This site was written from a terminal
No editor session was opened to build this website. Every page, every block, the brand tokens, the font slots, the contact form and the content type behind these articles were created by an agent calling tools over MCP.
Why that is possible at all
Every action the editor can take is also a tool. Not a subset chosen for a demonstration, and not a separate API written afterwards that drifts from the real one. The editor and the agent go through the same access checks and the same validation, because they go through the same code.
What the loop actually looks like
The useful discipline turned out to be narrow reads before writes. List the blocks on a page, fetch one subtree by path, validate the replacement with no side effects at all, then write exactly the node that changed. A whole page can land in a single all-or-nothing write when that is what you want.
Then look at it
The last step is the one that matters most. The platform renders the page through its own renderer and hands back a PNG, at desktop, tablet or phone width. An agent that cannot see what it made will keep producing things that validate and look wrong.
What it does not mean
It does not mean the editor is going away, and it does not mean you have to work this way. It means the surface is honest: if you can do it by hand, something can do it for you, and the checks do not change depending on which one is asking.
More from the journal.
Your account is the only thing standing between you and a site.
Create one, pick a starting point, and change anything on the page.


