New:Microsoft Teams Notifications Are Now Available in Socket.Learn more →

markdown-for-agents

Dependencies

Maintainers

Versions

Alerts

File Explorer

Advanced tools

License

Install Socket

Detect and block malicious and high-risk dependencies

Install

markdown-for-agents

Runtime-agnostic HTML to Markdown converter built for AI agents. One dependency, works everywhere.

latest

Source

npm

Version: 1.3.4

Version published: 6 months ago

Weekly downloads: 1.1K

Maintainers: 1

Weekly downloads

Created: 7 months ago

Source

markdown-for-agents

Runtime-agnostic HTML to Markdown converter built for AI agents. One dependency, works everywhere.

Try it in the playground — paste a URL or HTML and see the conversion live.

markdown-for-agents

Audit any URL — no installation required:

npx @markdown-for-agents/audit https://docs.github.com/en/copilot/get-started/quickstart

           HTML            Markdown        Savings
───────────────────────────────────────────────────
Tokens     138,550         9,364           -93.2%
Chars      554,200         37,456          -93.2%
Words      27,123          4,044
Size       541.3 KB        36.6 KB         -93.2%

Convert any HTML page into clean, token-efficient Markdown — with built-in content extraction to strip away navigation, ads, and boilerplate. Inspired by Cloudflare's Markdown for Agents.

Features

Runtime-agnostic — Node.js, Bun, Deno, Cloudflare Workers, Vercel Edge, browsers
Frontmatter — automatically extracts title, description, and image from <head> and prepends YAML frontmatter
Content extraction — strip nav, footer, ads, sidebars, cookie banners automatically
Content-signal header — opt-in content-signal HTTP header for publisher consent (AI training, search, AI input)
Framework middleware — drop-in support for Express, Fastify, Hono, Next.js, and any Web Standard server
Content negotiation — respond with Markdown when clients send Accept: text/markdown
Token estimation — built-in heuristic token counter for LLM cost planning, with support for custom tokenizers
Plugin system — override or extend any element conversion with custom rules
Single dependency — only htmlparser2 (no DOM required)
ESM only — modern, tree-shakeable, with subpath exports
Fully typed — written in TypeScript with complete type definitions

Install

npm install markdown-for-agents

Quick Start

import { convert } from 'markdown-for-agents';

const html = `
  <h1>Hello World</h1>
  <p>This is a <strong>simple</strong> example.</p>
`;

const { markdown, tokenEstimate, contentHash } = convert(html);

console.log(markdown);
// # Hello World
//
// This is a **simple** example.

console.log(tokenEstimate);
// { tokens: 12, characters: 46, words: 8 }

console.log(contentHash);
// "d-1a3b4c5" — deterministic, use as ETag or cache key

Content Extraction

Real-world HTML pages are full of navigation, ads, sidebars, and cookie banners. Enable extraction mode to get just the main content:

const { markdown } = convert(html, { extract: true });

This strips <nav>, <header>, <footer>, <aside>, <script>, <style>, ad-related elements, cookie banners, social widgets, and more.

Frontmatter

By default, metadata is extracted from the HTML <head> element and prepended as YAML frontmatter. This aligns with Cloudflare's Markdown for Agents convention.

const html = `<html>
  <head>
    <title>My Page</title>
    <meta name="description" content="A great page about things">
    <meta property="og:image" content="https://example.com/hero.png">
  </head>
  <body><p>Content here</p></body>
</html>`;

const { markdown } = convert(html);
// ---
// title: My Page
// description: A great page about things
// image: https://example.com/hero.png
// ---
// Content here

Extracted fields: title (from <title>), description (from <meta name="description">), image (from <meta property="og:image">).

Disable it or merge custom fields:

// Disable frontmatter
convert(html, { frontmatter: false });

// Merge custom fields (custom overrides extracted)
convert(html, { frontmatter: { author: 'Jane', title: 'Custom Title' } });

Middleware

Framework middleware is available as separate packages — they serve Markdown automatically when AI agents request it via Accept: text/markdown:

// Express
import { markdown } from '@markdown-for-agents/express';
app.use(markdown());

// Fastify
import { markdown } from '@markdown-for-agents/fastify';
fastify.register(markdown());

// Hono
import { markdown } from '@markdown-for-agents/hono';
app.use(markdown());

// Next.js (auto-unwraps /_next/image URLs)
import { withMarkdown } from '@markdown-for-agents/nextjs';
export default withMarkdown(handler);

// Any Web Standard server (Cloudflare Workers, Deno, Bun)
import { markdownMiddleware } from '@markdown-for-agents/web';
const mw = markdownMiddleware();

The middleware inspects the Accept header. Normal browser requests pass through untouched. When an AI agent sends Accept: text/markdown, the HTML response is automatically converted.

Package	Framework
`@markdown-for-agents/express`	Express
`@markdown-for-agents/fastify`	Fastify
`@markdown-for-agents/hono`	Hono
`@markdown-for-agents/nextjs`	Next.js
`@markdown-for-agents/web`	Web Standard (Cloudflare Workers, Deno, Bun)

Custom Rules

Override how any element is converted, or add support for custom elements:

import { convert, createRule } from 'markdown-for-agents';

const { markdown } = convert(html, {
    rules: [
        createRule(
            node => node.name === 'div' && node.attribs.class?.includes('callout'),
            ({ convertChildren, node }) => `\n\n> **Note:** ${convertChildren(node).trim()}\n\n`
        )
    ]
});

Custom rules have higher priority than defaults and are applied first.

Options

All options are optional. Defaults are shown below:

convert(html, {
    // YAML frontmatter from <head> metadata
    frontmatter: true, // false | Record<string, string>

    // Content extraction
    extract: false, // true | ExtractOptions

    // Custom conversion rules
    rules: [], // Rule[]

    // Base URL for resolving relative links and images
    baseUrl: '', // "https://example.com"

    // Heading style
    headingStyle: 'atx', // "atx" (#) or "setext" (underline)

    // Bullet character for unordered lists
    bulletChar: '-', // "-", "*", or "+"

    // Code block style
    codeBlockStyle: 'fenced', // "fenced" or "indented"

    // Fence character
    fenceChar: '`', // "`" or "~"

    // Strong delimiter
    strongDelimiter: '**', // "**" or "__"

    // Emphasis delimiter
    emDelimiter: '*', // "*" or "_"

    // Link style
    linkStyle: 'inlined', // "inlined" or "referenced"

    // Remove duplicate content blocks
    deduplicate: false, // true | DeduplicateOptions

    // Custom token counter (replaces built-in heuristic)
    tokenCounter: undefined, // (text: string) => TokenEstimate

    // Performance timing (populates convertDuration in result)
    serverTiming: false // true to measure conversion duration
});

Server Timing

Enable serverTiming to measure conversion duration. The result includes convertDuration (in milliseconds), and middleware adapters use it to set a Server-Timing header:

const { markdown, convertDuration } = convert(html, { serverTiming: true });
console.log(`Conversion took ${convertDuration}ms`);
// Middleware sets: Server-Timing: mfa.convert;dur=4.7;desc="HTML to Markdown"

Custom Token Counter

By default, token estimation uses a fast heuristic (~4 characters per token). You can replace it with an exact tokenizer:

import { convert } from 'markdown-for-agents';
import { encoding_for_model } from 'tiktoken';

const enc = encoding_for_model('gpt-4o');

const { markdown, tokenEstimate } = convert(html, {
    tokenCounter: text => ({
        tokens: enc.encode(text).length,
        characters: text.length,
        words: text.split(/\s+/).filter(Boolean).length
    })
});

The custom counter receives the final markdown string and must return a TokenEstimate object with tokens, characters, and words fields. It flows through to middleware as well — the x-markdown-tokens header will reflect your counter's value.

Deduplication Options

Pass deduplicate: true to use defaults, or pass a DeduplicateOptions object to customize behavior:

const { markdown } = convert(html, {
    deduplicate: { minLength: 5 } // catch short repeated phrases like "Read more"
});

The minLength option (default: 10) controls the minimum block length eligible for deduplication. Blocks shorter than this are always kept. Lower it to catch short repeated phrases, raise it for more conservative deduplication.

Content-Signal Header

Middleware can set a content-signal HTTP header to communicate publisher consent for AI training, search indexing, and AI input. This is opt-in — the header is only set when explicitly configured:

app.use(
    markdown({
        contentSignal: {
            aiTrain: true, // ai-train=yes
            search: true, // search=yes
            aiInput: true // ai-input=yes
        }
    })
);
// Sets header: content-signal: ai-train=yes, search=yes, ai-input=yes

Only explicitly set fields are included. Set a field to false to signal denial (e.g. aiTrain: false → ai-train=no). Omit a field to exclude it from the header entirely.

Supported Elements

Block

HTML	Markdown
`<h1>`...`<h6>`	`# Heading` (atx) or underline (setext)
`<p>`	Paragraph with blank lines
`<blockquote>`	`> Quoted text`
`<pre><code>`	Fenced code block with language
`<hr>`	`---`
`<br>`	Trailing double-space line break
`<ul>`, `<ol>`, `<li>`	Lists with nesting and indentation
`<table>`	GFM pipe table with separator row
`<script>`, `<style>`, `<noscript>`, `<template>`	Stripped

Inline

HTML	Markdown
`<strong>`, `<b>`	`bold`
`<em>`, `<i>`	`italic`
`<del>`, `<s>`, `<strike>`	`~~strikethrough~~`
`<code>`	`inline code`
`<a>`	`[text](url)` with title and baseUrl support
`<img>`	`![alt](src)` with title and baseUrl support
`<sub>`	`~subscript~`
`<sup>`	`^superscript^`
`<abbr>`, `<mark>`	Pass-through (text preserved)

Subpath Exports

The core package provides fine-grained imports for tree-shaking:

import { convert } from 'markdown-for-agents';
import { extractContent } from 'markdown-for-agents/extract';
import { estimateTokens } from 'markdown-for-agents/tokens';

Runtime Compatibility

Runtime	Version	Status
Node.js	>= 22	Tested
Bun	>= 1.0	Tested
Deno	>= 2.0	Tested
Cloudflare Workers	-	Compatible
Vercel Edge	-	Compatible
Browsers	ES2022+	Compatible

Audit

The @markdown-for-agents/audit package lets you measure token savings when converting HTML to Markdown. Fetch any URL and see exactly how many bytes and tokens you save:

npx agent-markdown-audit https://example.com

           HTML            Markdown        Savings
───────────────────────────────────────────────────
Tokens     48,291          12,073          -75.0%
Chars      193,164         48,292          -75.0%
Words      9,456           5,209
Size       188.6 KB        47.2 KB         -75.0%