next-with-text

llms.txt for Next.js — in 3 simple steps.

npm install next-with-text

// next.config.ts
import { withText } from "next-with-text";

export default withText(nextConfig);

# .gitignore
**/app/%5Fllms/

That's the entire integration. Your app now serves:

• /llms.txt — a spec-compliant llmstxt.org index of your pages, with real titles and descriptions
• /llms-full.txt — every page's full content as markdown, in one file, for bulk ingestion
• Accept: text/markdown — agents that ask a canonical URL for markdown get markdown; browsers are unaffected
• /<route>.md — a markdown twin of every page (/about → /about.md)

No route handlers to write, no middleware, no config files, no route lists to maintain. It reads your build output — not your source tree — so nothing it needs disappears in a serverless bundle. ChatGPT, Claude, Perplexity, and coding agents can read your site the way they want to.

Building locally leaves nothing behind: the generated files exist only for the deploy build that needs them (details).

Requires Next.js 16+ and the App Router.

How it works

withText hooks into the Next.js build and generates a private catch-all route. From there, the package has two paths that produce the same markdown:

1. At build time, it reads the prerendered HTML in .next/server/app, discovers the concrete routes Next actually emitted, converts each page to markdown, and writes llms.txt, llms-full.txt, and the per-page .md files into the deploy artifact. Metadata and export const md overrides are applied before the files are grouped and rendered.
2. At request time, the generated route serves the same surfaces when a static file is unavailable, including during development. For a dynamic .md request it evaluates the page's md export when present; otherwise it fetches the page from the same deployment, forwards the visitor's cookies, and converts the returned HTML to markdown.

This split is also the auth boundary. The static path only sees pages Next prerendered, then removes routes covered by compiled proxy matchers, so private rendered HTML never enters a deploy artifact. A dynamic request renders with the caller's own session and returns Cache-Control: private, no-store; a gated page reaches the public indexes only when its source explicitly exports publishable md metadata or content.

Options

Zero config is the intended config — the same one line works unchanged in every app you ship. When you need more:

export default withText(nextConfig, {
  md: true, // default true — the .md twins + Accept negotiation
  include: ["**/*"], // route-path globs (routes, not file paths)
  exclude: [], // excluded routes vanish from every surface: both indexes,
  // no .md file, and the on-demand route 404s them
  llmstxt: (ctx) => "…", // optional: its return value IS the entire llms.txt body
  // ctx is { title, description, sections: [{ title, routes: [{ title, description, href }] }] }
  llmsfulltxt: (ctx) => "…", // optional: owns llms-full.txt with the same ctx shape
  // every route in llmsfulltxt also has its finalized publishable content
});

llms.txt and llms-full.txt are generated unless your app already owns that route. The index groups pages by first path segment (/docs/** → ## Docs), root pages listed first — no section config to maintain:

Auth safety, by construction

By default, pages behind auth never leak into the static output — the generation mechanism can't publish them. Listing one is possible, but only as a deliberate opt-in written into the page itself.

• The static surfaces (llms.txt, llms-full.txt, the .md files) are generated only from HTML that next build prerendered. An auth-gated page reads cookies() or headers(), or redirects — so Next marks it dynamic and emits no build HTML for it. There's nothing to publish, so nothing leaks; there is no exclude list to forget.
• Routes guarded by your proxy.ts (middleware) are excluded too, even when they're statically prerenderable: withText reads the compiled matchers from the build output and drops every matching route from all static surfaces. Also automatic — a matcher: ["/admin/:path*"] means no /admin in either index and no published admin.md.

Auth-gated dynamic pages still get .md and Accept negotiation, safely: they're converted on demand by fetching the page with the requester's own cookies, so /account.md renders exactly what /account would show that visitor. Those responses are sent Cache-Control: private, no-store, so a CDN that caches without varying on Cookie can't hand one visitor's page to the next.

Listing a gated page on purpose

Sometimes you want an agent to know a page exists — an account area, a guarded reports section — without publishing what's behind it. A page opts in by exporting md with a title and no content:

// app/account/page.tsx
export const md = {
  title: "Your account",
  description: "Billing, plan, and settings. Requires sign-in.",
};

That entry appears in llms.txt; llms-full.txt carries the same metadata plus a [Requires session] link instead of a body; no account.md file is written, so /account.md keeps rendering live with the caller's cookies. Add a content alongside the title and that text becomes the published body — useful for a guarded page you want to describe properly.

The safety property is unchanged, because the opt-in is an allowlist and everything it publishes is text you wrote in the page file — the page's rendering is never the source. Three things hold it in place:

• Only literal routes. /users/[id] can't opt in: there are no concrete URLs to publish without enumerating your customers.
• exclude still wins. A route matched by an exclude pattern stays dark no matter what it exports, so config remains a reliable kill switch.
• It's never quiet. Every build prints the routes that opted in by name: 2 gated route(s) opted into the index via md export: /admin/reports, /gated.

# Acme

> Payments infrastructure for platforms.

- [Pricing](/pricing.md): Plans and per-transaction fees.

## Docs

- [Getting started](/docs/getting-started.md): Install and make your first charge.

Per-page override: export const md

For pages whose rendered HTML converts poorly, export md from the page file. It replaces the converted content for that page everywhere (.md file, llms-full.txt, Accept negotiation):

// app/docs/setup/page.tsx
import type { MarkdownPage } from "next-with-text";

export const md: MarkdownPage = `# Setup\n\nHand-written markdown for this page.`;

All of these work:

// plain string
export const md: MarkdownPage = "…";

// object — title/description also replace this page's entry in both indexes
export const md: MarkdownPage = { title: "…", description: "…", content: "…" };

// function, sync or async, with typed params for dynamic routes
export const md: MarkdownPage<"/tags/[tag]"> = async ({ params }) =>
  `# ${(await params).tag}`;

Functions also receive searchParams, just like standard PageProps.

Custom index: the llmstxt function

If the default template isn't right, take over the whole file:

llmstxt: ({ title, description, sections }) =>
  [
    `# ${title}`,
    `> ${description}`,
    ...sections.flatMap((section) => [
      "",
      section.title && `## ${section.title}`,
      ...section.routes.map(
        (r) => `- [${r.title}](${r.href}): ${r.description}`
      ),
    ]),
  ]
    .filter(Boolean)
    .join("\n");

sections is the final route list, grouped the way the default index groups it — filtered, auth-excluded, with any md overrides applied. The first section has an empty title: those are the root-level pages the index opens with, before any heading. A section is only present when it has routes, so you never have to guard against empty ones. Regroup them however you like; this only affects llms.txt.

Custom full text: the llmsfulltxt function

The full-text callback mirrors llmstxt: it owns the whole file and receives the same title, description, sections, route grouping, and ordering. Each route also carries its finalized content:

llmsfulltxt: ({ title, description, sections }) =>
  [
    `# ${title}`,
    `> ${description}`,
    ...sections.flatMap((section) => [
      section.title && `## ${section.title}`,
      ...section.routes.map(
        (route) => `${route.content}\n\n[Source](${route.href})`
      ),
    ]),
  ]
    .filter(Boolean)
    .join("\n\n");

content is exactly what the default llms-full.txt would publish for that route: frontmatter is removed, md content overrides rendered HTML, gated opt-ins become safe stubs, and excluded or non-publishable routes never reach the callback.

Generated files clean up after themselves

Building locally leaves your working tree exactly as it was. Nothing to gitignore, nothing to review, no diff noise:

• app/%5Fllms/ — the two-line route file that powers on-demand conversion. It's written when the build or dev server starts and deleted when that process exits (Next's typegen reference to it is scrubbed too, so your editor never shows a dangling import). While next dev is running the file stays put, and a next build in another terminal won't pull it out from under the dev server.
• public/llms.txt, public/llms-full.txt, public/<route>.md — the static tier, written only when the build is a deploy build. Locally they're never written, and any left over from an earlier deploy build get reclaimed.

A deploy build is one where CI or VERCEL is set — true on Vercel, GitHub Actions, and essentially every CI runner. If you build the artifact you actually deploy somewhere those aren't set (a self-hosted box, a release script on your laptop), open the gate by hand:

NEXT_WITH_TEXT_STATIC=1 next build

The same variable set to 0 forces local behavior anywhere, including CI. With output: "standalone" the files are also written into .next/standalone/public on every build — that tree is the deployable artifact, never your working copy — so a standalone build shouldn't need the override. That path hasn't been verified on a real standalone deploy yet; set NEXT_WITH_TEXT_STATIC=1 if you'd rather not rely on it.

Local next start serves every surface identically — the on-demand route covers what the static files would have. It's a request-time HTML conversion rather than a file read, so it's slower than production, and it's the only difference you'll see.

Files you edit by hand are never deleted: pruning removes a file only when its contents still match what the last build wrote.

Existing surfaces always win. A file such as public/llms.txt, or an exact App Router handler such as app/llms-full.txt/route.ts or app/about.md/route.ts, makes next-with-text skip that output entirely — it doesn't render the body, invoke its customization callback, write a static file, or install a rewrite over your route. This applies independently to both indexes and every per-page .md twin. Files recorded in the build manifest are still next-with-text outputs, so later builds update them normally; editing one by hand transfers ownership to you. A custom per-page .md route remains linked from llms.txt, and llms-full.txt references it instead of publishing a second rendering of the underlying page.

If a dev server is killed with SIGKILL (or your machine loses power), app/%5Fllms/ can survive — the next build or dev run clears it.

Why this exists

AI assistants increasingly decide what to cite by what they can read. Next.js has no first-class way to serve LLM-consumable content, so sites either hand-roll markdown endpoints or reach for tools that scan the source tree and serve title-and-description stubs at best — and a stub is indistinguishable from a broken endpoint to an agent trying to read the page.

next-with-text derives everything from your build. It walks the HTML that next build actually rendered — generateMetadata, generateStaticParams, MDX, whatever produced the page — and converts it to clean markdown: headings, code fences, images as absolute-URL references. Dynamic routes like /blog/[slug] show up as concrete URLs (/blog/hello-world), not patterns. If a page prerenders and isn't excluded, it's in the index.

Every surface is asserted by a test suite that builds a real Next app — file-level checks on the build output, HTTP checks against next start and next dev.

What it's not for

• Pages Router — the mechanism reads App Router build output; there's no Pages Router support and none planned.
• Fully hand-curated indexes — if you want a bespoke, hand-written llms.txt, put a static file in public/; you don't need a library for that. The llmstxt function covers the middle ground: your template over auto-discovered routes.
• Markdown authoring — export const md is an escape hatch for pages that convert poorly, not a CMS. If most of your pages need it, this is the wrong tool.

Development

bun install
bun run test   # builds the library, then runs the verifier suite against tests/fixtures/next

The test suite builds a fixture Next app, inspects the build output, and asserts over HTTP against next start and next dev — including auth exclusion, cookie-forwarded on-demand conversion, and Accept negotiation.

License

MIT