llms-txt-check

Validate a site's llms.txt against what the site actually serves.

Your llms.txt generator promises AI tools a map of your docs. This tool verifies the deployed site honors the map, and fails your CI when it stops.

npx llms-txt-check https://your-docs-site.com

Why a generator isn't enough

Generators run at build time. Breakage happens at the serving layer: host redirect rules, trailing-slash normalization, docs restructures, domain migrations. Your build can be green while every AI agent reading your llms.txt gets 404s.

Real examples found with this tool in August 2026:

• Drizzle ORM's llms.txt listed 71 URLs that returned 404 in production (16% of the file, the entire PostgreSQL section).
• Cursor's docs moved domains, and the old docs.cursor.com/llms.txt started redirecting to an HTML marketing page. Agents holding the previously correct URL now ingest a React shell.
• LiteLLM's llms.txt had dead links too, including its own intro page.
• A Docusaurus site once served a zero-byte llms.txt thanks to a trailingSlash config interaction with its host.

Every one of these sites had a green build while serving a broken llms.txt.

What it checks

1. Discovery. /llms.txt exists, returns 200, and is markdown rather than an HTML app shell.
2. Spec lint. One H1 title, well-formed - [title](url): description entries, no duplicate or relative URLs, no staging domains baked in.
3. Serving validation. Every listed URL is fetched against the live site and checked for HTTP errors, zero-byte and near-empty bodies, and .md URLs that return HTML.

Usage

# Check a deployed site (fetches <url>/llms.txt)
npx llms-txt-check https://docs.example.com

# Check a llms.txt URL directly
npx llms-txt-check https://docs.example.com/llms.txt

# Lint a local file without network checks
npx llms-txt-check ./static/llms.txt --lint

# Spot-check 50 URLs spread across a large file
npx llms-txt-check https://docs.example.com --sample 50

# Machine-readable output
npx llms-txt-check https://docs.example.com --json

Exit code 0 means healthy, 1 means problems were found, 2 means the tool itself could not run. That makes CI integration one line:

# .github/workflows/deploy.yml, after your deploy step
- run: npx llms-txt-check ${{ env.DEPLOY_URL }}

Library API

The parser and checks are exported for programmatic use, with zero dependencies:

import { parse, lint, checkSite } from "llms-txt-check";

const doc = parse(text);
// { title, summary, preamble, sections: [{ name, links: [{ title, url, description }] }] }

const issues = lint(doc, { origin: "https://docs.example.com" });

const report = await checkSite("https://docs.example.com", { sample: 50 });
console.log(report.failures);

Options

| Flag | Default | What it does | | --- | --- | --- | | --sample <n> | all | Check at most n URLs, spread evenly across the file | | --concurrency <n> | 8 | Parallel requests | | --timeout <ms> | 15000 | Per-request timeout | | --lint | off | Skip network checks, lint the file structure only | | --json | off | Machine-readable report |

License

MIT