llms.txt

Off by default. Written when llms is an object with a title. llms: true alone writes nothing and logs a warning.

What reads it

llms.txt is a convention from llmstxt.org: one Markdown file at the host root that states what a site is about and which of its pages matter, so a language model does not have to crawl the site and strip HTML to find out. It is a convention, not a standard — no crawler is obliged to fetch it and nothing enforces the format.

That decides what belongs in it. The sitemap lists every page, for a crawler that walks the whole site; llms.txt is the short answer for a reader that fetches one file and stops. Name the pages a person would point someone at — services, contact, legal pages, the main articles — and leave the long tail to the sitemap.

Options

interface LlmsOptions {
  title: string
  description?: string
  details?: string
  sections?: LlmsSection[]
  /** Async functions that return additional sections, merged after `sections` */
  sources?: LlmsSource[]
}

interface LlmsSection {
  title: string
  links?: LlmsLink[]
}

interface LlmsLink {
  title: string
  url: string
  description?: string
}

type LlmsSource = () => LlmsSection | Promise<LlmsSection>

Option	Rendered as
title	# title
description	A blockquote under the title
details	A plain paragraph
sections	## title with a list of [title](url): description
sources	Sections from code, appended after sections

Where the link URLs come from

The file is fetched on its own and read as text, so there is no page it was linked from. A relative URL like /lohnmosterei/ has nothing to resolve against: it only works for a reader that already knows the host. The integration therefore puts the site URL in front of relative link URLs, the way the sitemap builds its <loc> values:

Configured url	site and base	Written to llms.txt
/lohnmosterei/	https://example.com	https://example.com/lohnmosterei/
/docs/	https://example.com, base /app/	https://example.com/app/docs/
https://other.example/x	any	https://other.example/x
mailto:info@example.com	any	mailto:info@example.com

These links point at content, like the sitemap’s entries, so sitemap.siteUrl replaces Astro’s site here as well and base is appended — see Site URL and base path. A URL that already carries a scheme (https:, mailto:, tel:) or starts with // is written exactly as configured.

With neither site nor sitemap.siteUrl set there is nothing to put in front. The link is then written as configured and the audit reports llms/relative-links (warn) listing the URLs it affects. The file still builds — its links just stay unusable outside the site.

Example

With site: 'https://example.com' in the Astro config, the links can be written as paths:

siteFiles({
  llms: {
    title: 'Example',
    description: 'An example website about TypeScript tooling.',
    details: 'Documentation for internal tools and workflows.',
    sections: [
      {
        title: 'Documentation',
        links: [
          { title: 'Getting started', url: '/docs/start/', description: 'Setup guide' },
          { title: 'API reference', url: '/docs/api/' },
        ],
      },
    ],
  },
})

# Example

> An example website about TypeScript tooling.

Documentation for internal tools and workflows.

## Documentation

- [Getting started](https://example.com/docs/start/): Setup guide
- [API reference](https://example.com/docs/api/)

Sections from code

sources run once, in astro:build:done. Each returns one section:

import { readdir } from 'node:fs/promises'

siteFiles({
  llms: {
    title: 'Example',
    sources: [
      async () => {
        const files = await readdir('./src/content/blog')
        return {
          title: 'Blog',
          links: files.map((f) => ({
            title: f.replace(/\.mdx?$/, ''),
            url: `/blog/${f.replace(/\.mdx?$/, '')}/`,
          })),
        }
      },
    ],
  },
})

getCollection() is not available in astro:build:done; read files directly. Links from sources are resolved against the site URL like the ones in sections.

Audit hints

llms/no-description, llms/no-sections and llms/sections-without-links (all info), plus llms/relative-links (warn). See build-time audit.