<
Enterprise content supply chains run 15 to 19 steps. Joel Varty, CTO of Agility CMS, thinks AI can get that down to six, so people are the only bottlenecks left. In Episode 47 of Beyond the CMS, Chris Bryce and Joel dig into the MCP server that connects Claude directly to your CMS and kills the copy-paste layer of content operations.
Joel has spent 21 years building Agility, the Canadian headless CMS behind Cineplex (20 years on platform), Scotiabank country sites, school boards, and higher education. He explains why Agility deliberately did not build an AI agent inside the CMS, and instead meets teams inside the AI they already use: Claude, Copilot, or Gemini.
The MCP server story runs in three phases: content models for developers, then a content pipeline that pushes up to 250 items per batch (point Claude at an old WordPress site and it walks the pages straight into Agility), and now bulk publish, unpublish, and delete. Joel's biggest surprise: people don't want control over their content as much as they want the work done.
Key topics
- Why the AI agent visits your website before the human does, and what that means for AEO
- Metadata vs JSON-LD schema, explained for CMOs
- Why Agility connects into Claude instead of building an agent into the CMS
- The Agility MCP server: models, content pipelines, 250 items per batch, bulk publishing
- Cutting the 15-to-19-step content supply chain down to six
- Cognitive offload and the AI slop test
- Content ops through the CLI: governance and audit logs for regulated industries
Chapters
- 0:00 Welcome to Episode 47
- 1:17 Agility CMS: 21 years, SaaS since 2018
- 3:45 What headless actually means
- 6:22 From SEO to AEO
- 8:30 Metadata vs schema: JSON-LD in plain language
- 10:55 Why Agility didn't build AI inside the CMS
- 13:58 Killing copy and paste
- 14:37 The Agility MCP server: 250 items at a time
- 16:18 Bulk publish, unpublish, delete via MCP
- 17:18 Connect Claude to Agility: mcp.agilitycms.com
- 18:01 The MCP ecosystem: Figma, Chrome, and thousands more
- 19:00 Blogging from a phone with a Claude project
- 20:17 Shrinking the 15-to-19-step content supply chain
- 21:57 Cognitive offload and the AI slop test
- 25:11 The MCP UI experiment
- 28:06 Content ops through the CLI: governance
- 31:21 The audit log problem with pure MCP
- 31:52 Wrap-up
Frequently asked questions
What is an MCP server? An MCP (Model Context Protocol) server connects an AI agent like Claude to another system, such as a CMS. Once connected, the agent reads and writes to that system directly, with no copy and paste in between.
How do you connect Claude to Agility CMS? Go to mcp.agilitycms.com for setup instructions for Claude, ChatGPT, Copilot, Gemini, and Cursor. Some are one-click installs that add the Agility connector into your AI agent.
What can the Agility CMS MCP server do? It started with content model creation for developers, added a content pipeline that accepts up to 250 content items per batch, and now handles bulk publish, unpublish, and delete. Full content workflows run end to end from your AI agent.
Can AI migrate a WordPress site into a headless CMS? Yes. Point Claude at the existing site and it walks the pages and loads them into Agility through the MCP server, replacing manual copy and paste or custom import routines. Review the result before publishing.
Why did Agility choose not to build AI inside the CMS? Organizations already have an AI strategy built around Claude, Copilot, or Gemini. Another agent inside a product forces teams to work in yet another place. Agility connects into the agent teams already use instead.
What is a content supply chain? Every step between an idea and published content. Enterprise content supply chains typically run 15 to 19 steps. MCP-driven automation cuts that to between six and ten, leaving people as the only bottlenecks.
What is the difference between metadata and schema? Meta titles and descriptions are open text fields that show up in search results. Schema, written as JSON-LD, is structured data that tells crawlers and AI agents exactly what a page is: an event, an article, a podcast episode, with defined fields for each.
What is answer engine optimization (AEO)? Structuring content so AI agents can find, understand, and cite it. Agents crawl your site when building their knowledge base and again when researching on a user's behalf, so the agent visits your website before the human does.
Why does structured content matter for AI discovery? A headless CMS translates structured content into schema automatically. Sites now see fewer visitors as people spend more time in AI agents, but the visitors who do arrive convert at a much higher rate, so being understood by agents is what fills the funnel.
When do content ops need a CLI instead of MCP? Regulated industries that need sign-off and audit logs. The Agility CLI runs inside GitHub Actions or Azure DevOps pipelines, moving content through dev, staging, UAT, and production with every step logged. Pure MCP operations leave no log.
Related from Dotfusion
We have built on Agility for 18 years, across three generations of the Brookfield Properties website and enterprise builds from manufacturing to commercial real estate. See our Agility CMS development services.
Links: Agility CMS https://agilitycms.com | Connect your agent https://mcp.agilitycms.com | Dotfusion https://www.dotfusion.com
Full transcript
Edited lightly for readability.
Welcome
We're live, and we're at Episode 47 of Beyond the CMS. Thanks to everyone making time during their lunch on Thursday, Eastern time, a little early on the West Coast. The whole idea here is to chat about trends in content platforms and content operations, grassroots conversation in an ask me anything format. If you're tuning in live, leave a comment on LinkedIn and we see it right on the interface in the show. If you're watching in the future, leave a comment on YouTube or send us an email and we'll be happy to get back to you.
Joel, you've been on the show a few times before, and it's always a pleasure to get your insights and ideas around the world of content. You're the CTO at Agility CMS, one of our favorite CMSs and one we've been using almost since the beginning of time when we look back. Thanks for joining us and making time to share your ideas.
It's always great to chat. So many things are happening, and yet so many things are the same, so it's always good to catch up.
Agility CMS: 21 years, SaaS since 2018
Let's assume a lot of folks listening don't know anything about Agility. Tell us about the product, a bit of its history, and your role over there.
I've been with Agility for 21 years, and the company has been around a couple of years longer than that. Our anniversary is coming up, the 23rd or 24th. We started as an agency building web applications and websites. Jon Voigt started it with Mike Assad, his partner at the time. I've been around since we started having a CMS product that we used to build websites on. In 2018 we went 100% SaaS, no longer implementing websites as an agency, just providing the headless CMS product. We're headquartered in Canada with a global customer base that's growing fast, across a really diverse set of customers.
We fit best in the small to medium enterprise space, with customers like Cineplex here in Canada and a few Scotiabank sites. School board sites work really well on the platform, and we were just talking about the higher education folks who have been on the platform a long time. An interesting one: we're coming up to the 20th anniversary of Cineplex being on Agility. Cineplex.com and all their properties, the Rec Room, Playdium, they're all on Agility, which talks to the power of the platform.
Maybe I'll introduce myself too. I'm Chris from Dotfusion. We build large-scale, mid-to-large enterprise websites on Agility, from manufacturing to commercial real estate. Not only is the tool accepted and used globally, but there's an emphasis on Canadian data sovereignty, and we love Agility as a Canadian product. Anyone listening, I really recommend reaching out to Joel and the team at Agility, especially if being in Canada matters for your data infrastructure.
What headless actually means
Tell me about Agility's architecture. And I have to ask for your definition of headless again, for my inventory of quotes.
Headless CMS goes by a bunch of terms that mean roughly the same thing: API driven, decoupled. Your content data is delivered via an API, so there's a disconnect between where your website or app is hosted and where your content is hosted. The net effect is really cool: your site is usually a lot faster, a lot more secure, and it's a lot easier to build things. You don't have to build a website; you can build an app, or build one thing to get started and have it expand over time, into multiple properties or multiple departments of your company. It's a really nice way to solve something quickly, and a lot of the time that's how we get in the door with a company.
That's what happened with Scotiabank. We don't have all of Scotiabank's websites; we have five or six of their country websites from the Latin American companies they purchased. It was really quick to set those sites up and build them out without completely reinventing everything within a pretty large organization. Headless is the agile, nimble way to work with content. It's why we call ourselves Agility.
For folks who don't know, we have a great case study where we moved the Brookfield Properties website through three different generations, all with Agility at the back end. When you have a headless setup, you can keep changing the presentation of your content without ever touching the back end. It's a great choice for keeping all of that content in one place, especially the stuff that lasts forever and persists.
From SEO to AEO
Another reason we love headless and Agility is the ability to build on Next and React, which produces very discoverable pages. The Dotfusion website is on Agility, and if you type in best headless agency in Canada, we come up first. One reason is luck, but there are also pragmatic, scientific things we've been doing, and Agility is an infrastructural part of that equation, as is the formatting and structure of the data. What are your observations around discoverable content and helping your clients get discovered because they're on Agility?
We've always talked about SEO as a top strategy for discoverability, but AEO is the newer term: being discovered by AI agents as well as search crawlers. Your content needs to be discoverable by the agents building their knowledge base. Every time there's a new version of Claude or ChatGPT, they go and build their knowledge base, and you want to get crawled as part of it so your content is part of the core knowledge for an agent.
But then that agent does research after that. It reaches out and talks to your website, basically doing web searches. It's not necessarily Google's crawler hitting your site fairly often as part of its crawl; it's a person's agent doing web research. The person isn't hitting your website yet, their agent is. So the programmatic discoverability of your content is really important, and that's why structuring your content in a headless CMS matters. It's actually really easy to do if you know the hoops to jump through.
Of course you still want to build a beautiful website for humans to look at. It needs to be fast, so when a person actually comes to your site they have a good experience. What we're finding is that fewer people are going to a lot of websites now, because more people are spending time in AI agents. But the people who do land on your website are way more likely to convert and to actually be interested in that content. The traffic is much more valuable when someone does come, so you really want to capture their attention.
Metadata vs schema: JSON-LD in plain language
We could talk for a while on structured data, but can you discuss the difference between metadata and schema data? We typically work with CMOs who aren't necessarily that technical and have way better things to do than to understand all of this. How does Agility enable publishing that type of content?
From a very high level, SEO and metadata are the metadata of a page: its title, its description, maybe the Open Graph image, which is a little more on the social side. Those are the things that show up when you search for content in Google, usually your title and meta description, sometimes something from the page content if it matched. They're very open-ended fields. They're just text, without a lot of structure.
But you can also add a much tighter schema onto each page that tells whoever is crawling that page, agent or crawler, exactly what that data is. Google led the way on this; it's called JSON-LD. It used to be called Rich Snippets. Essentially, it's a JSON schema for your page. You can tell an agent: this is an event page, it has these two hosts, Chris and Joel on the podcast, at this time. You can even have a carousel of images that represent it. If you have an article page, there's a specific article schema. These are well-known schemas that say: if you have this kind of content, these are the standard pieces of data that need to be in there.
With a headless CMS, you can translate your structured content to that schema automatically. Agility has tools right in the product to help with this, and more AEO and SEO tools are coming down the pipe in the not too distant future.
Why Agility didn't build AI inside the CMS
You mentioned it was Agility's decision not to pack too much AI inside the product, which is contrary to some of the other CMSs out there. I thought you had a really good reason. Where does AI fit into the content operations of an organization and its platform?
What we have observed at Agility is that many organizations are figuring out their AI strategy right now with a particular agent. Most folks are using Claude, which I really like, or Copilot; some are using Gemini. Those are the three big platforms, and they all have coding agents plus non-technical sides for regular users. Organizations are figuring out how to use those foundational pieces of their AI puzzle. Having another agent in your product that you want them to use doesn't work very well, because now they have to go to this other place. There's a place for that at some point, potentially, but there's much more value in saying: you have all of these things in your digital ecosystem, and Agility is one of them. How can we get Agility into Claude for you?
That's what we're spending the most time on now with folks. From a development point of view, it's showing them how to cut development time from hours and hours down to minutes. But it's also getting rid of the copy and paste. People are copying and pasting out of Word or Google documents into fields in Agility, and those fields are super powerful because they lead to schema-driven content on your website. How can we use AI to get content out of those unstructured documents, which are what people actually use to create, and into the structured pieces of Agility? Maybe we translate to different locales at the same time, or follow brand guidelines, or ask questions about the content if there's a problem with it. Hey, we don't use the word patient, we use the word customer. All of those guidelines, which are really hard to follow, can run as part of a content pipeline. Setting that up for customers, whether migrating content from old systems or running their regular content pipeline, has been really valuable, and they can do it as part of their existing AI strategy. They don't need a whole new strategy just for Agility's agents.
Killing copy and paste
I love that you tapped the I don't want to cut and paste idea. That's been the feedback of the year over here at Dotfusion: there must be a way to solve for this. And one major option is an MCP server.
The Agility MCP server: 250 items at a time
I think you've gone through a couple of evolutions of the Agility MCP server. Tell us where you've been, where you're at, and where you're going.
We initially built our MCP server to help developers, so that when I'm building a website I don't have to jump back and forth into Agility to update content models and component models, making sure all the field labels are good and everything has a field description, all that extra stuff developers hate doing. Just being able to create models, and that works amazing.
What we discovered is that people actually want to create content in Agility too. Not just developers, although developers want it too. How do I just get content in there? And a lot of times they want to translate it, follow brand guidelines, all that. So we built a pretty amazing content pipeline into it, and it's not one piece of content at a time. You can send up to 250 content items at a time through the MCP server. If I have a folder of Word documents, drop them into Claude and it goes through them like nothing. Or even importing an existing site: point Claude at the old WordPress site and say, go through all my pages and just add them into Agility. That works amazingly well. It's crazy how much time that used to take, copying and pasting or trying to write a routine.
It blows my mind how much people were willing to let Claude do all that work for them. I thought people would want more control over their content. Turns out they don't; they just want to get work done. They're okay with getting it on the website so we can look at it, and then review it after.
Bulk publish, unpublish, delete via MCP
We held off on publish and delete operations, unpublish and approve and all that, thinking people would want to go into the CMS to do those manual things. Turns out they want to automate that too. Based on customer demand, we recently added the ability to publish, unpublish, and delete content and pages in bulk as part of the MCP servers. You can write content workflows now that do the whole gamut: creating your website or your apps, putting the content into Agility, approving that content, and publishing it out to your different channels. All of that can be automated, and Claude is by far the best at it. But Gemini, ChatGPT, and Copilot can all hook into the MCP servers and use them to differing amounts.
Connect Claude to Agility: mcp.agilitycms.com
Agility has a marketplace of connectors and integrations. Asking from a layperson's perspective: how do I get Claude to connect to Agility?
mcp.agilitycms.com is its own mini website with the instructions for how to connect Agility to your agent. We've got all the different clients I just mentioned and a few others, like specific instructions if you're using Cursor for development. Some of them are one-click installs. What that does is essentially add the Agility connector into your AI agent.
The MCP ecosystem: Figma, Chrome, and thousands more
There are a gazillion connectors out there to connect into everything. There's a HubSpot connector that queries your HubSpot data. There's one that queries your commerce data. There are tens of thousands of MCP servers, and not all of them are that good, because you really have to work with the use cases people actually want. I had high hopes for Figma's MCP server, and as far as I've used it lately, it's basically slightly less good than taking screenshots from Figma. It does work, but I don't love it. Chrome has a pretty good MCP server that works really well, so you can control Chrome as a browser. And there are things like OpenClaw trying to automate your whole computer.
Blogging from a phone with a Claude project
The really powerful version of these things is stuff where you don't have to be there. I have Claude on my phone, my iPad, and my laptop. I can create blog posts on my personal blog from my phone just by narrating into it. I was just in Scotland, and the views from the hikes we were doing were amazing, so I'd take some pictures, open Claude, narrate what was happening, upload a few images, and say go. That was in a Claude project I had tuned to my blog: send it to this Agility instance, create blog posts, figure out the category from my content, and create nice galleries for the images. It's kind of like a skill. At the time we didn't have our publish capability, so I had to publish manually, but now it publishes as well.
The more MCP servers you have, the more capabilities you have. It does slow down your agent to a certain extent, so people are now understanding: I have to pick and choose which connectors I'll get the most value out of. As they fine-tune that, it's amazing how fast these workflows get.
Shrinking the 15-to-19-step content supply chain
We call this the content supply chain. Content supply chains in an enterprise typically have between 15 and 19 steps, which is insane. How can we get that to maybe ten steps, maybe even six? Then it's only the people that are the bottlenecks, and we don't have all of these technological bottlenecks. AI can really help with reducing the content supply chain and automating that pipeline.
We have a shared client, a rural telco out in Manitoba, and we were trying to solve for the amount of cutting and pasting they were doing. A typical blog post or news post can have up to 17 or 18 different attributes: read time, author, schema data, meta. That's a lot to be porting over across two different screens. So we love the idea of streamlining all of this as your AI model becomes your real interface to everything.
Cognitive offload and the AI slop test
We've got a research agent writing our posts that goes through a human to tune it. It raises an interesting question: if I tune AI for three weeks so it knows me, at what point is the writing from me anyway? It's saving me the time. It knows my preferences and writing style very well. That's just a good discussion.
The term for that is cognitive offload. Instead of using a rock to hammer in a nail, you use an actual hammer, and it gets a lot easier. Now I don't have to focus so much. Eventually I get an air nailer and I don't even think about it; the nail just goes in. Why wouldn't I use the air nailer if I'm building a house? If I have a rock, well, this is how we always did it, but I'm going to hit my hand on that rock and it's going to be super slow. Why wouldn't you use the tool if it's providing value? Now, if I have a whole new hammer and the house falls down after, probably not a very good hammer. So the question is: is it AI slop that it's creating? Does that content actually have any value, and is it at the same level of value as stuff you typed out? Once you've got it trained and tuned, you just can't type fast enough. You can have ideas that fast; you can't type that fast.
I say the same thing to developers holding back on coding tools. I understand you have your craft, but you can't type that fast. You can't even talk fast enough to get out the ideas. So think about the ideas and less about how fast you can type them out. It's a tool for getting your ideas out there, and ideas can come from so many places. Our creativity should not be bounded by technology, only by how many ideas we can have. I've changed my mind on AI several times, but I've seen it help people become so much more creative, because their ideas now have this tool to take them to the world in a much quicker way.
It's almost like more Lego. A hack on how I create content: I'll rebound an idea with AI. We're kind of a Claude shop here, but with whatever. Its pattern is to interview me back. What's important about this? What's my particularly unique angle? What would I say if I was talking to a pal out on a bike ride about it? It colloquializes the mission of writing the content. I acknowledge I'm not writing something with my best craft of creative writing skills; it's my best craft of getting AI to do what I want. So I get this interesting relationship, and a lot of chatting back and forth.
The MCP UI experiment
Staying on this MCP server idea: you mentioned you had dabbled with a bit of an MCP UI in Agility. Not all product ideas come to light, but I really love that you're constantly experimenting and trying new things. I don't think people even know what that idea is.
MCP UI is the idea that after you connect an MCP server, that server could actually serve up some user interface inside Claude, for instance. You haven't seen this take off, but when ChatGPT was doing apps inside ChatGPT, that's essentially the same protocol, slightly different. I think they're trying to unify the implementations between Anthropic and OpenAI, as well as Gemini. It hasn't taken off very well because the first implementations were for retail, and people didn't really like the idea of buying things there.
My thought was: when we're getting some content in, maybe we surface a little piece of UI to say, yes, I want to publish my content. Turns out people don't even want to click anything. They just want it to happen, and agents are pretty good at asking for permission anyway. I think it could be useful for other pieces of user interface, for uploading files specifically or doing a little preview. I was really big on it as a thing, but I don't think MCP UI is going to be that big a deal unless there are specific use cases it's really for.
Interestingly, the other day in ChatGPT I used the Apple Music app, which is essentially an MCP UI, to create a playlist of songs from the 70s, 80s, and 90s. It only allowed me 25 songs. It was okay, an interesting branded interface. I thought, this could be useful. I haven't used it since. The idea was so that people would buy stuff, but we just got my mom to trust buying things on a website. Buying them inside ChatGPT is not quite there yet. The e-commerce thing didn't quite happen; whether it will, I'm not sure.
Cognitive offload we're great with. Financial offload, not quite so much. Not yet.
Content ops through the CLI: governance
DevOps is complicated, and now you're getting feedback from your customers that they're also starting to create rather complex content ops, with Agility in the midst of it as the publishing conduit. What are you seeing out there?
There's a continuum of complexity. At one end, we have a school board with one instance of Agility, where each folder of the sitemap is a different school. Three people edit all the content, one deployed website for everything, almost no DevOps. That's as basic as it gets. All the way up to folks with one instance of Agility in a sandbox and pretty good DevOps workflows, or a dev, staging, UAT, and production code environment. You only need one or two instances of Agility to do that, and we've got a docs page that shows all these flows.
On the far end, you get folks that have content in all those environments as well. A lot of the time, this is when our customer is delivering a product that another customer then uses, a B2B2B situation. It's often in regulated industries, where content needs sign-off between different environments, and often matched not just to websites but to apps or both. They need sign-off on the code moving up through environments, and also on the content that moves with it.
You can use our CLI, the command line interface, to do that. In the same way they use a DevOps pipeline to push their code, they invoke the CLI to move content to different environments in the same or similar workflows. What's neat is that DevOps keeps all the logs of everything that comes out of that command line interface. Instead of someone doing it on a terminal on their computer, which you can do, it's a GitHub Action or an Azure DevOps action invoking it, moving content from one instance to another so somebody can sign off: yes, that's good to have on QA, move it to UAT or customer acceptance testing, all the way up to prod. Some customers do it the most basic way, quick, get it on the website. Others have to jump through all these hoops with the code so everything is verified all the way along.
The audit log problem with pure MCP
One of the many challenges of pure MCP content ops is there is no log. It's just magically there. That's great in most cases, but as you edge into the more enterprise use cases, financial data, gated content, stuff where much more governance is required, this is a great concept I haven't heard anyone else talking about: managing content ops through a more governed pipe. I call the MCP way the magic version, which is great. But for folks in the mid to large enterprise space listening, this is another great reason to choose Agility.
Wrap-up
Time goes by so fast once we get rolling. Thanks to everyone who checked in over lunch on Thursday, and folks listening in the future. Can't wait to chat again, Joel, and looking forward to seeing you soon and getting caught up. Great info and insights. If anyone's interested in Agility, we've been building on it for 18 years or so now. Reach out to Joel, Gina, and the gang over there, a wonderful service team as well. Get yourself a trial and get in there. Joel, thank you again. Take care.