A website built with AI on top of a real CMS
A practical test of using AI agents with EmDash to build and maintain a real CMS-backed website, while keeping a human control surface and avoiding disposable AI-generated sites.

AI can build you a website now. Sort of.
That sentence hides the actual problem. Yes, an AI agent can generate a nice-looking page. It can write the code, create the layout, add copy, and give you something that looks like a website.
But a website is not just a snapshot. A real website has to change. It has to evolve with the business, author, or body of work it represents. The owner needs to edit it, correct it, publish new things, change the structure, and keep control after the initial build.
That is where a lot of AI website demos start to feel thin.
The WordPress tradeoff
I still use WordPress, and there is a reason for that. It is slow, tedious, and not built for an AI-agent world. But it gives me control: content types, admin screens, themes, plugins, and a way to maintain the site without rebuilding everything.
Most AI-generated websites land on the other side of the tradeoff. They are fast, but hard to maintain. Impressive for a demo, less convincing once a real business needs to update a page, add a content type, or fix something without calling a developer.
What I am interested in is the space between those two models: can AI help build a website that remains structured, editable, maintainable, and owned by the business or author it represents?
Why EmDash caught my attention
EmDash is a new open-source CMS from Cloudflare. It borrows some ideas that make WordPress useful: structured content, themes, plugins, and an admin interface. It is built on a modern stack with TypeScript, Astro, and Cloudflare's serverless infrastructure.
The part that made it click for me was not the feature list. It was the architecture around AI agents.
EmDash exposes tools, conventions, a CLI, skills, and an operating surface that agents can actually use. A normal CMS gives humans an admin panel. An AI-native CMS also gives agents a way to work inside the system.
To me, that is the core idea: use AI to build and maintain the site without making the site dependent on AI.
The human still needs a steering wheel
AI agents are non-deterministic. They will not do exactly what you want every time. The unacceptable risk is not that AI makes a mistake. The unacceptable risk is that the business owner cannot easily correct the result.
That is why the CMS matters. It gives humans a control surface. An AI-native CMS needs to give agents one too.
What I tested
This site is my current experiment with that idea.
It is built on EmDash and used to document what I am building around AI, automation, smart homes, self-management, and related systems. So far, almost everything about it has been made by AI agents: design, implementation, publishing, content types, taxonomies, and even article copy, such as this post.
That is the interesting part to me. Not that AI made a page, but that agents could operate through a real publishing system.
What is still unproven
I don't want to oversell it. EmDash is still early beta. It doesn't have WordPress's maturity, plugin ecosystem, or feature depth. I am watching two things closely: content modeling and visual design. A blog is one thing. A deeper information architecture for a real business is another. The design workflow still involves more manual trial and error than I would like.
Right now, I am still prompting every step. The agents are doing the work, but I am guiding the process. Despite the massive time savings already, this has not proven full automation or a reproducible workflow yet.
But it does feel like the right problem.
Why it matters
If this workflow becomes repeatable, it could matter for any organization that needs a website it can actually maintain.
For small businesses, the value is obvious. A real website still feels expensive because there is so much tedious work around design, structure, copy, revisions, and maintenance.
But this is not only a small-business problem. Larger organizations care about the same thing: lower costs, faster updates, fewer technical bottlenecks, and workflows that can adapt as the organization changes.
AI should be able to help with that. Not by generating disposable websites, but by helping create real websites that people and organizations can still own and maintain.
I do not know yet if this fully works. But it is the first version I have found that feels like it is testing the right thing.
OpenClaw Ops
A set of configuration, safety, and validation patterns for running agentic systems with less mystery and fewer footguns.
Dispatcher
A practical orchestration layer for triaging inbound tasks, generating issues, and handing implementation off to the appropriate agent or toolchain.