llms.txt v2: what an agent actually reads

llms.txt v2 is the latest revision of the llms.txt proposal, published in August 2026. It addresses the question of how an agent finds the llms.txt file for a given page. The llms.txt file is a Markdown file that serves as a curated entry point for agents to access LLM-friendly content on a site. It includes a name, short summary, context, and a list of links to relevant content. Jeremy Howard published the original proposal in September 2024, and llms.txt is meant for on-demand inference, not training.

In v2, the changes include allowing any path, with the most specific file taking precedence. The file can now be located at /docs/llms.txt or any subpath, and agents should use the most specific file if multiple exist. Two Markdown URL forms are introduced: appending .md to the path (e.g., page.html.md) and replacing the extension with .md (e.g., page.md). URLs without a file name fallback to index.html.md or index.md.

Discovery through link relations is another improvement, using rel=alternate type=text/markdown to point to the Markdown version of a page and rel=describedby to link to the relevant llms.txt file. This can be implemented using link elements in the page's head or as an HTTP header.

The optional header form of llms.txt is no longer special and is now just a convention for secondary links. The header format allows adding the link information to server or CDN configurations without modifying individual pages. The demo in the proposal showcases how an agent walks from a page to its llms.txt and Markdown version using these new rules.