A Journey Integrating AI Agents in WordPress Using MCP

Dave:
Hello everyone, I am your host today, Dave Lockie hosting Emerging Tech and with me I have Tole. So Tole, maybe you could say start by correcting my pronunciation, giving us your full name and a little hello, hi to everybody.

Tole:
My nickname is Tole and full name is Tihomir Dmitrović. I know it’s a bit hard and I’m an old school web developer.

Dave:
Perfect, thank you. So I asked on Twitter who’s doing interesting stuff with WordPress that should come on Emerging Tech and talk and a bunch of people shouted your name and you were kind enough to respond and be willing to come on the show. One of the things that I know is when we cut you, you don’t bleed WordPress. You know, you’re not kind of heavily embedded in the WordPress ecosystem. So why don’t you give us your story starting wherever you’d like and ending with why you found yourself Suddenly building a WordPress MCP.

Tole:
All right, so I would like to say that I’m classic developer, I always want to to be programmer and as soon as I graduated I started working and I had a luck to work in mobile industry during 2000 and if you remember it was also a crazy time those mobile phones, logos, pictures, polyphon memory, polyphon melodies, games and we were in that business and what we were developing those kind of quizzes and delivery services those are web services and administration GUI and there was that need for shipping a lot of small code customizations and then we were our solutions, we were changing and developing them as no code solutions. And I spent something like 10 years in that environment and kind of specialized building no code solutions. But this is all non WordPress related. But you know that classic backend developer, something like that. But I was hanging out with some younger guys and they were showing me look, I’m building websites about movies, about games and it’s WordPress great thing. And I was, I always liked and the web and the PHP and I saw that as a great opportunity and started building putting some interest in the WordPress and I even developed my first plugin in something in 2011 it was first iPhone optimized mobile gallery.

Dave:
Amazing. So you came to my attention because you’ve been building a, an MCP for WordPress. So maybe let’s start out by defining terms. What is an MCP?

Tole:
MCP is just protocol API. You know, we are all working with the APIs. So it has some actions that you can call with some arguments and get back. So some kind of set of rules how two platforms, two applications are communicating or over the net. Right, so basically it’s just that.

Dave:
And how does it, why does it have a different name from API? What’s an MCP compared to an API?

Tole:
It’s optimized for use with AI agents. Okay, so what basically it means when you are working with those AI agents. AI APIs actually, because every that agent that we are communicating is actually making API requests to, I don’t know, OpenAI or Google, whoever is provider of the AI model. And inside of that we can put some instructions how to behave, but we can also put descriptions of the actions it can take.

Dave:
Okay, so MCP stands for model context protocol, correct? Yes, and it was created by Anthropic who make Claude, but it’s also now adopted by lots of other models including their OpenAI.

Tole:
Yeah, right now it seems like the most prominent standard, but it’s good to know that there is maybe almost 20 different variations of the standard by different players. And we are not talking just about MMCP. Second most famous is Google’s initiative agent to agent. It’s a slight different thing, but again is trying to define how some smart agent that people are using, like in a cloud desktop application, will communicate with some end external system like your WordPress website, maybe even your Komoot shop.

Dave:
Okay, so we can think of it simply as an API that’s designed for AI agents to be able to use.

Tole:
Yes.

Dave:
Okay, great. All right, so what initiated your thought process of huh, I need an MCP for WordPress. How did that start?

Tole:
Okay, well, curiosity. Okay, so let me just go step back with what am I doing in the WordPress? Right, so in one moment there was that Alexa thing and smart speakers and they also advertised that it is that it was some AI stuff, but it’s not even, not even nearly comparable with what we have right now with these large language models. Yeah, and we created that no code tool called Convoworks for managing that kind of stuff and it failed. That market failed completely. Me as an entrepreneur and marketing guy, it’s all also didn’t go too good and I was about to end up everything but that was the moment that ChatGPT came out. And I said, okay, no, we have a great no code tool, let’s continue with the GPT. And since then I’m doing that actually in my something like free time. And in all that, what’s happening on this, that AI market MCP came out, it’s actually very something that was missing and no wonder that people are using that so much. And I said, why wouldn’t I have that in the converse once I started? Okay, so it’s not just like simple rest API. It is kind of bit more difficult to implement protocol when you are programming your own server. And it’s the main issue, for example, comparing to the classic APIs, it’s using two channels.

Dave:
Okay.

Tole:
You have one channel to issue comments and on another channel you are receiving responses. This way actually you can trigger, you know, asynchronous operations. That’s why that kind of stuff is interesting. So anyway, I implemented that. It fits quite well with what we already had and it’s a nice thing.

Dave:
So you had, did you say it’s called ComWorks?

Tole:
Convoworks?

Dave:
Yes. Okay. And so this platform that you’d built, you wanted it to be able to integrate with WordPress using this conversational AI interface, is that right?

Tole:
Yes.

Dave:
Okay, fine. And there was not an MCP available or at least known to you at that time. So you’re like, well, I’ll just build one.

Tole:
Yeah. So this, what we have with that converge is a universal workflow builder and it has as ability which I can add new platforms, new endpoints, because in the end it’s always some endpoint that you have to expose and you will receive some request on that endpoint. It’s something that we are doing all the time.

Dave:
Great. So it’s a way for your platform to talk to WordPress, basically to use WordPress as a client or channel or something. Okay, cool. And so how does your MCP work under the hood? What’s the actual technical integration? You know, is this a. Is it a plugin? Is it like a API direct? Like tell us how it works technically.

Tole:
So actually you have to first install converts.

Dave:
So Convoworks. So C O N V O W O R K S the no code powerhouse.

Tole:
Okay, yeah, so it’s actually you can build different workflows and when you are building that workflow, then we call that service, you can, it’s just a workflow JSON definition which components are executing. And then there is another thing, you can enable access to that workflow from different platforms. Like one platform can be chat, but one platform can be MCP server which receives a request and starts that workflow. So what I’ve done here, I’ve done just specialized component for MCP and you put that in the Merc flow and here you can, you have some basic MCP stuff is already implemented. And in that converge service there is a place you can add your own components and implement what they will do. Add your functionalities tools that you want to expose.

Dave:
Okay, so you’ve got a ConvoWorks platform like SaaS and you then have like a ConvoWorks Worker plugin that you install in your WordPress site, which comes with some I guess like generic WordPress MCP functionality. And then you can also within the plugin extend that to customize it for your needs. Is that accurate?

Tole:
But almost. So there is no SaaS. Right. Everything is that plugin. So it’s all localized, private. Imagine that you can have for example multiple websites with one Elementor. You go into Elementor plugin and create new and then create new. So and everything is working on the one website. Here you have those workflows and you just create new in that plugin and you enable. Okay, this one is connected to MCP, this one is connected to some hooks and that kind of stuff. But anything anyway, everything is inside that plugin.

Dave:
Okay, so Convoworks is an application that runs within WordPress via a plugin and can connect out to other services from your WordPress site.

Tole:
Yes. And you defend that in those workflows. And there are templates for workflow. You don’t have to start from the beginning. And one of the templates is MCP server. You just create that and you can modify try it.

Dave:
So it’s all open Source. Cool. So ConvoWorks gives you your own AI powered workflow generator and. Yeah, like systems integration engine.

Tole:
Yes. So basically, so it’s a visual programming because we have quite low level components. You can have foreach, you can have it then and that kind of stuff. You can also call any PHP function which is defined on the on that website and all that you need to define what will happen in which moment, what converge solves to you that you can easily hook to some webhook to some WordPress hook or filter or action. Right. You can also create specialized endpoints and then in no code. Right. So you just declare for example converge test and it will catch that request and execute what you defined in the no code interface.

Dave:
Cool.

Tole:
And you can mix that with the GPT in the meanwhile and whatever you want. And like we have that connection to WordPress hooks, we have connection to the API custom API endpoints. We also have that new connection, the MCP server.

Dave:
So it sounds a little bit like a WordPress host. Well like a WordPress based NA10. Okay, cool. Very cool. So what was the hardest part about building this thing? Like what was. What were your biggest challenges technically?

Tole:
All right, so these two channels, keeping those two channels in the WordPress on the backend. So you have to synchronize that and then so one channel will write that. Right now my implementation is writing on the file system and other channel has to have that infinite loop and check for example every 200 milliseconds is there any new messages. So it’s not, you know, it’s not nice solution. PHP is not optimal platform for this kind of stuff. Okay, so this is something challenging.

Dave:
Fine. So you’ve got these two channels and that is so that you can keep giving like pushing messages or commands into WordPress and then like it takes some time to do that. And whenever the reply comes from that, it gets written to this file and then the outbound channel is reading that file and talking back to the user.

Tole:
Yes, sending messages actually. So this MCP is actually like it’s implemented over RCP 2 standard and actually and combined with that streamable response. So basically it’s sending a message.

Dave:
Okay. And you read those messages in your cloth.

Tole:
The application. So claw. The application reads that.

Dave:
Yeah.

Tole:
Response and then generates full nice response to you as a user.

Dave:
Cool. All right, well that sounds like. Sounds like non trivial challenge.

Dave:
Installs Convoworks and are prepared to do some workflow building. What can they do now with ConvoWorks that they could not do before?

Tole:
I like to troubleshoot issues with that. So for example, I have the chat again. There is a template. It’s a chat that has access to all WordPress, all PHP functions in that system So I can talk to it and try to fetch some data from the database. It will build query and it will execute WP database and it will execute query and it will tell me what are the results or I can ask it to fetch some posts and it can for troubleshooting. It’s a great thing what I’m doing right now for myself as an administrator and developer. But also I created a couple of implementations for my test customers and one of the examples was automated support ticket system. Okay, so customer had website and he has a community and each day he has a couple of support tickets and most of them are already answered. Everything is on the website. You know how it is. 90% of those are easily to handle. So we catch up the hook when the new ticket is created, we compare to database what we have already answered and we generate a new answer and insert answer in the. It was built with the Fluent support system.

Dave:
Okay. And so this is not a ticketing system that’s based on WordPress. Like the end user’s website doesn’t need to be on WordPress, they just need to build an interface to ConvoWorks and then ConvertWorks is using WordPress to store all of the conversations and feed back into the LLM to generate an answer.

Tole:
No, this is a full WordPress solution. So it’s a website with a fluent support fluence, CRM, CRM and that kind of stuff. And everything works in the WordPress, everything stays in the WordPress. Okay, we just use for example OpenAI as a model for agent and we are using Pinecon for storing and semantic search.

Dave:
Okay, fine. So when a new ticket comes into the WordPress database, then your workflow takes that support ticket and does a vector search against all of the previous tickets and answers, grabs the top five responses, chucks all of that OpenAI with a templated prompt that says here’s the question, here’s what we think are the best answers. Please generate a nice message to go back to the customer. It generates the response, answers the support ticket and that’s the kind of end to end process.

Tole:
Yes, there are a couple of more steps here, right? So we made it semi automatic. We generate a response, we schedule that, save that and we notify administrator.

Dave:
Right.

Tole:
And right now administrator actually has to click on the link. And interestingly that link leads to custom endpoint which is also built with the which just changes the status of that prepared response to approved and the scheduled job will catch it in the next run and will actually create a response to the ticket using the Fluent API. So everything in the process will be triggered like additional hooks or additional plugins that can add some additional things to the process.

Dave:
My brain wants to add a YOLO mode where the customer can say I want a faster response but it might be crazy. And then it just bypasses the administrator and get, they get a response straight away back but it might be a bad response.

Tole:
So for example, this is here also a thing because people have work time, right? You don’t actually maybe want to respond immediately in some cases.

Dave:
Yeah. All right, I feel like that’s a good demonstration of how Convo works here. You know, as basically as like an agent workflow for Fluent CRM and a support ticketing system based around that. So it sounds like this is reasonably early stages commercially. So are you like who do you want to start using this tool? Is this like a very developer focused tool? Are you going to go and pitch this to everyone that has Fluent CRM like, you know, or is this just for fun and really just you’ve built it for yourself and you’re not, you know, you’re not planning to get thousands of users for it.

Tole:
Yeah, as I said, I quite failed as a business businessman this time. I would like to stick more to the developer role converge right now. And for example, this Fluent implementation is nothing, you know, custom add to converge. It’s just you are clicking and writing in the gui. So it’s not something hard coded for the fluent, it can be any support system.

Dave:
So like if you’re running Gravity Forms for example and you could the same thing. Okay, that’s very cool. That’s very cool.

Tole:
Yeah. So you see when you are building no code tools there is those two opposite directions. One makes it simple but then you are very bounded. And if we all know each website has some special things, each customer have some, okay, we are using this plugin, but we are using that in different way or something like that. You always have to apply some additional customizations and when you have no code tool which allows unlimited customizations then it’s not that much easy from the beginning. So that’s why I like to call this more like visual programming tool.

Dave:
Do you have like a step before this where you can get an LLM to help you build the workflow in the first place? So you can say I’ve got Gravity Forms and I would like for when somebody asked me a question to compare it against the FAQ content on my site and then if the answer is there, just like send them a polite message. Could I just like vibe code, the workflow that then runs in convoworks right now?

Tole:
No, but this the direction I’m looking into.

Dave:
Cool.

Tole:
Exactly that. So this is. I can click and quite rapidly create prototypes with converge right now. But I would rather speak with the chat.

Dave:
You’d rather not have to click?

Tole:
Yeah. And so what’s actually interesting in this kind of approach, how it will look and what it implies, it’s actually like more like project management. And this is something what I’m actually already doing in my implementations with the converge. So I create some kind of block with some kind of functionality adding function for create new ticket or search tickets and that kind of stuff. Grouping together in one single block and adding some additional information about. So we are in the WordPress and when I need to stop store data into database, I will create custom post type. And in that instruction I tell to the bot system prompt I tell, okay, we are using this field in the options for authorization token, for example. We are using this custom post type with these custom fields for this kind of information. And so basically when you are doing that, you are describing the system some like product, like product requirements. And that is the idea to go into. To be able to nicely structure and save in the WordPress requirement specifications and then use that as a starting point for our AI to generate actually the complete solution. And later it. Generating the solution is the easiest part. You know, the hardest part is actually to maintain that later to update to test it. Is it working still not to break it with adding some other feature. So that’s why this product definition, product requirements and that kind of stuff. This is something we will have to look more in the future.

Dave:
Yeah, it’s like a description or like a supervisor agent. Right. Like I think my. Well, my theory of how the AI stuff starts to show up more is like a team of agents that work together. So you have like the supervisory agent, you have the like job creation agent, you have the like moderation or like whatever it is, you know the quality control agent. And they are just like working together.

Tole:
But that supervisor agent has to have some base document which will explain actually what they are doing.

Dave:
Yeah.

Tole:
And exactly what to expect from those sub agents that they should deliver. Right?

Dave:
Yeah, exactly.

Tole:
So it’s very, very, very interesting thing. I’m very excited and just waiting to grab some time to go more into that.

Dave:
So talking about time, how long did this take to build and what was your development process? Did you just go to cursor and say please build me an MCP for a WordPress site. How did you go about doing it.

Tole:
And how long did it take for MCP?

Dave:
Right, yeah, okay.

Tole:
It was just a couple of days and it was working quite easily and quickly. Some initial steps and of yes a lot of talking with the ChatGPT but since I’m implementing that in the converts which JGPT doesn’t know much, you have to adapt how you are talking with it. So actually I was talking about plain PHP WordPress implementations and then reusing snippets into what I’m actually doing. And there is also interesting thing when I was doing that MCP server actually I was pasting the full classes converts from the core and it was quite easy for it to understand and replicate and that was great experience.

Dave:
And have you. What do you. What’s your development environment like? Do you use Cursor or Windsurf or are you just on VS code or are you like super hardcore backend developer Notepad and I don’t know VIM or something like where are you on the scale?

Tole:
Most of the time I’m using ChatGPT so since the beginning I bought a premium and I’m using that quite skilled in that how to, you know, not to spend too much time on the copy pasting but I’m trying also in visuals as math code or also to use Klein which our copilot but I’m not skilled that much yet on that part.

Dave:
Well, OpenAI bought windsurf so if you haven’t tried that, maybe just like chucking the whole plugin into Windsurf would be interesting to see how it manages it as well. Obviously like you know, version control, take it back up because it’s also likely to mess stuff up. But yeah, I’ve been building various things for WordPress in that way. It’s been pretty interesting. So I just create like a virtual like a SIM link from my WordPress like local thing over here to the plugin folder that’s in my cursor project and then like it has access to the whole WordPress plugin without having to worry about the whole WordPress application. It works pretty well.

Tole:
I’m also using Synlinks to put my plugin into the WordPress and there is also one interesting thing I was doing actually with converge. There is that admin bot that we have template which can do everything and I was talking with it and it was actually reading files from my plugin and installation and created. Okay, here is your example. I was wrapping up WordPress rest APIs and I create this example with the posts endpoint and then I create. Then I said okay, read those files and create me the same for the users settings and whatever we had there. And it was just done that inside of the installation. Something like that. Windsurf is doing nice.

Dave:
You built your own Windsurf as well as convokes. That’s very smart. So did you run up against any limitations apart from the fact that PHP is not event driven and you’re having to read and write from the file system? Were there any other workarounds that you had to get this up and running?

Tole:
Not exactly. What you actually have to care is about how large those files are, how many those files are. And so you kind of like a log file, right?

Dave:
Easy to get out of control.

Tole:
Yes. And you have to care about that. So it’s still not that much user friendly that you can really talk however you want and ask forever whatever you want. You, you still have to have in mind those limitations when you are doing.

Dave:
That’S one of my biggest frustrations with the current generation of AI tools is like I will chat with Claude, will make a whole ton of progress and then you like run this final thing and it’s like conversation limit reached. And so then you have to start a new conversation and try and fold all of that understanding and learning in and then go again. Yeah. So anyway, I digress. So have you had to think differently at all about security considerations and user permissions and stuff? Like if you’re creating new endpoint, how does that show up? How do you lock that down? How do you keep a combo works running site safe?

Tole:
So when we are talking about this custom API builder, well everything works through WordPress and however WordPress is working, it’s a default infrastructure however it works. So if you are going to call that API endpoint from the jquery from the front end and your user is authenticated, this call will also be right automatically.

Dave:
So however you defined it, great, that seems very sensible. And so have you had much interest from the WordPress community? Have you heard like been talking to other people who want to use it? Do you have other people using it or is it primarily something that you’ve built for your own services delivery today?

Tole:
Yeah, I haven’t been much successful on spreading the word and that kind of stuff. I have couple of test customers and I’m shipping solutions for them and they’re quite happy what they have with that and what we can do. One of my colleagues, he’s actually creating APIs for his mobile games that Was nice implementation and he’s combining those API calls with the GPT. So he. It’s a bacon from his mobile games and applications.

Dave:
Nice.

Tole:
So yeah, right now everybody, you know it’s not easy to attract new users. We know all that and it’s not that much easy learning Core with the Convoworks it’s completely different than anything else. And yeah, it’s quite a slow process, but I’ll stick around and play with it.

Dave:
And are you using MCP elsewhere? Like does ConvoWorks have MCPs for other applications or are you using like for example, I run Obsidian and I use Claude desktop and there is an MCP for Obsidian that Claude can connect to. So when I’m having a conversation with Claude, it can also check my Obsidian notes to enrich its response. Do you use any other MCP stuff for combo works or in your own workflows?

Tole:
No, not actually right now.

Dave:
Cool. Lots of exploration to do. All right. And then I guess the last set of questions are about what comes next. So where do you think that WordPress and AI go from here? How do you see those two technologies intersecting?

Tole:
And you know, when I’m looking at those vibe coders, what they are doing with Vercel Replit and all that kind of stuff, I would say it’s actually pretty much their situation.

Dave:
So people vibe coding applications built on.

Tole:
WordPress that would be great if we end up with that.

Dave:
And do you think it’s likely?

Tole:
I’m don’t mull.

Dave:
Good answer.

Tole:
Okay. WordPress has that huge amount of the existing businesses and website. So this is a good thing and it will not just fade away just like that because people are thinking about business, not technology. It was the all the time. So we have some time here mcps and actually that feature API which brings WordPress more closer to AI’s is a very good step to keep us in the track in the in the game. But it would be not. You see, when you are using for example these MCPs, you are using clothing, you are using Windsor, you are using other things. So WordPress is just maybe if you are going to use it, it’s just a side player. So. So central role is in the cloud central. So what we are maybe missing here in the WordPress in the future, what I would like to see is something like MCP client which uses other things.

Dave:
Right.

Tole:
So why wouldn’t I have in my WordPress background process that searches the Internet through one MCP, generates images or something through another MCP whatever. Why WordPress wouldn’t do that? I Don’t want. I don’t like that things are executed somewhere else.

Dave:
Yeah, okay.

Tole:
But that was, for example, why we actually started with the converge and Alexa. Everything was outside. Why would that be outside? I want that everything in the WordPress. Here is the data, here is everything we have. I want that to run inside WordPress. So this is the idea behind Converse. It’s an O code tool, but not in the cloud, in your WordPress.

Dave:
That’s a great vision. And last question then. What advice would you give to other developers that are thinking about building an MCP or messing about with WordPress and MCP? What’s been interesting?

Tole:
Okay, it’s still early. MCPs are cool. But try just to use it.

Dave:
Okay.

Tole:
Just to use it. That’s enough for now.

Dave:
Okay.

Tole:
Use it for everything, for anything. Just to train yourself. Actually, it’s a tool when I. I remember earlier when new guys come to the company, so first thing we were learning, teaching them to use the Google. Right. You have to spend some time in learning the tool. Just that for now, everything else will come up by itself.

Dave:
I think that’s solid advice and thank you for giving us that accessible way into the wonderful world of mcps. Todd, it’s been really interesting to talk to you. I think ConvoWorks is really interesting. I love your vision for people being able to democratize their own AI engines and agents alongside their publishing and ecommerce. So, yeah, everything feels very aligned and I appreciate you taking time to talk us through it today.

Tole:
I’m very glad to be here.

Dave:
And if people want to get hold of you online, how do they get hold of you? Tole.

Tole:
It’s easiest way on the Twitter, but they can also send a contact message on the Converge website.

Dave:
And what’s your Twitter handle?

Tole:
Tole. T O L E, underscore C A R. Okay, perfect.

Dave:
And I hope that people will get in touch and there’ll be some fruit from your time today.

Tole:
Thank you.