The open-source Docsy project, created by Google to build documentation sites using the Hugo static site generator, is moving to the Linux Foundation at a time when technical documentation has become a direct source for AI agents, not just developers.
Erin McKean, a developer relations engineer at Google and a member of Docsy’s steering committee, announced the move during a talk she gave Wednesday at the Linux Foundation’s Open Source Summit Europe in Prague. The move is partly based on Docsy’s adoption across the foundation’s associated project ecosystem; by the end of 2024, approximately 2,200 projects were using it, including projects in the Cloud Native Computing Foundation such as Kubernetes, OpenTelemetry, gRPC, and Jaeger.
A documentation project turning toward an automated audience
Docsy was first announced in 2019. It is an open-source template designed for technical documentation and can be used with open or private projects. However, recent changes focus on how AI tools access content and on the form of instructions these tools need to carry out project-related tasks.
Since the 0.15.0 release in May, Docsy has been able to generate a Markdown version of every page alongside the usual HTML version. It can also generate an llms.txt file that provides an index of the site’s content. Both features are optional and experimental.
In version 0.16.0, released in July, an upgrade guide was added that was formatted so an automated assistant could follow it, including prerequisites, steps, and verification procedures. Version 0.17.0, released in August, added a hidden directive at the top of every page for users who enable llms.txt, directing visiting agents to the site index. This feature also remains experimental.
What changes in practice?
Docsy shifts part of the task of improving documentation from merely formatting pages for human reading to organizing content so models and agents can find, understand, and follow it. For projects that rely on contributors or external users, this could mean directing AI tools to a clearer version of the instructions instead of leaving them to infer the instructions from public HTML pages.
McKean says better documentation can reduce routine questions reaching maintainers because answers become available to users before a maintainer intervenes. This is a potential operational benefit for projects, but it does not in itself mean that agents will understand the content correctly or that the documentation will automatically become accurate.
Next step: measuring readability for agents
The Docsy team is working on what it calls AF documentation scores, meaning scores for how suitable documentation is for agents. The goal is to give maintainers a metric for evaluating AI tools’ ability to discover, navigate, and consume documentation, rather than relying solely on guesswork.
The source does not yet explain the form of this metric, its final criteria, or when it will launch. Markdown and llms.txt features and the directives associated with them are also optional and experimental, so their practical impact remains dependent on projects adopting them and on the quality of the original content they produce.