nuxt-llms-txt

Make your Nuxt site readable by AI. A Markdown mirror of every page, plus a properly sectioned llms.txt index — zero config, and source-aware.

npx nuxi module add nuxt-llms-txt

export default defineNuxtConfig({
  modules: ['nuxt-llms-txt'],
})

That's it. Build the site and you get:

| Route | What it is | | --- | --- | | /about.md | Clean Markdown mirror of /about, with title/description frontmatter | | /about + Accept: text/markdown | The same Markdown, via content negotiation (Vary: Accept) | | /llms.txt | A sectioned index of the whole site, per llmstxt.org | | /llms-full.txt | Every page inlined, for one-shot ingestion (opt in with full: true) |

Every HTML page also advertises its mirror with <link rel="alternate" type="text/markdown"> and a matching Link: response header.

Why another one of these

There are two existing options, and neither does the whole job:

| | nuxt-llms-txt | nuxt-llms | @mdream/nuxt | | --- | :---: | :---: | :---: | | llms.txt | ✅ | ✅ | ✅ | | Per-page .md mirrors | ✅ | ❌ | ✅ | | Routes discovered automatically | ✅ | ❌ hand-written config | ✅ | | Publishes your original Markdown when a page has a source file | ✅ | n/a | ❌ HTML round-trip | | Sectioned index with real descriptions | ✅ | manual | ❌ flat list | | Per-locale llms.txt | ✅ | ❌ | ❌ | | Works on a plain Vue-pages app (no @nuxt/content) | ✅ | ❌ | ✅ |

The differences that matter in practice:

It publishes your source, not a round-trip. If a page came from a Markdown file, that file's Markdown is what gets served — MDC components, fence labels, tables and all. Converting rendered HTML back to Markdown loses every one of those. Pages that aren't Markdown-backed fall back to HTML conversion automatically, so a mixed site works either way.

::alert{type="warning"}
Keep your key on the server.
::

```bash [npm]
npm install @acme/sdk
```

Keep your key on the server.

```
npm install @acme/sdk
```

::alert{type="warning"}
Keep your key on the server.
::

```bash [npm]
npm install @acme/sdk
```

The index is organised. Sections are derived from your route tree, and a top-level segment only becomes its own heading once it has children — so a flat marketing site gets one clean list, and a docs site gets ## Docs / ## Blog:

# Acme Inc

> Acme builds tools for building tools.

## Pages

- [Acme Inc](https://acme.test/index.md): Acme builds developer tools that get out of your way.
- [Pricing](https://acme.test/pricing.md): Simple, transparent pricing for teams of every size.

## Docs

- [Documentation](https://acme.test/docs.md): Guides, API reference and recipes.
- [Getting started](https://acme.test/docs/getting-started.md): Install the SDK and make your first request.

## Optional

- [Terms of service](https://acme.test/legal/terms.md): The legal agreement between you and Acme Inc.

Descriptions are real. They come from useSeoMeta / og:description / frontmatter, falling back to the first substantial paragraph — and optionally from an LLM (see below).

How it works

.md files and llms.txt are written during prerendering, so a static build (nuxi generate) needs no server at all. For SSR routes that are never prerendered, a Nitro middleware renders and converts on demand — so /team/ada-lovelace.md works even though that page only exists at request time.

Content is isolated before conversion: <nav>, <header>, <footer>, <aside>, <script> and anything marked data-ai-ignore are stripped, so the Markdown is the page, not the chrome.

<template>
  <main>
    <h1>Pricing</h1>

    <div data-ai-ignore>
      <CookieBanner />
    </div>
  </main>
</template>

Content negotiation

Markdown is served from the HTML route only when the client explicitly asks for it and does not ask for HTML:

| Accept | Response | | --- | --- | | text/html,... (any browser) | HTML | | text/markdown | Markdown | | */* (curl, most HTTP libraries) | HTML |

The */* case deliberately keeps returning HTML — guessing otherwise would break ordinary API consumers. Disable the whole behaviour with negotiate: false.

Configuration

Everything is optional. Defaults shown.

export default defineNuxtConfig({
  modules: ['nuxt-llms-txt'],

  llmsTxt: {
    // Absolute site URL, used for absolute links in llms.txt.
    // Inferred from `site.url` or NUXT_PUBLIC_SITE_URL when omitted.
    domain: 'https://example.com',

    // llms.txt header. `title` falls back to `site.name`.
    title: 'Example',
    description: 'What this site is.',
    notes: 'Free-form prose under the summary.',

    // Which routes to map. `exclude` is *merged* with the built-in defaults
    // (/api/**, /_**, /200, /404), so you never have to re-list them.
    include: ['/**'],
    exclude: [],

    // Routes listed under the conventional `## Optional` heading.
    optional: ['/legal/**'],

    // Override the derived sections entirely.
    sections: [
      { title: 'Documentation', match: '/docs/**', description: 'Guides and API reference.' },
      { title: 'Blog', match: '/blog/**' },
    ],

    // Content isolation.
    mainSelector: ['main', '[role="main"]', 'article', '#__nuxt', 'body'],
    stripSelector: ['nav', 'header', 'footer', 'aside', '[aria-hidden="true"]', '[data-ai-ignore]'],

    // Outputs.
    markdownFiles: true,   // write /foo.md next to every prerendered page
    full: false,           // also emit /llms-full.txt — or { maxChars: 400_000 }

    // Runtime behaviour.
    runtime: true,         // serve .md for SSR routes
    negotiate: true,       // honour `Accept: text/markdown`
    linkAlternate: true,   // <link rel="alternate"> + Link: header

    // Prefer original Markdown sources. Auto-detected with @nuxt/content;
    // point it somewhere else with { dir: 'docs' }, or turn it off.
    source: true,
  },
})

AI-written descriptions (opt in)

Pages with no description can have one written for them at build time.

llmsTxt: {
  ai: {
    provider: 'anthropic',              // or 'openai'
    model: 'claude-opus-5',             // default; 'gpt-4o-mini' for openai
    apiKey: process.env.ANTHROPIC_API_KEY,
    only: 'missing',                    // or 'all' to rewrite every description
    concurrency: 5,
  },
}

Three things make this safe to leave in a shared repo:

• No key, no calls. Without apiKey the step is skipped entirely, so contributors and CI build the site without an account.
• Results are cached in node_modules/.cache/nuxt-llms-txt/ keyed by a content hash, so a rebuild costs nothing and only edited pages are re-billed.
• Failures degrade. A rate limit or refusal leaves the existing description in place and logs a warning; it never fails your build.

Install whichever SDK you use — they are optional peer dependencies:

pnpm add -D @anthropic-ai/sdk   # or: openai

Or bypass both providers entirely:

llmsTxt: {
  ai: {
    summarize: async page => myOwnSummarizer(page.markdown),
  },
}

i18n

With @nuxtjs/i18n installed, locales are detected automatically. Each gets its own index at /{locale}/llms.txt listing only that locale's pages, and the primary index cross-links the others under ## Optional. Locale prefixes are stripped before sections are derived, so /fr/docs/intro lands under ## Docs alongside its English twin — not in a section called "Fr".

Hooks

Both run at build time, after prerendering.

export default defineNuxtConfig({
  hooks: {
    // Mutate the page corpus: drop pages, rewrite Markdown, fix titles.
    'llms-txt:pages'(pages) {
      for (const page of pages) {
        if (page.origin === 'html') page.markdown = tweak(page.markdown)
      }
    },

    // Mutate, add or drop output files just before they are written.
    'llms-txt:files'(files) {
      files.push({ path: '/llms-extra.txt', contents: '...', kind: 'llms.txt' })
    },
  },
})

Using the core without Nuxt

The conversion and llms.txt machinery is framework-free and published separately, for scripts, CI checks or another framework's adapter:

import { buildLlmsTxt, htmlToMarkdown } from 'nuxt-llms-txt/core'

const { title, description, markdown } = htmlToMarkdown(html, {
  domain: 'https://example.com',
})

const index = buildLlmsTxt(pages, { title: 'Example', domain: 'https://example.com' })

Notes and limits

• llms.txt needs prerendering. The index is built from prerendered pages, so a pure-SSR build with no prerendered routes produces none. Set nitro.prerender.crawlLinks: true (or list routes) to get one. Per-page .md mirrors do not need this — the runtime middleware covers them.
• Content source mapping follows @nuxt/content's default path rules (numeric 1. prefixes stripped, index folded into its parent). Custom source prefixes may not map; override individual routes with sourceMap, or set source: false to always convert from HTML.
• Pages marked noindex are excluded from both the index and the mirrors.

Contributing

pnpm install
pnpm dev:prepare      # generate module + playground types
pnpm dev              # playground at localhost:3000
pnpm verify           # lint + types + tests + build

Two playgrounds cover both paths: playground/ is a plain Vue-pages app, playground-content/ is a @nuxt/content site.

License

MIT