llms.txt

llms.txt is a markdown index of a site: an H1, a > summary, and sections of - [title](url): description links, with /llms-full.txt for the whole content and a .md twin of each page (/docs/intro.md; the root and routes ending in / at <route>/index.html.md). The library reads it from your site and serves one generated from its site map.

#Reading your site's own

With llms.read: true (the default) the crawl first reads your /llms.txt, anonymously and through your app, never the hook's generated one:

• its title and summary describe the site in the prompt;
• its links — page routes or their .md twins, which are mapped back to the page route — seed the crawl and set the outline's order, in the order you wrote them; links under ## Optional come last. Their descriptions fill the outline before any page is rendered, so a site with a usable llms.txt has a complete outline on the first request;
• pages the crawl finds that your file does not list are merged in after your links;
• your file is used verbatim as the outline's base; over limits.outlineChars it is cut at the last whole link line, ## Optional first, and the cut is marked.

/llms-full.txt, when present, is split at its page headings straight into the page cache, so read_page needs no render for those pages. And with llms.preferMarkdown: true every page read asks for <route>.md first and uses it when the answer is markdown.

Every link goes through the route rule and your excludes. Absolute links must be on origin; without origin configured they are dropped (use relative links, or set origin).

#Serving one

Through the aide() hook, with llms.serve: true (the default), and only where your app answers 404:

Nothing is generated at build time and no file is written: each is the site map, serialised on request, with cache-control: public, max-age=<cache.ttlSeconds>. Excluded routes are never listed. Pages from pages.source are left out unless llms.serveSource: true.

A request to any of these paths counts against the visitor's warm-up bucket (rateLimit.warmups, 429 rate_limited past it) and starts the crawl if it is not running, waiting at most limits.crawlBudgetMs for it.

A site that already serves /llms.txt keeps it: the hook calls resolve first. llms.serve: false turns the serving off entirely. To mount the documents as your own routes, use llmsTxt, llmsFullTxt and pageMarkdown from @chat-aide/sveltekit/llms (Getting started → mounting by hand).