Markdown Fetch

Markdown Fetch

When fetching web content for AI processing, request markdown instead of HTML. Cloudflare's "Markdown for Agents" feature (and similar services) convert HTML to clean markdown on the fly when the request includes an Accept: text/markdown header. This dramatically cuts token usage since HTML markup, nav bars, script tags, and wrapper divs are stripped out.

How It Works

Any Cloudflare-enabled site with "Markdown for Agents" turned on will honor content negotiation. The pattern is simple:

1. Send Accept: text/markdown, text/html in the request headers
2. If the server supports it, you get back content-type: text/markdown
3. If not, you get standard HTML as a fallback

The response may also include:

• x-markdown-tokens header: estimated token count of the markdown content.

Use this to decide on chunking strategy or check context window fit.

• Content-Signal header: declares the publisher's AI usage permissions

(training, search, agentic input).

When to Apply This Pattern

Apply this when ALL of these are true:

• You are fetching a web page (not an API endpoint or binary file)
• The content will be processed by an LLM (summarization, analysis, RAG, etc.)
• You don't need the raw HTML structure itself

Skip this when:

• The task requires specific DOM elements or CSS selectors
• You're calling a REST API that returns JSON/XML
• You need to render the page visually

Implementation

curl

curl https://example.com/page \
  -H "Accept: text/markdown, text/html"

Python (requests)

import requests

response = requests.get(
    "https://example.com/page",
    headers={"Accept": "text/markdown, text/html"}
)

is_markdown = "text/markdown" in response.headers.get("content-type", "")
token_estimate = response.headers.get("x-markdown-tokens")
content = response.text

TypeScript / Workers

const response = await fetch("https://example.com/page", {
  headers: { Accept: "text/markdown, text/html" },
});

const isMarkdown = response.headers.get("content-type")?.includes("text/markdown");
const tokenCount = response.headers.get("x-markdown-tokens");
const content = await response.text();

Checking Whether a Site Supports It

Not every site supports this. After fetching, check the response content-type header. If it contains text/markdown, the conversion succeeded. If it returns text/html, the site either doesn't use Cloudflare or hasn't enabled the feature. Your code should handle both cases gracefully.

Presenting Markdown to the User

When a fetch returns text/markdown, save the content as a .md file and present it to the user in addition to using it for the task at hand. The markdown is already in context from the fetch, so this adds negligible cost.

1. Complete the primary task (summarize, analyze, etc.) using the markdown content
2. Write the markdown to /mnt/user-data/outputs/{descriptive-name}.md
3. Call present_files to share it

The user gets a clean, rendered, downloadable copy of the page content alongside whatever they originally asked for. This is a side effect, not a replacement for the primary task. If the response came back as text/html (site doesn't support markdown negotiation), skip this step and proceed normally.

Alternative Conversion Methods

If markdown negotiation isn't available from the source, Cloudflare provides two other conversion options:

• Workers AI AI.toMarkdown(): Converts HTML, PDFs, Office docs, and other

formats to markdown server-side. Useful for document ingestion pipelines.

• Browser Rendering /markdown REST API: Renders dynamic pages (SPAs,

JS-heavy sites) in a real browser first, then converts to markdown.

Example: Deciding Based on Token Count

token_estimate = response.headers.get("x-markdown-tokens")
if token_estimate and int(token_estimate) > context_window_limit:
    # Content too large - chunk it
    chunks = split_markdown_by_sections(content)
else:
    # Fits in context - use directly
    process(content)

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: tonysina
• Source: tonysina/claude-skills
• License: MIT

Install and usage instructions live in the source repository linked above.