Ernest Koe: 00:14

Alright. Welcome back to another episode of The Context Podcast. And today, I've got Todd and Kyle. And we have a bunch of things we wanna talk about today. Let's see.

Ernest Koe: 00:24

What do we have? Some topics. I think we wanna talk about some features that may be coming on Ottomatic. Right, Kyle? I think there's some cool AI things

Kyle Duval: 00:32

and then some servers. Some existing features, some old features.

Ernest Koe: 00:38

And then some news in AI. I think there's been a lot of talk about slop cannon recently. We have some thoughts about that. My favorite topic is SaaS dead? Maybe.

Ernest Koe: 00:51

Maybe not. We shall we shall see. And what else, Todd? What's on what's on your mind?

Todd Geist: 00:57

Well, there was a bunch of new model drops this in the last couple weeks. We have

Ernest Koe: 01:02

Oh, yes.

Todd Geist: 01:03

Opus four seven. We have ChatGPT five five, which is, I guess, out in the API. And also DeepSeek v four, which is making waves. Just came out, I think, today with Frontier Model Performance that's much smaller, yeah, lots of good stuff happening. The other other thing I think is interesting is Claude or Anthropic released Claude Design, which is, you know, we we're gonna talk about whether SaaS is dead or not, but and we don't maybe spoiler, we don't think it is, but Claude is doing a pretty good job of slaughtering startups in their infancy here with basically releasing, you know, features that startups are shipping with all over the place.

Todd Geist: 01:53

So why don't we dive in? Where do we wanna start? Wanna start with Slop Cannon? Or

Ernest Koe: 01:59

Let's yeah. Let's start with some Slop Cannon talk. How about how about

Todd Geist: 02:03

that? Yeah.

Ernest Koe: 02:04

So yeah. Go for it.

Todd Geist: 02:06

So well, I gotta find there's a guy another another person you should follow in this space talked about slop canon this this week. I think his nickname is Moe, we'll include a link to the videos that he does on AI, and they're fantastic. Really funny and really important to watch actually just to get an understanding of sort of what AIs are good at and what they're not good at and what you should be very careful about. And he's in this video, he's saying embrace the slop, which is true, but what that means really is like you have to know it's also slop. Like like, there's a there's a place and a purpose for what these AI models can do, but they also do just sling slop.

Todd Geist: 02:51

And you get things like what happened to me this week is sometime earlier in the week, think on Monday, '1 of my AI agents decided that in order to deal with these pesky linters rules that I set up, Instead of fixing those rules, it should just basically relax all the rules.

Ernest Koe: 03:15

It's gonna fix the linter.

Kyle Duval: 03:16

We don't need those rules.

Todd Geist: 03:17

We don't. Yeah. It just said, like, ah, you know, I I didn't catch this, so it got committed. And this week I am dealing with, like on Friday I realized that this happened and none of my rules are being enforced. So luckily this isn't like functionality that was broken, but it was there's now something like a thousand linter errors that I have to deal with because I did also a major refactoring and so you know these AI agents will happily disable tests to turn off lint rules and if you're not catching it, end up with basically a week's worth of slop that just got fired out of a cannon into your code base, and like if you catch it early enough, it's not a problem, at least like I figured out what it was and I could go back and fix it, but I think this is this gets to the point that I think the industry is dealing with right now.

Todd Geist: 04:15

I'm seeing this in other other people on on the various socials, coming to the to the same realization that there are things that AI is really good at, and it has definitely changed the landscape of software development, no question that we're living in a new world, but I think the idea that a person with no engineering experience is going to be able to maintain a complex code base that is running a business that has a lot of like, a lot of revenue being generated

Ernest Koe: 04:47

And risk.

Todd Geist: 04:48

Without engineering expertise, not happening. Not gonna work. That's that's my my hot take for the week.

Ernest Koe: 04:57

I I'm curious, Kyle. I mean, I I've seen this happen to me, but the things that I commit to are certainly far less risky than the stuff that you're committing to with Automatic and we've paying customers, the threshold and risk tolerance changes quite dramatically, right? But I've chuckled when at times and this was with like Opus 4.6 even. It would do something and tests keep failing, you can't figure it out. They say, okay, I'm just going to rewrite the I'm just going to remove the test and we're going to ignore it and test pass.

Ernest Koe: 05:33

Everything's great. You know? It's just keeps going. The test passed.

Kyle Duval: 05:36

I removed the failing ones.

Ernest Koe: 05:37

I removed the failing ones. Exactly. Yeah. So I'm curious. Are you dealing with it?

Kyle Duval: 05:46

With things like AutoFMS, a lot of the testing is actually not done by AI. So, a lot of the testing for AutoFMS is a is a it's an automatic process, but it's a very long running automatic process.

Ernest Koe: 06:01

Right.

Kyle Duval: 06:02

So I think the full test suite for auto FMS takes like a half hour or something. And those tests, a lot of them are written with the assistance of AI, but very few of them are run by AI. So when they fail, I can say, usually, it's I go look at the test and I say, what's going on? And I fix it. Or I say, oh, this is because I'm working on a new server and I forgot to set up the additional database folder, you know, which failed 40 of my tests last week.

Kyle Duval: 06:35

And so there's a lot there's much more human intervention in these kind of high risk applications. So especially for automatic and especially for auto FMS, the changes I make are very small normally. So, know, it's I change something in the data API proxy that's like a three line code change. I, you know, I usually, it's hey, say, I wanna do this thing. Claude walks walks through what we need to do, figures out what's happening, says or changing this.

Kyle Duval: 07:07

And I look at it and I say, that's a horrible idea. We shouldn't change that. That's gonna break everybody's data API proxy. You know? So What's your daily driver, Kyle?

Kyle Duval: 07:17

Are you I use Cloud Code.

Ernest Koe: 07:19

Are you using Cloud Code? Yes. So that's that's what I'm doing. Yeah. It's just on on the topic of Cloud Code, and this is something I think is worth talking about in terms of AI news.

Ernest Koe: 07:28

Claude 4.7 well, I mean, recently, it really did feel like there was some kind of regression. And I think they made some announcements on

Kyle Duval: 07:35

day. They went

Todd Geist: 07:36

they came clean and admitted there was. Yeah. Some they had to do a pretty heavy investigation to figure it out.

Ernest Koe: 07:41

But It got dumber for sure. Yeah. You know, it wholesale forgot, like, what it was working on it on at times. And they're like, what? I just we just talked about this, like, two minutes ago.

Ernest Koe: 07:52

Yeah. The little brother, Opus 4.6, would have totally gotten this right.

Kyle Duval: 07:56

I've actually like, I think this week, I haven't used well, this week, I've been doing a bunch of secret rotation because of the Vercel breach. But the the when I've been doing Claude stuff, I've almost exclusively been using Sonnet four six Yeah. Because Opus four seven. Well, a lot of the times the work I do, like, you know, I'm not touching 600 files at once. I don't need something like Opus.

Kyle Duval: 08:21

I I think that's that's a thing to be clear about too is that a lot of the times I think people use Cloud and they're like, oh, I need Opus for this. And it's like, no, you're changing three files. Use Sonnet. Right. It's gonna be fine.

Kyle Duval: 08:30

But so I I've been using Sonnet, so I haven't seen it as much. But definitely, last week, I touched on I used to Ottoman four seven, and it didn't it didn't feel quite as good. I agree.

Todd Geist: 08:41

No. I I I think it was a regression. In fact, I'm certain that Ottoman four seven is the one that turned off all of my all of my gates. I've noticed it being very obstinate too, like very certain audits.

Ernest Koe: 08:55

Yeah. It's very High conviction.

Todd Geist: 08:58

Yeah. High conviction. Which sometimes, I mean, I like a model that pushes back. I mean, I don't want a super sycophantic model, but it definitely felt like a regression. I think they've made some changes and maybe it'll get better, but currently if I'm using the anthropic models, I'm sticking with four six.

Ernest Koe: 09:19

Were you were you seeing the regression with the pure model, or were were you using Claude code? Because So

Todd Geist: 09:24

I'm using Cursor, actually.

Ernest Koe: 09:27

Oh, interesting.

Todd Geist: 09:28

And four seven a lot this week, and I really think it's what caused a lot of these problems that I'm having now.

Ernest Koe: 09:33

Because Anthropic is is saying that it's a hardness problem and not a model problem. I'm not Well, the model

Todd Geist: 09:40

is different, and for sure, like, has a different feel. It feels much more certain.

Ernest Koe: 09:45

The model is different. Yeah. For sure.

Todd Geist: 09:48

And I found it to be also be slow. Yeah. Like, was really surprised the amount of how long it was taking to get things done. I think here's, you know, here's something to say about this stuff. Like, you cannot give up control, and yet there's so much of how human nature and what our brains are wired to do that just kind of makes you just want to keep pulling the lever, know, and like you really have to think, am I using, should I pick a different model here?

Todd Geist: 10:22

Like this doesn't feel right. Like you have to learn to trust those intuitions and switch and do something else. That's one of the reasons why I still use cursor is I can actually switch to use like GPT-five 0.4 or 5.3 codecs, which are very good models, so I can get some different opinions on that stuff and it definitely feels like it's worth doing. So this stuff is still, you really have to, even despite the like everything in your human nature saying, just just, you know, pull the lever one more time, it'll fix itself. Like, you really need to insert yourself into these things still.

Todd Geist: 11:03

I think that's just true. Just the way it is. So, yeah, like, the model, choosing the right model, like, still having still, like I think we should also we should talk about, like, why SaaS isn't dead also because Vercel breach and, like, the number of breaches that are happening, the number of things, the number of security issues that are that are coming due to AI making the hackers better just means that every single code base that you maintain is something that you're going to have to at some point do stop whatever you're doing and go fix it because there's a breach. And for a company like us, this is gonna become part of what we just have to do. Right?

Todd Geist: 11:49

Like, as a professional software development company, we are going to have to have systems and protocols and runbooks in place that say when something happens for all the repos that we're managing, all the projects we're managing, we now have to stop whatever we're doing and deal with that. And like, again, like if you're not a software engineer, if this isn't part of your practice, I'm not trying to I hope I this doesn't come across as I'm trying to scare people. I want people to realize that yes, you can make prototypes, and yes, you can make simple apps without a lot of experience, and you should, like you should be experimenting, you should be playing with these things, you should be learning with these things, but if you you think it's a good idea to ship into an app that has significant risk associated with it and you're not a software engineer and you didn't set up the test, didn't go through any of this stuff, realize that this is the world you're stepping into. It is not safe. Yeah.

Todd Geist: 12:58

You should know that.

Ernest Koe: 12:59

I I wonder if the framing is also something like this. I mean, in theory, you could with copious amount of time and some copium, you know, work with Claude or work with Codex and get upskilled and become a security expert and become a threat analyst and become a infrastructure maintainer, become all these things that you need to to to deploy and put software in a place where it's actually working for you and your team consistently. The problem is that that may be true, but it becomes another job you have to do or multiple jobs you have to do. And this is the thing that I keep coming to lately, which is the job of the the thing of making software is is not one job. It's not solely the business of like building the thing.

Ernest Koe: 13:58

It's maintaining the thing. It's fixing the thing. It's dealing with bug fixes, dealing with support, dealing with new features, dealing with feature requests, dealing with all of that stuff. It really becomes, the mountain of work that we often take for granted doesn't actually go away. So I think we have to ask ourselves a really hard question when we say, I can build this too.

Ernest Koe: 14:23

Because what you're really saying is I'm gonna take responsibility for maintaining it, owning it, making sure it works, keeping it up, doing all that stuff as well. Right? Which leads me to think that I'm not sure SaaS is dead because I have a lot of things that I know I can build, but I don't want to have anything to do with keeping them around or alive or whatever, you know?

Todd Geist: 14:46

Yeah. I think if you're in the business of software development today, like if you're a FileMaker developer and you're entering this space or if you're any kind of developer and just starting to get into AI, you just said about learning all these things is actually really important. You need to these tools can help you get better at this and you need to you need to learn it because that's where the edges are gonna be. That's where you're gonna

Ernest Koe: 15:12

Right.

Todd Geist: 15:12

That's where you're gonna potentially cause real problems. Like, this Vercel breach that happened this week, essentially, a bunch of secrets got exfiltrated by a hacker. Not all of Vercel's customers were impacted, but a number were, and in fact, they're not even sure they found all of them yet. They're continuing to work through that, But those secrets give access to, well, whatever they did, you know, API keys to whatever service they were connecting Full

Kyle Duval: 15:44

access to databases or Yeah.

Todd Geist: 15:47

Full access to databases. Right. Like, this is serious stuff. This is not like, you know, no big deal. This is serious, and you have to be able to respond to that.

Todd Geist: 15:57

And if you don't understand if you don't understand things like how do I what are the difference between sensitive secrets and just config? Like a lot of that time in, you know, if you're talking to Claude, it's going to equate the public site URL, is essentially config, like that site lives at this URL with a QuickBooks auth token or something like that. Those will get managed in the same way by your code. They're gonna be in some way an environmental variable, but the threat the threat and the risk of each of those things is dramatically different. One is no risk, the other has like risk to your bank account risk.

Todd Geist: 16:43

So you have to understand the differences and how to manage those differently and what tools you can apply. One password, Doppler, other kinds of secrets managers to deal with that stuff. So you need to learn all this stuff. I guess what we're saying is the space for place for human expertise is not gone. In fact, it might be accelerating in some ways.

Todd Geist: 17:05

Don't know what the future's gonna bring. Maybe the models will get good enough, but I actually don't think so. I don't think they have the kind of understanding currently that is required to manage this kind of complexity. I think they're very good at generating code, and they can be guided towards generating good code, but to manage this surface, I don't think they can do it today.

Kyle Duval: 17:29

I think the the problem the problem I foresee is that five years ago, right, when I was learning how to be a developer, in order to write software, I had to know how to write software.

Ernest Koe: 17:41

Right.

Kyle Duval: 17:42

And I had to learn all of the things that I had to learn so that I could write software. Right? I had to learn how to write JavaScript. I had to learn how to write TypeScript. I had to learn how to run an application.

Kyle Duval: 17:50

I had to learn how to etcetera, etcetera, etcetera. And if you're learning how to be a developer now, you don't necessarily need to know all of those things. You can say, you know, pull down a template of a NextJS application and set it up with this, this, and this. You know, I want it to be able to do this. And CLAUD or Cursor will just run that for you and set it up.

Kyle Duval: 18:11

And if you don't have the background to know what's going on, you don't know what's going on. Right? You don't know if it's doing something wrong. You don't know if the the code that's being written is useful. So if you're learning to be a developer now, my concern is not that I mean, I am concerned that people are gonna make things that are insecure.

Kyle Duval: 18:32

But my concern is that the people who are learning to be developers won't actually learn how to be developers. And then because the the where previously the gate to making software was learning to be a developer, now the gate to learning to be a developer is wanting to learn to be a developer and wanting to learn how to work these things because you have to take the steps you need to take to, like, put that information into your brain instead of just letting it instead of offloading that thinking.

Todd Geist: 19:01

Yeah. Yeah. I I

Ernest Koe: 19:03

don't know what that line is. I mean, I think two things ring true to me. One is complexity is definitely going asymptotoptical. You know? Like, it's just going crazy.

Ernest Koe: 19:16

I actually have a really hard time myself, like, dealing with the the levels of complexity that I'm generating for myself. Like Yeah. It's so fun to have all these agents and all these skills and all these services. And, like and I was like, what did I do yesterday? You know?

Ernest Koe: 19:31

Like, you know, which Asian did I work on? And where where I've refactored my personal Asian and skills, like, so many times. Like, the the drift has been so great that I I just wanna throw it all away and start start again, you know, because, like, it was such a simple idea before. Just do my time. And now it's, like, doing all these things, descending briefs, it's, like, getting it wrong, and I've gotta refactor it.

Ernest Koe: 19:54

And I was like the the the cone gets wider and wider. You know? And, like, then I'm I'm deploying more agents to, like, get ahead of my, you know, my problem space. So I don't know. I I think that that's definitely one problem that I do see, like, you know, on the horizon how we deal with just, like, the slop of everything coming at us.

Ernest Koe: 20:17

The other thing, though, Kyle, is, like, I'm wondering and maybe we are in the thick of things now, and we haven't crossed the the the the the, you know, the ocean yet to get to the other side. Right? But this I I can safely say that I have not written a lick of rust code in my life ever. And I I maybe wrote very primitive swift scripts and code before, you know, 2025 or 2026. But I don't actually feel incompetent in Swift or in Rust, even though I don't fully understand syntactically all the things that are going on in there.

Ernest Koe: 21:04

But the the the logical primitives and and the processes and the concepts that I've gained from whether from TypeScript or from from JavaScript, from even PHP and, hell, even Pascal, you know, going back to college days. Those feel, like, durable. You know? Like, I don't think I need to be a new kind of programmer, but I do think that my DSA concepts, the data structures and algorithms concepts that I that carried with me over the years are actually very useful. Right?

Ernest Koe: 21:35

And I sometimes wonder if the conceptual mental models and frameworks are more important than actual syntactic fluency, if you will. And and but but maybe maybe you don't even need that. You know, I think that's the question. Right? Can you vibe code your way through by just trusting Claude to fix its own issues?

Ernest Koe: 22:01

That's where I'm kind of not sure about, you know?

Kyle Duval: 22:05

Yeah. I think I agree, which is that like I don't need to know how to write a Rust program in order to create a Rust program. Right? But I should probably know what a loop is. You know?

Kyle Duval: 22:19

And I should probably know looking I I can probably reading Rust for like, I agree. I've never written Rust in my life. I've never gone through and hand handwritten the Rust program. I've read a lot. I've read of a number of them at this point because I have I've I have a couple things I'm spinning up and I'm like, this would be fun in Rust, know.

Kyle Duval: 22:37

Might as well. And reading Rust, not that different from reading TypeScript it turns out. It's a different language. It does different things. But like the syntax and the structure, it's a programming language.

Kyle Duval: 22:47

It's not all that different.

Ernest Koe: 22:50

So But your strategy doesn't change that much even though there's different different capabilities of these things. I mean, rest with, you know, memory handling being automatic and all that. That's very, very nice. You don't have to worry about pointers and nonsense like that, And you can be more conceptual. But I just wonder, like, how much can you ignore?

Ernest Koe: 23:10

Can you purely be product driven, no understanding what the algorithm is or whatever? And we're at a point where that that can be done. That's not clear to me. You know?

Kyle Duval: 23:21

Yeah. I think it depends a lot on what you're building and who it's for. Because if it's for you, right, if it's if it's very on your computer and no one else is gonna use it, then, yeah, be product driven. Maybe write a couple tests to make sure stuff doesn't break. But, like, as long as it's doing the stuff you care about, that's all that matters.

Kyle Duval: 23:40

Yeah. But if you're writing if you're writing, you know, a TypeScript linter, which I was just reading more about this morning, like, it's very important that you know how the performance of that linter is and what things are happening inside that Rust based linter that are maybe slowing it down. Right? Because the value proposition for that is performance and is speed and is correctness. So there's an element of like I I think for products that ship to people, you need to have tests.

Kyle Duval: 24:13

You need to have tests that you write and that you do not even if you're writing them with the assistance of an AI tool, it needs to be a test that you said this thing should not break if it's run like this. Right? Something with this input should get something with this output always. I think that's an that's an important thing for for production ready code is to have regression testing that is overseen by a human. Because if it's overseen by an AI, there's there's nothing to say that it won't break other than like like we were talking about earlier, my test breaks.

Kyle Duval: 24:54

Claude's like, wow, the test broke, whatever. Let's just delete the test. That's not important. Even though, you know, three weeks ago, you made that test with Claude because it was really important, but it doesn't know about what happened three weeks ago, but you do.

Todd Geist: 25:09

Yeah. I think I think the other thing that's super important is like how is it being deployed, right?

Kyle Duval: 25:13

If you're

Todd Geist: 25:14

deploying a web app and you don't understand the threat vectors that and how to protect against, like you don't have to know every single possible threat vector, but you have to know that there are them and that they exist, and you have to know how to at least use the tools that protect you against everything that's known. Right? And so things like secrets management, we touched on before, is one of these things. Like, you really have to understand that. I mean, if you don't, you're you're running a big risk.

Todd Geist: 25:43

And other things like different kinds of like cross site scripting and other other kinds of attacks that could come in and, you know, just cause havoc. You need to have at least some understanding of how that works so you can protect against it. So that's a whole different set of things you got to know than let's say if you're building something for the App Store, you know, where they have a different, there's a whole different way of just deploying it, distributing it and the different security issues, and the fact that Apple's probably gonna prevent you from doing anything dumb, at least, well I don't know, we'll see what Claude Mythos can do to that story, but for right now it's pretty hard to ship something into the Apple Store that has major security flaws. That doesn't exist on the web, there's nothing you have to pass.

Kyle Duval: 26:32

There's nothing stopping you from putting up a website.

Todd Geist: 26:34

It's stopping you from putting something on the web. So I think there's like, those are all things you just have to learn. And so that's why SAS is not dead, that's why software development is not dead. But whether you're, the particular syntax being an expert in any given language, probably not as necessary anymore. But you need to know the other parts.

Ernest Koe: 26:57

I I do think that it's making it easier for people to compete in the SaaS and software space because it's easier to to bring your own version or your own taste to the party quickly. So speed is no longer a real moat per se. Right? The the pure coding engineering part, if you have all the other capabilities lined up and the ability the the ability to bring things to market more quickly, then it does put pressure on Yeah. SaaS to innovate and do something different.

Ernest Koe: 27:32

Yeah. I think we are still moving towards an agent first world, which I I think seems transformative. I think it's going to have profound sort of pressures on traditional software companies and how we think about pricing and how we think about the features we deliver and all that stuff. But it's not clear to me that that everybody is gonna be building their own building and running their own software just because they can.

Todd Geist: 28:02

I I don't think that's going to be the the way at all. It's not gonna work. I there's gonna have to be I think it's probably not possible with, as long as LLMs are the model that's driving all this. I don't think they bring a level of understanding that you need to actually do this top to bottom. People are working on other types of models, world models, models that have actual understanding, not just the ability to predict the next most likely token.

Todd Geist: 28:34

So I think we're all, all of us in this are professional software developers. We have a little bit more life left in us. Just Maybe

Ernest Koe: 28:44

too quick. Pretty clear,

Todd Geist: 28:46

which is great because it's super fun to be able to build and like to to remove a lot of the roadblocks that we have in the past just because, you know, you know, we couldn't we can't write Rust, so we can't, like, explore these areas where Rust speed might really unlock something. In fact, we can now. So that's pretty great.

Kyle Duval: 29:05

Yeah.

Todd Geist: 29:08

What was I had a pivot I was about to make. Something to do with so I guess a couple things. So what we're saying, I mean, maybe to sum it up is like, agentic engineering is definitely the way forward and like you should be if you're in this field, whether you're a Famic developer, any kind of software developer, you should be exploring this. However, do not listen to the idea that you can just safely go out and do all this without learning some of these hard things. You may have you're gonna have to learn them in some way or another.

Todd Geist: 29:45

And so the pivot I wanted to make is that if you are using something like FileMaker, you already understand a lot of the security footprint of or the security landscape around FileMaker, and so one of the things is if you're thinking that I'm gonna recreate whatever my FileMaker thing is into a new app that's not running in FileMaker, that's actually more possible than it's ever

Ernest Koe: 30:11

been

Todd Geist: 30:11

before, But don't think you're not gonna have to deal with learning and understanding a whole new security landscape that you didn't have to deal with before. And so one of the focuses of our group here is to bring agents to FileMaker in any way that we possibly can. So in the last episode, we talked about something we're working on, which I think we were calling ProofKit MCP. I don't know what the name of it's gonna be when we get it out there, but it's coming very soon. ProofKit AI or something.

Todd Geist: 30:47

I don't know, Ernest. We're gonna have to figure that out. And I think we're gonna have Eric on to talk about that in the next in another podcast. But we can build apps. We can use agents with FileMaker, either using the methods we talked about last week with with ProofKit or some of the things we have with, Ottomatic enabling you to do AI assisted either DevOps or adding tools to Claude and other things using Automatic.

Todd Geist: 31:20

And so we've done some of that stuff. We've shipped things with AutoFMS and Automatic that are taking advantage of AI in one form or another, so that's actually the reason we have Kyle on this week, to talk about that. We just took a long time to get around to it, but we should talk about some of the things that have shipped recently in automatic and auto FMS that are part of pretty much everything we do is like, are we any innovation or anything that we're trying to ship into our product space or into like the tools that we build, we're thinking is this a agent first? Is this making agents more powerful? And so that's what we have here.

Todd Geist: 32:02

So Kyle, give us a rundown of what we've shipped into the our automatic AutoFMS ecosystem that are basically from the idea that agents are powerful and we need to use them. Let's start there.

Kyle Duval: 32:18

So I don't remember exactly which version, but a few versions of AutoFMS ago, we shipped MCP servers in AutoFMS. So I'm gonna assume you know what an MCP server is, but there's two different kinds in AutoFMS. Every share server just, like, ships with an admin MCP. So the admin MCP is intended for basically developer use to do DevOps. So it lets you work with files.

Kyle Duval: 32:44

It lets you work with your server. There's a number of tools on it that let you do, like, encryption of files and listing and closing and opening and restarting services. So it it's kind of the the I want in Claude to be able to say, hey. My server's being weird. Let's restart the data API instead of going to open your AutoFMS and going to the services tab and restarting the data.

Kyle Duval: 33:10

Right? It's a productivity tool to help you do DevOps better.

Todd Geist: 33:15

Yeah. So you so you can say things like, give me a backup of this file, it will just get a link, you click it

Kyle Duval: 33:21

and download. Give me a download. Right. Give me a copy of this file, give me an XML of this file, that kind of thing.

Todd Geist: 33:27

And so let me just jump in there for a second. So one of the things that Claude in particular, like MCP servers work with all of the coding agents and all of the AI tools, but like in recent months Claude has shipped, like if you get Claude desktop, there's now three modes, you've got the chat app, which you can do, we've talked about, we were just mentioning like give me a download, you've got Claude CoWork and you've got Claude Code, and these are sort of like three different levels of how technical do you want to get to build things, And so not only can you just chat with your FileMaker server and tell it to do things like close this file, open this file, back up this file, blah blah blah blah blah blah. You can also use something like Claude Code or Claude Cowork to make an automation, so it is relatively easy to do something with Claude Cowork that would be something like this every morning at 4AM make a backup of this file and download it to this folder on my computer, And co work will happily write that for you, it'll schedule the script and you just have that, right?

Todd Geist: 34:36

And so that's what we mean by when we say agent first, what we wanna do is make whatever agents you're using including like say OpenClaw, which got a lot of press recently. If you're a happy OpenClaw user, you can use the AutoFMS MCP server to automate stuff on your with your FileMaker admin stuff Yeah. Using those tools.

Kyle Duval: 34:59

Yeah. And I will also add the AutoFMS developer API exposes almost all of these features as well. So if you want to I guess the the alternate form of doing what Todd just mentioned about getting a download every night, Right? If you don't want to get a ping at four in the morning that says, hey, have a tool thing I need to approve, you know, or whatever, you know, oh, I upgraded to a new model and it doesn't have access to this. It would also be exceedingly simple to say, hey, write me a cron job that runs a bash script that hits this URL with this API key and does a download to this folder.

Kyle Duval: 35:38

Right? So I'd say the the m t pool MCP tool enables a lot, but just having the ability to basically build some custom tooling around running a script every night is also just a thing that is enabled by having an LLM. I do this I'd I I bring this up because I do this a fair bit, which is, like, a lot of things that I do require, like, restarting my local database for one of my projects and, like, resetting stuff requires, seven commands. So instead of running seven commands, I was like, hey. I have this whole flow.

Kyle Duval: 36:14

I'm not super great with bash syntax, but I can read it. So I said, hey. Write me a bash script, you know, and you can do that. And the developer API enables all of that. So if there's something on your server you wanna do that's not exposed by the MCP, you could do that as well.

Kyle Duval: 36:29

It's great for yeah. The admin MCP is great for, like, DevOps management and that kind of thing. If you want to build a application, an MCP that works with your data. So let's say you have a super user at your customer who wants to be able to ask Claude what's happening. You know, get me the stats for users in the past day or what's our revenue for last week, that kind of question.

Kyle Duval: 37:01

AutoFMS also exposes an Odata MCP. If you watched I think that Clarus had a community live about something similar. Clarus is releasing MCP's, which are similar but built into Clarus Connect and not quite the same. But the goal is essentially that you can query live data in a database using an MCP. So AutoFMS exposes this via scripts.

Kyle Duval: 37:25

So you set up an OData key, in AutoFMS and then can set up an MCP that will run scripts that do specific things. And you can set up custom input schemas so that you they the LLM knows what to send in, and, you can coordinate that whole tool in MCP building in the AutoFMS UI, and then it will expose that as an MCP. So you can say, you know, create me a customer from this business card. You can say, what are our analytics from last week? You can say, show me, you know, files that were uploaded today, whatever.

Kyle Duval: 38:05

You know, anything that you can do in a file maker script that you can return data from, you can use the Ottomatic P4. And, you know, your your imagination can run wild about what can happen in there. But because it's literally anything that can run-in the script. It can be an action. It doesn't have to be querying a database.

Kyle Duval: 38:22

You can say, you know, you can I don't even know? Like, if you're using the CRM, for example, it can be like, you know, delete this user. I don't like them anymore. That kind of thing. Right?

Ernest Koe: 38:33

This is yeah. This is insanely cool. I I I know I know it sounds like it's this MCP server, but I I don't wanna I don't wanna bury the lead here because your existing FileMaker databases have a lot of scripts. And a lot of them drive, you know, business logic and business rules and do things that are very important. And and, really, anything that you can any any of those scripts now can be accessible to AI by using auto FMS MCP server to enable that.

Ernest Koe: 39:07

So it turns your FileMaker databases into an AI agentic surface, which is very, very cool. Very early on when I discovered this on AutoFMS, Kyle, I mean, I was late to the party. I had I was managing a whole bunch of domain names, and I wanted to do some analysis on the right domain name. So I was like, let me see what I can do. So I I turn I turn on AutoFMS MCP server on that on the database of, like we have so many domain names that we've been sitting on, you know, just because we're trying to do brand stuff with it.

Ernest Koe: 39:40

Mhmm. And build, a a domain name analysis tool by just running a very simple script to, like, just talk to all my domain names and have do the sentiment analysis. It was like it was amazing. I was like, ta da, this is great. So good job on that.

Ernest Koe: 39:54

I loved it. But really anything, if you have scripts that tell you or your databases is your family good databases, where your customers are, your invoices are, you know, your performance data is, turn it on and let Claude go to town, you know, and see what you can do with it. Yeah.

Kyle Duval: 40:15

And also, like, it wouldn't it's not hard jump to say the the value prop of a lot of the MCP stuff is not the I can do a query that I tell it to run this tool. It's Right. I can build this into some sort of larger ecosystem of tools. So I can have a connection to GitHub and I can have a connection to my FileMaker app and I can have a connection to Google. Right?

Kyle Duval: 40:43

And all of those things can work in some sort of concert to give me all the information I need to make some sort of decision or to do something. So if you're

Ernest Koe: 40:50

finally This is a key point. This is a key point. I mean

Kyle Duval: 40:53

It's it's not an island.

Ernest Koe: 40:55

Yeah. The ability to get multiple streams of data to your AI is one of the most important tangible gains we're gaining in this AI world. You know? When I bring in my Tazorac data and I bring in my meetings data from from Rome or from Zoom, Zoom actually has a new connector out for cloud.

Todd Geist: 41:13

It's really awesome.

Ernest Koe: 41:15

It's fantastic. I love it so much. And then you just build, like, this context. Like, you have your own little, like, Palantir for your your private Palantir because all this data's coming into you, and you're the fusion center for all this stuff. You know?

Ernest Koe: 41:28

And then it just goes to town and and does summaries for you, write briefs for you. That that's that's very, very powerful. So I think I think that's absolute absolutely spot on. The other thing I was going to say about MCP and and automatic and auto FMS is that yeah. I I think that we often take for granted so one of the hard things about MCP servers is dealing with auth and authentication.

Ernest Koe: 42:01

It's a pain. And I don't know how many custom OAuth things I've been I've I've had to spin up just to, like, tie two services together and give internal people access because we're using Google auth. And you don't wanna have to, like, just ship API keys everywhere in cloud desktop config and you want it as a web connector. But the cool thing about doing this in the FileMaker space is that if your FileMaker database already has permissions, your scripts are already permission under the same security schema. So you're not you're writing scripts that you or that in theory adhere and comply to your security permissions that have already been set up.

Ernest Koe: 42:45

You're not going behind. You're not going over the top. You're not, like, bypassing things. I think that's really important. You know?

Kyle Duval: 42:51

Yeah. And as long as you have a data a data API key for enabled for Odata that is scoped to the right user.

Ernest Koe: 42:57

That's right.

Kyle Duval: 42:58

So you say, this user gets this API key. This user gets this API key. Right. Yeah.

Todd Geist: 43:03

Yeah. It's we've been talked a lot about this in the beginning, but just again, I think people should take seriously what they know about FileMaker security as a very important thing because security is going to get more difficult. And so don't undersell what you've already built. You have a system that has has, in some cases, decades of iteration in it, and it should be secure based on the FileMaker security schema that we've all had for two decades now. The last time it switched was essentially 2006, I think, when we went to FileMaker seven.

Todd Geist: 43:46

So that's a real thing that you can still rely on. I wanna move, so that's where we sort of I think that's where we are currently. So that stuff all exists. You can find that going you can find that stuff by going to our docs where we have a new feature.

Kyle Duval: 44:07

New feature.

Todd Geist: 44:07

Should be out by the time this podcast hits which is

Kyle Duval: 44:11

It's releasing on the day of this recording.

Todd Geist: 44:13

Day of this recording, we have AI assisted what? Kyle, tell us.

Kyle Duval: 44:19

Yes. AI assisted, I guess, chat. It's chat, doc, searching. Chat, doc, reading. I'm not really sure exactly what to call it.

Kyle Duval: 44:27

But it's basically, do you have a question about how AutoFMS or automatic works? Type it in. It does doc doc based querying and information gathering for you. So if you don't know exactly which words to search for because the bay built in search is like basically word word search and not semantic search. You can say, hey, I want to set up a publicly accessible build so that anybody can download a demo of my FileMaker application.

Kyle Duval: 44:58

It can say, that's great. Here's the public builds documentation. You 'll need to set it up on an HTTP server somewhere whether that's on FileMaker or a separate web server. And then you can do FTP based publishing or SFTP publishing from AutoFMS to send to that web server. Here 's the steps you need to follow, set it up in auto deploy.

Kyle Duval: 45:20

Yep. And instead of having to find the public builds page, which is in like a three menus deep in the auto FMS Yeah. It will just it'll link you to the docs. It also, I will say, is the the the documents it has access to includes the community forum. If you have a a bug that you are like, oh, I have this problem where whenever I run an auto f m s deployment, the next time I restart FileMaker server, all of my encrypted files don't open.

Kyle Duval: 45:52

Which is a known issue with FileMaker Yeah. And but there's like four doc, there's four forum threads about it. And you say, hey, have this problem. It it pulls up those four forum threads and it's like, hey, it looks like this is a known issue. Follow these troubleshooting steps.

Kyle Duval: 46:08

But

Ernest Koe: 46:11

Do we do we have the feature that McDonald's included in the chatbot where it will also write code for you when you

Kyle Duval: 46:19

I think I've tried and it won't.

Ernest Koe: 46:21

Oh, okay. So

Kyle Duval: 46:23

so hopefully, no.

Todd Geist: 46:26

This is a The forum search is really cool because it'll produce a really nice document of what it discovered and it'll link back to the original forum post. This is super powerful way to figure out what's going on with your automatic or auto FMS stuff. So that's shipping, and then we have a couple things which are in the works with which we're just rolling out. These will be automatic features, so we're gonna build MCP servers and into automatic that will do multi server DevOps kind of thing. So if you have an if you have all your servers in automatic, you'll be able to, like, transfer files from one server to another with an agent, for example.

Kyle Duval: 47:13

Right. With an agent move files around or with an agent do log querying. So Mhmm. You can do one of the tools we're building for it is multi log querying using essentially like a grip like tool. So being able to say, hey, I'm having this problem on my FileMaker server.

Kyle Duval: 47:34

Can you show me errors that happened in the last twenty four hours? And it will search across the event log and the auto info log and the access log and the script error log. And maybe it'll search for, you know, it kinda depends on what the AI chooses to do. But it might search for the word error and then pop up. Hey.

Kyle Duval: 47:50

Here's the last 100 errors that happened in the last twelve hours. It looks like you have an auth error in auto f m s. Right? Which it will show a lot because there's an auth error that logs there when you fail when you try to log in, it doesn't work.

Ernest Koe: 48:04

But

Todd Geist: 48:04

But but it's also gonna be Yeah. It'll also be intelligent enough to to notice, like, yes, there's an auth error, but this is just somebody logging into automatic with the wrong Sure. Password, so you don't need to worry about it. Right? Yeah.

Ernest Koe: 48:17

I can't wait. I can't wait for when, you know, we always get asked like, my database is really slow, what's happening? Yeah. And maybe it's gonna be able to tell you.

Kyle Duval: 48:30

Right.

Todd Geist: 48:30

Top Call Stats too.

Kyle Duval: 48:32

Yeah. I'm trying to build The Top Call Stats by itself isn't a very AI friendly

Todd Geist: 48:36

No. We have to build on it.

Kyle Duval: 48:38

Yeah. It's not a very human friendly thing. Right. Right. And it turns out those things are almost the same.

Ernest Koe: 48:42

That's right.

Kyle Duval: 48:44

And so I'm trying to build a tool that will help with that. But like the the win the win I see is, hey, I'm having a problem with my server. I don't know what's going on. The stuff in the data API seems a little weird. It says, oh, it looks like your data API isn't responding to calls from the auto FMS web hook receiver.

Kyle Duval: 49:01

Let's restart that. And you say, yes. Let's restart that. And it restarts the data API and then everything works.

Ernest Koe: 49:06

Right.

Kyle Duval: 49:06

Right? That's the kind of like tools all coming together. Because right now, like as a human working with auto FMS, you say, something on my server is weird. Let me go look at the logs. You walk into the logs, you look at the web receiver log, you see some incoming requests, but none of the, like, success messages, and you say, that's kinda strange.

Kyle Duval: 49:27

And then you go to postman, and then you send a couple of requests to the data API, and the data API is not really responding. And you say, okay. Interesting. And you click back three different tabs because you have to click back to the FileMaker server tab and then over into the processes tab and then you say restart the data API and it says, are you sure? And you say, yes.

Kyle Duval: 49:43

Right? And now you've spent seven minutes reading a bunch of logs and clicking around and clicking buttons and waiting for stuff to happen and then testing again. And all of some of those things will still happen probably. You'll probably still wanna run a test in Postman that says, does my data API work? You'll probably still wanna actually read the logs.

Kyle Duval: 49:59

But the goal is to get you in a direction faster.

Ernest Koe: 50:03

That that was really that was really, well rehearsed, the whole spiel that you did. We need to clip that. No, I think that will be amazing.

Todd Geist: 50:13

One of the things that agents are really good at actually is parsing logs and going through lots of data and collating it and bringing it together and having and even having opinions about it. Log data made it into the training data for all the models, so they're pretty good at it. And now we'll be applying our own context to it to make it smarter about the particular configuration of the automatic ecosystem. So it should be able to give you good advice and also tell you like, just restart your this or that. And then again, the other thing that's really cool is this now opens it up for agentic automation, right?

Todd Geist: 50:54

You could, in the future, we're gonna experiment with, can we build in alerting systems that send an alert to an agent that when certain criteria are met, we've got too many errors in the data API or the data API hung, what can we do about that? And the agent could proactively make a decision to say, restart the Data API, which restarting the Data API is a low risk kind of activity. Only gonna kill whatever processes are running. It's not going to damage the database or anything like that. Restarting it can often have a big impact on whatever's going on in your system if one of those scripts is hung, for example.

Todd Geist: 51:35

So the main thing is giving you access to build, agentic, tools on top of

Kyle Duval: 51:46

Yeah.

Todd Geist: 51:47

Whatever Ottomatic Yeah. Is

Kyle Duval: 51:49

And there's not there's not like if you're the other example I could see is saying, I have this process that runs, you know, every day. And over the past three weeks, I've been having this problem with it. Let me add some logging to the script using the AutoFMS app log. Right, that prints out to the to the app log log, right, that will be queryable. And then you say, okay, every night at seven or every morning, I guess, at 08:00 in the morning, Look over the logs, see if there's anything weird, and then if there is, tell me about it.

Kyle Duval: 52:25

Right? And that's Claude, co work, you know, recurring task that gets run every morning, and it gives you a summary of what happened. It says everything looks fine, or it says if it's still not working we should try something else. Right? It tells you was running the process that's failing.

Kyle Duval: 52:41

You don't have to read 400 logs of this process. You have to see what happened.

Ernest Koe: 52:46

I think the main message here is that you know, this is the starting point to enable a more more more agentic way of working with all these services. Yeah. I can imagine a future, which I don't think is that far away. I think we've been experimenting on some things here where right now in the story, you're responsible for bringing your own agent to the party and we'll give you the tools and access. So your agents could be Claude or Codex or whatever you want, you know, goose or anything else you want to bring to the party.

Ernest Koe: 53:20

But I can imagine a near future where we have agents too that just live on the platform and that you can spin up and they're little helpers that do smart things. And sometimes you can, you know, configure the kinds of access or the kinds of permissions that you want to grant them. So a little log analyst agent that just happily sits there and just watches all your files and you give the permissions to do a few things on your behalf and other things you say always tell me, get my approval on Slack or on Telegram before I do this, that kind of thing. That would be where we're deploying agents that you can access rather than bring your own agent to the party. Right?

Ernest Koe: 54:07

So I think the the sky's the limit here, and I I don't know how this will evolve. But I think the first step is make things agentic, accessible, and then see what happens.

Todd Geist: 54:19

Yeah. Right? Yeah. We still don't know where where this is all gonna land, obviously. But Yeah.

Todd Geist: 54:25

We do know that you need to kinda hit your your wagon to this wagon train, which is agents. And that's basically the main focus of everything we're doing really as a development shop. I mean, from the products we ship to what our services teams what our services teams work on. And and of course FileMaker is a key part of that. Opening up FileMaker to to agents is, I mean, I just think it's the most important initiative that we have.

Todd Geist: 54:58

We have thousands of FileMaker customers who have worked with us over the years one way or another, big community out there, and we need to provide the tools and the expertise and the training to all these folks so that we can keep going for as long as it makes sense to keep going with FileMaker. And you know, that decision about whether to stay in FileMaker, I think a lot of people are asking, well, why don't I just vibe code up something else that replaces this? Again, that bar is lowered, that's a thing that is achievable in a way that it wasn't before, but don't get fooled by the hype around like you're gonna be able to vibe code something and ship it securely and have your 25, 50 or 100 users to start to go to town on it. That's risky and you should know that and you should either get the training you need to overcome that, or alternatively, if FileMaker is working in every way but you just wanna be connected to agents, follow what we're doing because that's part of our goal. FileMaker as a database is there's nothing wrong with it in terms of it being a database.

Todd Geist: 56:16

The main issue right now is that agentic access is low, and it needs to be improved.

Ernest Koe: 56:24

When I think about what I'm paying for when I pay for software today, I I think I used to have an idea that I'm I'm paying for features. And I think maybe that's still a little bit true. But I think I I'm I'm much more partial to the the the framing that I'm I'm actually paying for the years of testing that's gone through the thing, the the the, you know, the the security architecture that survived multiple generations and multiple years of, you know, battle hardening in in the real real world and all that. Because I don't think those things are easy. And I think those things accrue only with experience and some very, very hard lessons along the way.

Ernest Koe: 57:09

And I feel like we're entering a world where we're maybe a little bit we we need a little bit of, what's the word, humility to make sure that we don't, you know, take this naive approach to to all these things and wind up in a kind of a sorry state. You know? So I I I don't know. And I think that there's a lot of value in existing systems that work, but they must be agentic for us to really leverage their capabilities.

Todd Geist: 57:38

Yeah. Yeah. Otherwise, you are gonna get left behind. There are ways forward within FileMaker, and that's kind of one of our main focuses and we continue to ship and to work on that. And yeah, like this is still a craft you have to develop expertise in.

Ernest Koe: 58:03

Alright. Well, I think with that, let's I think this is a wrap, and we will see you next time on the next episode of The Context Podcast.

Todd Geist: 58:12

Yep. Thanks every thanks very much. Thanks.

Kyle Duval: 58:14

Thanks, y'all.

Todd Geist: 58:15

Bye bye. Bye.