Weaver

Problem

A small business needs the menu up to date, a post about the next event and an email to regulars. An AI agent can draft all of that, but most website builders and CMSs keep the site in their own format. AI models have seen little of it, so they invent syntax. They know plain HTML, Markdown and JSON well. And the owner still has to see each change before it goes live, fix the details and undo anything.

What I built

Developers already work with coding agents this way. The code is plain files, the agent proposes a change, a person reviews it and git keeps every version. Weaver gives a small business owner the same loop, without needing a developer.

• The site is a folder of files. Pages, texts, events and the brand live in HTML, Markdown and JSON in git. Weaver adds no format of its own, so Claude Code or Codex can work in the same folder without it.
• The brand is the source of truth. The owner describes it in plain terms, and Weaver turns it into a spec the agent reads and the styles the site uses. The website, posts and emails come from it, so they stay consistent. Prepared sections let the agent fill in content without touching the design.
• A full website, not one page. Several pages, collections like events, rooms and articles with their own pages, an inquiry form and more than one language.
• The page is the review. Each change shows on two full pages side by side, live now and after approval, with the changes listed in plain words. The owner pins comments on the page, and the agent comes back with a new version. Each proposal lists what the agent used, like a date from the events list, so a guess is easy to spot.
• The owner edits content by hand. Any text or photo can be changed on the page. Layout and styles stay with the brand.
• Every published version can be restored.

From builder to agent

It started as a web builder. I turned it into an agent: it keeps what Webflow or a CMS can do, but the main input is a sentence, and the screens are for checking the result.

Result

The demo runs the whole path, from a proposal through review to publishing and restoring an older version, on a phone too. Over 10,700 automated tests check what the coding agents write.

Next is a version stable enough for a pilot.