md-json-ai

Lightweight markdown → structured JSON parser. Designed for consuming AI-friendly markdown endpoints (/index.md, llms.txt, GitHub READMEs) and turning them into clean, navigable APIs — without a server.

Runs in the browser, Cloudflare Workers, and anywhere that speaks JavaScript.

https://docs-api.language-models.workers.dev/workers-ai/platform/pricing/index.md

Output shape

Headings become object keys. Content nests under them. Mixed sections (content + sub-headings) become arrays where strings, tables, and callouts are plain values and sub-sections are {"Heading": ...} entries.

{
  "meta": { "title": "Workers AI Pricing", "description": "..." },
  "Pricing": [
    { "note": "Updated pricing is more granular." },
    "Workers AI is priced at $0.011 per 1,000 Neurons.",
    { "Custom requirements": "Complete the form for higher limits." },
    { "LLM model pricing": [
      { "Model": "@cf/meta/llama-3.2-1b-instruct", "Price in Tokens": ["$0.027 per M input tokens", "$0.201 per M output tokens"] }
    ]}
  ]
}

Tables → array of objects (or 2D array if no headers).
Lists → plain string array.
Code blocks → { code, lang } — JSON blocks also get a parsed key.
Callouts → { note: "..." }, { warning: "..." }, etc.
Leaf strings that are valid JSON → parsed in place.
Nav chrome (Skip to content, Edit page, Was this helpful?) → stripped automatically.

Parser

Built on mdast-util-from-markdown + micromark-extension-gfm — the same CommonMark-compliant core that powers the unified/remark ecosystem. ~37KB gzipped with GFM support.

Usage

import { parseMarkdown } from './parser.js'

const json = parseMarkdown(markdownString)

Or fetch a doc’s .md endpoint directly:

const md = await fetch('https://developers.cloudflare.com/workers-ai/platform/pricing/index.md').then(r => r.text())
const json = parseMarkdown(md)

Cloudflare Worker

import { parseMarkdown } from './parser.js'

export default {
  async fetch(req) {
    const url = new URL(req.url)
    const target = url.searchParams.get('url')
    if (!target) return new Response('?url= required', { status: 400 })

    const md = await fetch(target).then(r => r.text())
    const json = parseMarkdown(md)

    return new Response(JSON.stringify(json, null, 2), {
      headers: { 'Content-Type': 'application/json', 'Access-Control-Allow-Origin': '*' }
    })
  }
}

Handling real-world docs

Sites increasingly publish markdown for AI consumption but render it server-side first, baking in nav chrome. The parser handles this with a pre-pass that strips known patterns before handing off to mdast:

• [Skip to content], Was this helpful?, YesNo
• [ Edit](url) [ Issue](url) nav link lines
• ## Was this helpful? headings
• Copy page

New patterns are easy to add to stripCruft().

Caveats

• Multi-value table cells (separated by 2+ spaces, common in CF docs) become arrays
• Blank table header columns are merged into the previous named header
• Bold-only paragraphs (**Title**) followed by a body paragraph are treated as informal callouts: { "Title": "body text" }
• Frontmatter goes into meta and is excluded from section keys

Demo

Open demo — live editor with collapsible JSON tree and copy button.

License

MIT