Google moves its Docsy documentation theme to the Linux Foundation with opt-in AI agent support

A coding assistant is asked to call an API it has not used before. Before composing the request, it goes looking for the authentication instructions and the endpoint reference ^[7]. On a Docsy site with the new features switched on, it can consult an index, select the pages relevant to the task and fetch their Markdown versions. Those pages become context for the language model ^[6]. Anthropic calls this just-in-time retrieval ^[6]. Nothing is retrained. The agent reads the published docs during the task ^[7].

Teams tell themselves users land on the home page and click down through the sidebar. An agent has no use for the navigation and layout the HTML provides, and takes the explanations, headings, links and code examples in a simpler text format ^[8]. "You can add Markdown right alongside your documentation for tools that prefer that format," McKean said in her keynote ^[12].

The coverage frames the move in larger terms than the release itself does. "No one ever wanted to pay for humans to have good reference materials, but with AI agents, it's a different story," devops.com wrote in its report ^[14]. Docsy, built by Google's technical writers, is a theme for the Hugo static site generator, licensed under Apache 2.0 ^[2][9]. Its new home is the foundation whose projects, ranging from sandbox-stage efforts to Kubernetes, already use it, according to McKean ^[3]. The agent features are experimental and opt-in ^[4]. Docsy's release documentation says they help agents "discover and use your site content" ^[5].

So the evidence supports half of the shared-infrastructure story. The theme is shared: it now sits under the Linux Foundation, and the foundation's projects already rely on it ^[1][3]. Agent dependence is a forecast. The report does not say how many sites have enabled the features or whether agents answer more accurately when they do.

A Markdown copy also leaves the older problem where it was. "There is not a single person whose job it is to write documentation for the kernel," Jon Corbet, editor of LWN and overseer of Linux kernel documentation, has said ^[10]. At Linux Plumbers in 2026 he said that was still true ^[10]. An out-of-date page exported as Markdown is out of date in two formats, and devops.com notes that AI without good information is prone to garbage in, garbage out ^[15]. McKean's case starts with maintainers: better documentation cuts the questions users could have answered themselves ^[11]. "But when we're making technical documentation, it really doesn't matter how the information becomes useful to humans, as long as it does," she said ^[13].

If a team already runs Docsy, I'd turn the features on, since they are an opt-in on a theme it already has ^[4]. The tradeoff is that every page now reaches a reader that acts on what it finds. Before flipping the switch, sort pages on two axes. One is ownership: a named person has checked the page against the current release, or nobody has. The other is action: an agent would act on the page, as with authentication steps or endpoint references ^[7], or the page is background. Owned, actionable pages are ready, and owned background pages carry little risk. Unowned background pages can wait. The unowned, actionable quadrant needs fixing first, because an agent will build requests from those pages. If most of the API reference lands in that box, the project has the staffing gap Corbet described for the kernel ^[10].