The State of AI in Craft CMS 5: Agents, Plugins, and AI-Ready Sites

A while back I wrote about practical AI integrations for Craft CMS — the kind you build yourself with a custom module, a Guzzle client, and an API key. That post still holds up. But the landscape around it has changed fast. What used to be "roll your own" is now a real ecosystem: dozens of plugins in the AI category on the store, agent tooling that lets Claude or ChatGPT operate a Craft install directly, and a whole conversation about making your site itself legible to AI.

I get asked about this constantly now — by clients, by agencies I subcontract for, and at meetups. "What's the deal with AI and Craft? What should we actually be using?" So this is my attempt to map the whole thing in one place, honestly, from the perspective of someone shipping Craft 5 sites every week.

The way I've come to think about it, AI touches Craft at three distinct layers. They get lumped together in marketing copy, but they solve completely different problems and carry completely different risks. Sorting them out is most of the battle.

Three Layers of AI in Craft

• Layer 1 — AI as a content assistant. Editor-facing features inside the control panel: alt text, summaries, drafts, translation, on-site chatbots. This is where nearly all the plugins live today.
• Layer 2 — AI as a developer and operator. Agents that build and manage Craft itself — creating sections and fields, editing templates, moving content — through MCP servers, CLI skills, and in-CP copilots.
• Layer 3 — AI as an audience. Making your published site readable and citable by the AI systems that now sit between your content and your visitors. This is the GEO conversation.

Layer 1 is mature and crowded. Layer 2 is the frontier and the most interesting. Layer 3 is the one most teams haven't thought about yet. Let's take them in order.

Layer 1: AI as a Content Assistant

If you filter the Craft plugin store to the AI category, the majority of what you'll find slots into this layer. It's the direct productization of the exact patterns I described building by hand last year. The clear winners by sheer count:

• Alt text generation. This is the crowded niche — there are easily a dozen plugins doing nothing but this: AI Alt Text, altpilot, Altomatic, Lens, and several that wrap third-party services like alttext.ai and AltTextLab. It tells you something: alt text is the perfect AI task. It's repetitive, editors hate it, accessibility and SEO both demand it, and a vision model does it well enough that a human just reviews. If you're on a fresh Craft 5 build and don't want to write the module, one of these will pay for itself in a week.
• Content generation and rewriting. GPT Content Generator, OpenAI Content Writer, Content Buddy, Loki, and AI Assistant add "generate/rewrite/expand" buttons next to your fields, usually with a choice of provider.
• Summaries. AI Summary and the summary field that ships with Sidekick turn an entry into a bullet list or SEO description on save.
• Translation. Multi Translator and Weglot handle multi-site translation, with the usual "machine draft, human approves" flow.
• On-site chatbots. Chatagent, Asyntai, and Angie Chat train a widget on your Craft content so visitors can ask questions in natural language.
• Search and vision. Dexter manages Algolia/Meilisearch indexing; Spectacles (one of my own plugins) uses computer vision for similar-image search across your asset library.

My honest take on this layer: it's genuinely useful, but it's also the easiest place to over-buy. Every one of these is an API call wrapped in a Craft UI. If you already have a custom module, you may not need the plugin at all. And installing five separate AI plugins that each phone a different provider is how you end up with five API keys, five bills, and no idea what's costing you money. Pick deliberately.

Layer 2: AI as a Developer and Operator

This is the part that actually excites me, because it changes how the work gets done rather than just bolting a feature onto an entry. The idea: give an AI agent structured, permissioned access to a Craft install so it can perform real operations — spin up a section, add an entry type, wire up a field layout, bulk-edit content — from a natural-language instruction. Three things are driving it.

MCP servers

The Model Context Protocol is the standard that lets an AI client (Claude Desktop, an IDE, ChatGPT) call external tools. Several Craft plugins now expose your site as an MCP server — Craft CMS MCP and Craft Editor MCP among them — so the agent gets typed tools like "create entry," "list sections," or "update field" instead of guessing at your database. You point your client at the server, and the model can operate the CMS end to end.

{
  "mcpServers": {
    "craft": {
      "command": "php",
      "args": ["craft", "mcp/serve"],
      "cwd": "/path/to/your/craft-project"
    }
  }
}

Once that's connected, "add a testimonials channel with a title, a quote field, and an author field, then create three sample entries" becomes a single prompt the agent executes against your actual install — with the results showing up in the control panel like any other change.

Skills and CLI agents

Happy Cog's craft-skill takes a different, lower-level route. It's a standalone PHAR — under a megabyte, no bundled dependencies — that bootstraps against your existing Craft install and exposes content operations as CLI commands returning JSON. It's packaged as a Claude Skill, so an agent can drive entries, sections, entry types, fields, drafts, assets, users, and even Commerce, across multiple sites.

# Drop the PHAR next to your project and point it at the install
chmod +x agent-craft
./agent-craft --path=/path/to/craft sections/list

# Commands return JSON, which is exactly what an agent wants to parse
./agent-craft --path=/path/to/craft entries/create \
  --section=blog --title="Draft from an agent"

The reason this approach is clever: because it's a self-contained binary that uses Craft's own autoloader, there's nothing to composer require and no plugin to keep in sync with your project config. You can hand an agent access to a site without modifying the site.

In-control-panel copilots

Then there's the fully integrated end: Sidekick from Double Secret Agency. It lives inside the Craft control panel as a chat window and edits templates, entries, fields, and sections through conversation. It's extensible with custom skills, ships an AI summary field, and — at least as of writing — runs on OpenAI with Claude and Gemini support planned. Plugins like CoPilot and Autofill occupy the same space, turning messy input into fully populated entries.

Layer 3: Making Your Site AI-Ready (GEO)

The third layer flips the direction entirely. The first two are about you using AI to work on the site. This one is about the AI systems your visitors already use — ChatGPT, Claude, Perplexity, Google's AI answers — reading your site and deciding whether to cite it. The umbrella term is GEO: Generative Engine Optimization.

The clearest writeup I've seen on this makes a point I keep coming back to: LLMs understand text, not HTML. When an agent fetches your page, everything that makes it a webpage — nav, scripts, cookie banners, the whole chrome — is noise it has to fight through. The single most effective technique is to serve a clean Markdown version of the page when an agent asks for one, via content negotiation:

# A normal browser gets HTML. An agent that asks for markdown
# can get a clean text version of the same URL.
curl -H "Accept: text/markdown" https://yoursite.com/some-article

Craft plugins like LLM Ready and LLMify automate exactly this — Markdown delivery plus an auto-generated llms.txt at your site root, which is essentially a sitemap pointing AI crawlers at your text content with descriptions. A nice touch from that same writeup: don't invent new fields for the descriptions, just reuse your existing SEO meta descriptions. Less duplication, same signal.

# Justin Holt — Craft CMS Developer

> Official Craft CMS Partner in Charlotte, NC. Practical
> articles on building, deploying, and tuning Craft CMS.

## Blog
- [Practical AI Integrations for Craft CMS](/blog/ai-integrations-craft-cms/): Building AI features into Craft with custom modules.
- [The State of AI in Craft CMS 5](/blog/state-of-ai-craft-cms-5/): A map of the AI ecosystem around Craft 5.

I'll be honest about the caveats, because the GEO space is full of overpromising too. No major AI provider has officially committed to reading llms.txt yet — it's a bet, not a guarantee. But it costs almost nothing to generate and maintain, and the underlying discipline is just good hygiene. If you want to see it in practice, this very site serves /llms.txt and /llms-full.txt, linked right in the <head> of every page.

So What Should You Actually Do?

Here's how I'd advise a client or agency partner sizing this up in 2026, depending on which layer the need lives in:

• If you want editor productivity (Layer 1): Start with the single highest-friction task — usually alt text — and add one focused plugin, or reuse a module you already have. Resist installing a suite. Keep a human review step on anything that gets published.
• If you're a developer who wants to move faster (Layer 2): Wire up an MCP server or craft-skill against a local Craft install and use it for content-model scaffolding and bulk edits. This is the most underrated productivity win in the whole ecosystem right now, and it stays safe as long as it's not pointed at production.
• If you care about being found by AI (Layer 3): Add an llms.txt and Markdown delivery, reuse your existing SEO metadata, and fix your semantic HTML first. It's cheap insurance against a shift that's already underway.

And the constant across all three, the thing I said in the last post and will keep saying: treat AI as an enhancement, never a dependency. The site has to work when the API is down. The generated content has to survive an editor's eyes. The agent has to operate inside guardrails. Everything good about this ecosystem is real; everything dangerous about it comes from skipping the human in the loop.

Where This Is Heading

My read on the near future: Layer 1 will keep consolidating — the twelve alt-text plugins will become three good ones — while Layer 2 becomes normal. Within a year I expect "connect your CMS to an agent" to be as ordinary as connecting a deploy pipeline, and content modeling to be something you increasingly describe rather than click through. Layer 3 is the wildcard; it depends entirely on whether the AI platforms formalize how they consume sites. But the sites that already write clean, structured, accessible content are positioned for it either way.

What hasn't changed is what makes Craft a good fit for all of this in the first place: a real content model, Project Config that makes changes reviewable, and a codebase you actually control. AI is a lot less scary when the platform underneath it is one you can reason about.

---

If you're weighing where AI fits in your Craft build — which layer you actually need, and which plugins are worth the money — let's talk it through. Sorting the useful from the hype is exactly the kind of thing I like helping teams figure out.