actually direct it to say, here's the things that we know you're likely to do and you can customize any of them and you can add anything to it. Well, there's only certain data sources or applications that they have access to. And then the actions that are available or the tools that are available within those applications are limited as well. Their permissions are certainly limited. And by the way, let's double click on, but that's what I'm talking about. So let's just double click on permissions for a second because I think there is a widely misunderstood, a grossly misunderstood concept that somehow agents are bypassing security.

Ronnie, please speak to that. Yeah, absolutely not. I don't know why people think everyone's so reckless doing this. Yeah. Just as Michael said, all these tools are just, they're running scripts. These are regular FileMaker scripts. And so that means that they're still in the current user session, in the current user session. And so they are running with the exact same privileges that the current user has. And so if that script is set up to run as that user, then that's all the permission to have. If you set up that script to run as full access, then that's a whole other thing. That's on you. That's on you. But as a developer, if you've implemented the FileMaker security, following all the guidelines that we've been talking about for decades, then the AI has no more access than the current user has.

Has absolutely none more. It is only able to access the data that the user has access to. It can only run the scripts that the user has access to with those exact same permissions. It cannot bypass and do things that the user can't do. I never even heard or would have thought that an agent would have more access than a user would have. It is a widely held misconception. And by the way, also the other applications that you have that the agent has access to, not just FileMaker, because you're doing API calls that require authentication in those cases. Michael, sorry, go ahead. There is a caveat to that in that there's a lot of developers out there that use security through obscurity, you know, where they can't see that layout, and so they can't see that data.

But they're not actually using the actual FileMaker security to limit what they can see. So now if you use the tool, I have a tool in there, the Data API tool that you could use instead of XQSQL, And then it could only go to the layout that you have, you know, given them permission for. Again, permission not in your navigation script. So it's like... Well, none of those users listen to this podcast, so I think we're safe. So just want to make sure that you're actually locking it down because that is a requirement. Absolutely. Yeah, if you're doing it wrong, then the agent's going to do it wrong. Like you're just basically, you're just exposing the same risks that you did to humans.

if i write that way test test test the other thing you got to do i'll make another point having really played with this and taught it i'm not going to build regular filemaker reports anymore i'm done same i'm not going to make i'm not going to go in and make a layout that's a list view that has sub summaries and stuff like that because those are the most deterministic things that we do that's not why i create those that is why that is exactly why it can be your but it's not mine chris but you you no longer have to do that those are actually restriction with old old programming paradigms so we don't need to do those anymore it's actually frankly i mean i think there's there's going to be kpis that people are going to want forever oh you're kpis yeah i hate that term there's going to be things people are going to want to know about the database forever but you're really what you're zeroing in on that is a good point and that is that there's there are some outcomes that are rigid uh there are some like as a matter of fact If you built an agent system that just allowed a person at work to be able to express themselves and do whatever they want, ultimately what you'll end up finding out is that person probably does like 15 things over and over again all year long.

And then they're at certain different intervals and timeframes. So those are rigid. Those are those things they need to do are those outcomes are going to be rigid. But like, let's say they're just walking into a meeting. It's a one off or whatever they need. They're going to be able to express what they need in that one ephemeral moment. And it's going to be able to react. then that just kind of disappears and goes away so there are also times when that occurs as well so in the same way we all use chat gpt right like michael you were talking about it i used to learn greek i use it you and i chris learn it to use it for our pets to keep information uh ronnie uses it for all of his like collections of various star wars toys and coffee recipes yeah coffee recipes yeah uh that's a whole other it's a whole other episode but yeah you have to get he has a he has a filter machine that has 400 steps because he has to get that exact right grain, and he keeps all of his coffee in a room with a precise humidity of 46.2471%.

You and I were talking about it earlier. That's a bit off, but yeah. Measuring total dissolved solids and extraction. What's the coffee you were drinking earlier? Do you have that still? Yeah, it's boiling over there. I'm almost done, but this was an espresso tonic with a little bit of lemon honey double shot of espresso and we're taking an aside here espresso tonic tonic water and espresso yeah i've heard about it is it is fantastic it's a great summer drink um lemon honey and rose so it's lemon honey double shot of espresso and some tonic water and it's just really fantastic i have some fever tree tonic or is that fever tree tonic uh this one is actually Yeah, so there's a fever tree.

That's the good stuff, man. That's really good. So we've reached the, but wait, there's more part of the podcast episode here. So one of the things that we talked about was tools. Really, I mean, the lifeblood of aging capabilities are what tools we make available to them, right? Like they don't have, they can't just go out and do whatever they want. We have to provide what's possible to them. One of the challenges, though, with tools and really data sources as tools to be able to provide context to these agents is that all the different contexts could be different. You know, like you could have like a SQL environment or you could have a FileMaker environment.

And they sort of format or provide their data in different ways. And then you're doing a lot of programming if you have, let's say you have 20 different data sources and they're all disparate systems. you would have to do a lot of programming to be able to communicate with those. And so one of the more fascinating things, Ronnie, that has come out in really not even it's not even six months at this point is a protocol to unify that. Can you talk a little bit about what that is? And then I think Michael's got a surprise for us on that. Yeah, so Anthropica, a little while back, literally hasn't been a year, released a proposal of a way to kind of unify the way or kind of everybody get to expose these resources and tools in the same way, right?

So if we all speak the same language, interoperability will be a lot better and move forward to it. And they call it MCP, which is short for model control protocol. And basically, it's just a way of how do we expose these tools to AI in the same way? And so if we all speak the same language, then, right, they make this a lot easier. And, you know, I remember when they first came out, it was, you know, I've been keeping a close eye to development with it. but it was kind of a chicken and the egg type of thing where it's fantastic, but the value really is if everybody uses it. And of course, you and I were talking about, I think the pivoting moment, right, was when OpenAI said that they were supporting it.

And I think it's been a huge flood, right? And that was just the week of Engage, just to let people know. Yeah, I know. This short while ago, right? But it's been an explosion after that and it's becoming the de facto way of how tools are exposed to AI to the point where even Google announced that with their release of another protocol that they're going to be using MCP as the backend, as the kind of building block for it as well. So I think the industry is colorlessing into exposing these tools to AI through this. Yeah, the timeline there is fascinating. It was actually November. Yeah, it was in November. It was in 2024 that Anthropic came out. They literally just blog post and said, like, this is something we probably need.

The problem that I stated before is the problem they were stating that they were trying to solve. They came out with MCP after three or so months. The moment, that epicenter or that explosion moment that Ronnie was just talking about was when actually OpenAI came out and said, look, we were working on our own protocol as well. But we're going to acquiesce to the MCP approach and use that model and use model context protocol for everything. And then the whole industry was like, oh, well, now we don't have to create our own or wait for everybody to create their own. Let's actually just adopt MCP. And now it's everywhere. And actually, for those that are for some reason still listening to this part of the podcast, but are not compelled by how important having a protocol is.

Believe it or not, you are experiencing a context, a CP of your own, and you don't even know how important it is. That was when TCP actually came out. TCP, unifying a protocol, is what made networking computers even possible back in the day for all the kids. Well, there's no kids listening. You know, all the geezers listening. You remember when that happened. Yeah, exactly. When you said MCP, I'm like, Tron. Is there a Tron reference? Oh, no. That's like the master thing at the end, I think. Anyway. So then now the challenge is, though, so Ronnie, if I want to connect to an MCP server. So MCP, much like in TCP, there is a client and server relationship in these exchanges.

So there's MCP servers and then there's MCP clients. So, for example, could our FileMaker systems be an MCP server providing data to a client? Absolutely. It can also be an MCP client at the same time. But as a server, right, we can expose all the, as resources, we can expose the tables, right, the entire databases. We can expose tools, scripts as tools, or even custom functions as tools. So it's basically, if you think about it, it's an API for AI, right? It's how to expose resources, how to expose things to AI so that it can use in the best way it believes it should use them to accomplish a particular task. And obviously, we have a ton of data and resources and business logic already in our applications.

It makes a lot of sense to expose those that way. And what Michael has done, by the way, you don't need MCP. MCP really is great once you start splitting up the intelligence and the resources in two different places. How do you make those two things talk to one another? What Michael's built is everything is in one place, so he may not need MCP in there. But if you want to expose those things to the outside world, to another intelligence or to another resource, they need to speak the same language. And that's where MCP comes into play. Michael did the exact same thing, but internally. He's got his own way of talking, of exposing those scripts, for example, to the large language model.

In addition to providing the protocol and the uniformity of the exchange, there's one really interesting thing that's also happening with MCP. And that is that we're able to describe to the agent why this source or why these tools or these actions, whatever it is, are important and relevant to the actions that they might be providing. And that's a big deal. I can say that in some of my own early explorations where I was just like giving data to, you know, I'd like copy and paste data into ChatDBT and see what it would come up with. It was left to its own devices to try to interpret the meaning of the data, for example, by the column names or the index of the data or evaluated data.

And it would kind of have to guess, right? Much in the same way that much in the same way that we like we leave it if you leave like a lot of people perceive hallucinations as it just is wrong. It doesn't have the right intelligence. But the reality of it is it has multiple different answers that all could be correct. It's whether or not it fits your question. And so the way that you can dial in the aperture a little bit is by saying, OK, this is actually what this source of information is for. So if you ever run into a question about finances, this is where you're going to find those answers, that type of thing. So providing that context to the context is an efficiency, a piece of this that actually allows those exchanges to be a lot more efficient.

Now, the problem is you have to spin up the MCP server. Michael, is there an easier way to do this than having to manually create these MCP servers when you want to connect to these sources? Which we don't even know how to do. Yeah, let me, so what I did was I created two different versions of an MCP server than when I say I, you know, the coding agent and I, so, but, so for example, I made this, this extra agent here that all I did, I didn't give it any information here. I just gave it a couple of tools here. Then I have a Python script running here, which this you could, I'm planning on putting this sentence out, so it'd be easy to use with a container where you can just, it'll be containerized and you can just run this on your server, and then I also have a local version that you could spin up and run locally because Claude will use either one.

By the way, Claude, in this example, just so we're bringing it all together is the mcp client in this case and you're talking about creating an mcp server actually you know it's funny because it's the reverse and i had the hardest time to wrap my head around that for that reason but they call the the llm that is using the service the host and then the case and then the and then like file maker in this would be the client and then what's connecting the two is the server um and i had to bang my head against that But if I look here, I can see this FileMaker MCP, and I can see these same tools here that I see over here. And in fact, if I were to close Claude, and I'm going to change this so that I don't have to stop the server in the future.

But right now, I just stopped the server here. I'm going to add two more tools here. I'm going to leave the layout so I make sure I committed that. And then I'm going to just run that again. And it's going to go through and say that it's running again. And then if the demo guides are willing, then when I open Clod up again and I look to see, and there we go, we have five tools now. So we're providing all this functionality. For example, if I were to say, I'll use the same one, give me the top 10 products that were sold in 2024 by dollar amount and how much were sold the previous year in a bar chart. so I don't like typing either so I use super whisper to so this is running these tools they got it got the layouts and it's looking at the at the products layout but again this is not what it needs to do because you know what it and

so it's looking at these different things but it doesn't really need to do that what it needs to do is a sequel and get the information there but it doesn't know that so now it's look but it says that okay I have this information I'm I'm going to write a SQL query, and so now it has the database, now it's going to start writing the HTML and stuff for that process, this is all live, it's just doing it on its own, I have no idea what it's going to do, because I've decided to change the prompt here, but one thing while this is writing this out, that I wanted to bring attention to is that this is writing out all the code that you could take and put into a web viewer and file maker yourself so in other words you could take all this code and then set it up so that it substitutes out your current thing so you can make a new tool that would do a new type of of display or whatever by just having claude write the code for you and then here you go here's our here's of sales and give us some other stuff here um so uh so the drill press did a whole lot better than the previous year etc etc so um so yeah so this all is just by giving uh clod which is just

like open ai desktop app this is the clod desktop app and because i connected it to filemaker with that then it has access to all of those tools whatever tools i give it now this is running on the server because using the data api so it's it's making a call to the server to do things so without special um extra things which i've thought about how we might want to do you can't run it um on it's not getting client information in other words like who's currently signed in and and what layout are they on and all that kind of information like that user context the user context that you could get other permissions right exactly but you have in this case exposed your FileMaker data as an MCP server and then allowed, in this case, you're using Claude's IDE, the desktop client, to be the client to connect to that.

You're saying, though, that it doesn't carry with it the advantage that FileMaker has to be able to know who the user is and what permissions they have to certain data? Not off the cuff. I actually thought about adding in where you add your own credentials to the setup, And in that case, it would log in to the Data API with your credentials, and it would respect those. So for just to expose data, you know, the easy way is to expose data that everybody can see. Because that's a hard problem to solve. Yeah, but also to bring it into like a FileMaker metaphor, it's not too different than running a script within FileMaker versus running a script on the server where it doesn't know which user with all that context and permissions that it might need.

So it would be similar resolution at the same time. Yeah. I have one big question for you, Michael. What's with the second most high-selling product, Demolition Hammers? What are people using those for? It's completely random stuff that I had. It's a library that I coded something that would generate through bytecoding, generate a bunch of data. I've done that too. I just, no press I get. But Demolition Hammers, come on. I love sample data a lot. So we've really touched on a couple of, I'd say probably the powerhouse features of 2025. Here we are halfway through 2025, and it's the year of the agent. So we've talked about agents, what they are, and the multi-agent systems, and then also now MCP.

These are really the big topics of 2025. And I think the most important part here is that we're talking about these in the context of FileMaker and talking about all the things that you don't need to go out and source to be able to pull these off because you already have them in FileMaker. And now FileMaker is firmly in that conversation about some of the most important and relevant things within AI discussions and FileMaker can operate there. So super exciting stuff. And now, and Ronnie, you're going to have to really get that thing going because Matt puts these podcasts out very quickly. So you said by the time this podcast comes out, well, by the time it comes out, it's going to be in about an hour, actually. So you're going to have to clean that thing up.

I've been secretly messaging the person who's going to help me put it on the file. Two hours. So there you go. We're recording on the 6th of June at 6.30 PM my time. Oh God, now he's time stamping too. don't give up all our secrets but uh but yeah i think and i think really so as far as uh plugs are concerned i just want to give you guys a chance to talk again we'll try to we'll try to our best to get the links here in the show notes but uh we'll start with ronnie ronnie where can people find this specifically and what are some of the things that might be coming uh that you're working on the labs too that we might be talking about on future episodes um so yeah i'm posting this since this was part of uh the engage my engage sessions um i'm gonna be posting it there so if you go to community.clarice.com go into the engage uh section there for my session that will be part of the content there so it'll be posted in there and a couple of people have been asking for it so i'll be posting it there are those out yet um yes i mean the session the session entries there in community are absolutely there and so presenters can been um sharing their content in there i'm just late to this because it took me a bit longer but something else i'll um i am going to i'm working on i had kind of teased at at our training day ai session i've got this little

application that basically is an agent agentic playground so it's a place where people could go in and create agents manage different accounts put into some instructions for for the agents expose some tools just to play around and see you know kind of just work play with the with the concept of agents just in a small place play around with those things um use look at the script something a lot smaller because you know what what michael and i are exposing has a lot has a lot of moving parts but we can just compress this down into kind of this the smallest portion that maybe people can start learning around how these how these uh how these things work together so i'll be um i want to uh release that um agent play uh into playground and maybe next week or so and start playing with that yeah awesome um but that's super exciting and then michael uh again people can get this on your website but also your uh ronnie's going to help accelerate getting this onto market claire's marketplace too if i'm not mistaken you said that earlier yes and i did put the uh link in the show notes it goes directly to the page that where they can download it uh if you can add that to the video and then uh and i will i'm going to be releasing a an update today um within the next hour to uh have the extra uh mcp stuff in

there uh as well sweet so and uh matt what about you you got some you got some classes coming up here you just were talking about um the the ai classes yeah i did a full day i class in toronto at the conference with alexis allen oh yeah that's right a week ago that's right yeah and that was great we didn't get this deep in but it was very hands-on and everybody left with basically some ai implemented in their solution which i love i hear your class was the most popular um well it's not like i was up against whim to court or anything you were up against them to court wow we tied nice job damn dude anyway well hey you know you know well what's next you brought it up last um i've got the advanced class coming up it's actually advanced cloud maker plus a lot of ai uh starting in a couple weeks i think um i can't don't you have some stuff uh clary's can or there's some some collaboration there's some free classes coming in fact there's going to be two free ai classes i'll be doing one for reporting and one for like data summarization nice uh and then i'm not really sure you know i i think that and maybe you and i have talked about this it seems like the world is changing from zoom training classes to in-person

training classes oh i hope so i miss in person very much makes me sad that well actually i'm doing a couple in-person private trainings in two weeks what am i saying so yeah it's uh nothing like getting in front of people in the classroom and really leveraging that dialogue so yeah um yeah i hope to see i hope to see more of that happen and then um matt of course do you want to give us an update on your um on the application for your uh that you're working on for your your funding your greek funding that you uh we were waiting on you for i'm sorry you're having way at the end of the podcast yeah so going through the process of what we started before you started the thing here of getting a building permit to build a house in greece is something i do not recommend i strongly recommend against building a house in greece i i will heed your advice where there's enough water behind us now that at least that pain is gone and we only have a future pain coming up um uh but yeah it's it's a thing but we're you know the thing i like to say uh is while you're waiting for things to happen in greece you get to live in greece so there's that huge benefit well put well uh and i love it here so yeah we we this is fun to be able to visit

with you this is just for so we're clear this is the uh international clarice talk ai uh podcast uh every episode is international just so everybody knows here so this is true this is uh definitely an exciting one uh absolutely appreciate ronnie always great to have you back these are the most popular episodes whenever you're here so thanks for gracing us with the extra clicks so happy this is awesome this is awesome anyone was listening in their ears on this one you can absolutely go back and check out the youtube version there's so much cool stuff to see and michael did a great job of taking us through his code so uh you might want to give it another listen uh but really important topics here today um as always uh certainly appreciate the opportunity to talk to people about ai uh and matt i'll throw it back to you to close us out i gotta say it this has been so great for me to be able to talk to geniuses like you guys and it's really it's really empowered like my mission to learn more about it i cut you off again um and i just want to say thank you guys this is great you have uh and watching your evolution as a trainer i mean i don't know if anybody else is really consistently doing bringing ai to file maker people at the entry level in particular uh you're you're moving the needle dude like keep

doing it uh really appreciate that very much i hear feedback on your courses and so this is all part of the journey so um yeah this is a lot of fun um thanks for showing up and uh it's just if It's just Matt and I, it's like an old married couple. So it's great to have other people in the room to keep us straight. So who do you want to say goodbye to at the very end here, Matt? It's becoming kind of a trope. Alexi Folger, we love you. Bye, Alexi. Thanks, everybody. Cheers. Thanks. Take care.

Back to top ↑

Have something to share?

Join me for a conversation on ClarisTalk AI.

Schedule a recording →