#Backstory
If you’ve been following along the past few years, you know we’ve been on a bit of a cost-cutting adventure. First, we ditched Help Scout for FreeScout and saved $3,000/year. Then we dumped ActiveCampaign for FluentCRM and saved another $2,700/year.
But bringing everything in-house has its challenges. You miss out on the shiny new features. If you want something, you need to build it yourself.
We created a FreeScout to WordPress integration at the end of 2023. We needed it at the time— but it took a lot of trial and error.
By 2025, it seemed like SaaS apps were adding AI features daily, while WordPress was stuck in the slow lane— we still did most things by hand.
#Our original support and development workflow
This is the same basic workflow that’s been in place at WP Fusion for over 10 years.
- Ticket arrives → Try to understand the problem.
- It’s a bug → Manually create a test to reproduce it.
- Start on a fix → Create an issue in the bug tracker, label, and assign it.
- Fix the thing → This part’s fine, we’re developers.
- Release the fix → Test it, write what changed, send out an update.
- Send reply → Tell the customer it’s been resolved.
- Document → Update the website, take screenshots.
Each step means switching apps and copy-pasting text. Even something simple takes at least 20 minutes.
A large part of your time starts to feel like it’s being wasted on repetitive processes when you really just want to be writing code and solving problems.
#The grass starts to look greener
Soon after we switched away from Help Scout they started adding some AI features, but it was pretty underwhelming. It could correct spelling and grammar, or adjust the tone of your messages, but not actually write any content or propose solutions.
But then in 2024 Help Scout added AI-powered drafts, 24/7 self-service AI support bots for beacon, and AI documentation assistance in the docs editor. We were starting to feel a little left behind.
Coming into 2025 several things started to align:
- WP Fusion saves you time via marketing automation— but we were still held up via manual processes.
- Everybody outside of WordPress-land was adding AI features to their apps, seemingly every day.
- The LLMs (like ChatGPT) were getting a lot better. They started creating mostly-usable code.
I started to wonder how AI could help us automate some of these time-consuming manual processes.
You can see the journey reflected in the plugins we’ve released this year.
#Phase 1: AI-assisted development (Spring 2025)
I first signed up for Cursor in November of 2024. It was really impressive at the time— a code editor that had AI built into it from the ground up.
It could intelligently complete lines of code you were writing, and fix simple bugs via a chat panel in the editor.
We were working faster using Cursor, but still working the same way.
In January 2025 it occurred to me that I could ask Cursor for solutions to little problems that had been bothering me.
I’d admired the fun and colorful badges some SaaS companies were using on their changelogs. I didn’t know how to do this for WP Fusion, and never had a spare day to figure it out.
I asked Cursor if it had any ideas— it offered to create a new WordPress plugin, following our existing plugin template, that would scan our changelogs for keywords like “Added”, “Improved”, and “Fixed” and automatically apply colored markers.
I clicked “Accept” and let it do its thing. It looked at the way we structured release notes in WP Fusion’s readme.txt and copied that logic over to the new plugin. A few minutes later we had a neat little changelog solution.
Normally, it’s too much work to publish internal tools. I like to share, but it takes extra time I don’t have. Cursor offered to publish the plugin to GitHub, I clicked “Accept”, and we released EDD Changelog Badges.
This kicked off a few weeks’ deep dive into what might be possible.
- My attempts to build an AI-powered support agent didn’t work. It still made stuff up and we didn’t end up using it, though DocsBot AI was released around the same time and it’s become really powerful.
- In March I used AI to build a custom analytics dashboard for WP Fusion. We still use that one today.
#Phase 2: AI-Assisted Workflows (Summer 2025)
A few months later, I was deep down the AI rabbit hole. A lot had changed. I go into it in detail in my post from July. TLDR we’d:
- Added Claude Code which could handle larger tasks than Cursor without user feedback.
- Crafted custom rules for AI code editors
- Written commands (i.e., fixing a GitHub issue by providing the ID)
- Integrated AI into automated testing
#Phase 3: Approaching Full Integration (Autumn 2025?)
August 2025 has been transformational in terms of how much we’ve been able to automate:
At a technology level we’ve incorporated:
- MCP connections for giving AI’s additional tools and resources 🔌
- Agents tasked with specific tasks 🤖
- Using AI to build new AI’s ♾️
Up until now, we were using AI you would use a chatbot— ask for some code, get the response, and test it like normal.
MCP servers vastly increase their capabilities, allowing things like:
- Searching the web
- Reading support tickets and customer history
- Creating, labeling, assigning, and closing GitHub issues
- Writing and running automatic tests
- Taking screenshots and writing documentation
All of these increase the amount of context an AI has to work with when it’s asked to do a job.
More context = better results.
Meanwhile, we use specially trained “adversarial agents” to critique any output. That means if one AI writes code, another one tests it for compatibility, accessibility, and security.
This significantly reduces the likelihood of “hallucinations”, where an AI agent confidently declares they’ve solved a problem when they haven’t done anything at all.
This becomes a self-perpetuating cycle of efficiency as the AI systems start to propose and then develop their own new AI systems.
#Changes to Our Support and Development Pipeline
#1. Instant acknowledgement
After a customer initially contacts us, an AI analyzes their message in the context of common setup errors, known conflicts, our plugin code, recent changes, and planned features.
The customer gets a reply letting them know that someone will respond soon, with personalized resources that might be helpful now.
#2. Faster replies
We created a Chrome extension that reads the customer’s message in the context of their license status, order history, active sites, and our documentation to draft replies with the click of a button.
#3. Automatic bug reports
We developed a module for our support inbox called FreeScout GitHub, which automates the creation of GitHub issues from customer support tickets.
#4. Automatic fixes
The bug reports contain a technical assessment of the problem and whether or not it needs an automated review. If it does, the issue is labeled claude implement.
This triggers the Claude Code bot via a GitHub action.
#5. Automated review
When Claude (or a human) is finished working, this triggers a third-party AI audit from CodeRabbit.
CodeRabbit provides feedback on any potential bugs or conflicts, with steps to test the new feature or fix.
The ticket is automatically reactivated in FreeScout, with a link to the summary by CodeRabbit.
#6. Frictionless testing
Everything needs a final check by a human before it’s released.
We use the Git Updater plugin to allow our support team to switch the plugin code between GitHub branches (i.e. work-in-progress features). Including those generated by Claude Code.
This means they don’t need to set up a development environment on their computer to test changes that haven’t been released yet.
#7. Continuous documentation
When a pending feature or fix is merged into the main WP Fusion plugin, another AI assesses it to determine whether it affects existing documentation or screenshots.
If so, it proposes updated documentation text, and identifies any images that will need to be replaced (for example in case of settings change in the admin UI).
If new screenshots are required, it uses Playwright to open a browser, navigate to the relevant screen, and take a picture.
While implementing this we ran into an annoying problem: the admin pages on our test websites are super cluttered (we test a lot of stuff 😅).
So we asked the AI to create a new browser extension to clean it up. It’s called WP Clean Admin. It lets you temporarily hide everything but the plugin you’re working on in the WordPress admin. Perfect for taking screenshots or screencasts.
#MCP FreeScout Server (The actual magic 🪄)
As a part of this push to better integrate AI into our product development workflow, we also created a custom MCP server for our help desk software, FreeScout.
This may sound unexciting, but it’s actually the most important part of the whole setup.
Traditionally, when you want to connect one web service to another, you use what’s called an API (Application Programming Interface).
This is what WP Fusion is doing when it connects your WordPress site to a CRM like ActiveCampaign.
MCP (Model Context Protocol) is a relatively new concept. It standardizes the way that systems like websites, apps, and services can communicate with AIs.
Without it, our current workflow would be a series of daisy-chained API integrations— from FreeScout to GitHub to Claude to CodeRabbit and back to FreeScout again.
This is difficult to set up and hard to maintain. If any one system breaks, the whole thing falls apart. It also makes it very difficult to change out any part of the process.
Having an MCP server for our support desk not only simplifies the process of connecting it to GitHub, but it also allows all of our other AI services to connect to FreeScout as well.
As an example, when I’m working on fixing a bug, I can ask Cursor to bring the customer’s support history into the conversation context. In many cases, this alone is enough to solve the problem
We also have an “orchestrator”, another AI, which sits in the middle of all these systems.
Because it connects via this same universal protocol, it can load data from any of our sites or services at any time. It can then identify areas for improvement or point out where the other AI’s have made mistakes.
The orchestrator and many of our other tools also connect independently to an MCP called Memory Service.
This serves as a big brain for everything that’s happening across our whole platform. When one agent solves a problem, the solution goes into the brain.
When another agent fails because it can’t access a resource (maybe it needs a premium plugin license to continue testing), this also goes in the brain.
This gives the orchestrator and the other services a real-time overview of everything that’s happening. They can see practical things like who is working today or whether a bug has been reported before, but they can also do things like invent new plugins, browser extensions, or even new MCP integrations autonomously, with the goal of improving our efficiency.
#This is exciting stuff
It’s exciting to be working with this technology. We’re using tools today that didn’t exist at all a few weeks ago.
Things are changing quickly in the world of AI, but I think it’s worth it to spend some time trying to see how it can help you. You don’t need a technical background to get started. Sit down for an hour with your favorite model and have a conversation about ways it thinks it could help you. You might be surprised what you learn 🤓
