mdrip

Fetch clean markdown snapshots of any web page — optimized for AI agents, RAG pipelines, and context-aware workflows.

Reduces token overhead by ~90% compared to raw HTML while preserving the content structure LLMs need.

Why

AI agents and LLMs work better with markdown than HTML. Feeding raw HTML into a context window wastes tokens on tags, scripts, styles, and boilerplate. mdrip solves this by fetching any URL and returning clean, structured markdown.

• ~90% fewer tokens than raw HTML
• Automatic HTML-to-markdown fallback when native markdown isn't available
• Works everywhere — CLI, Node.js, Cloudflare Workers, or via remote MCP
• Token-aware — reports estimated token counts so you can manage context budgets

Sites that support Cloudflare's Markdown for Agents return markdown natively at the edge. For all other sites, mdrip's built-in converter handles headings, links, lists, code blocks, tables, blockquotes, and more, while filtering hidden/non-visible content (including hidden attributes, aria-hidden, inline hidden styles, templates/forms, and HTML comments).

Installation

npm install -g mdrip

Or use directly with npx:

npx mdrip <url>

CLI Usage

Fetch pages

# Fetch one page
mdrip https://example.com/docs/getting-started

# Fetch multiple pages
mdrip https://example.com/docs https://example.com/api

# Custom timeout (ms)
mdrip https://example.com --timeout 45000

# Strict mode — only accept native markdown, no HTML fallback
mdrip https://example.com --no-html-fallback

# Raw mode — print markdown to stdout, no file writes
mdrip https://example.com --raw

List fetched pages

mdrip list
mdrip list --json

Remove pages

mdrip remove https://example.com/docs/getting-started

Clean snapshots

# Remove all
mdrip clean

# Remove only one domain
mdrip clean --domain example.com

Raw mode for agent runtimes

--raw prints markdown to stdout and skips all file writes and prompts. Useful for piping content directly into agent loops.

mdrip https://example.com --raw | your-agent-cli

Programmatic API

npm install mdrip

Method reference

| Import path | Method | Returns | Purpose | |---|---|---|---| | mdrip | fetchMarkdown(url, options?) | Promise<MarkdownResponse> | Fetch one URL to markdown with metadata | | mdrip | fetchRawMarkdown(url, options?) | Promise<string> | Fetch one URL to markdown string only | | mdrip/node | fetchMarkdown(url, options?) | Promise<MarkdownResponse> | Node entrypoint alias for in-memory fetch | | mdrip/node | fetchRawMarkdown(url, options?) | Promise<string> | Node entrypoint alias for markdown-only fetch | | mdrip/node | fetchToStore(url, options?) | Promise<FetchResult> | Fetch one URL and persist to mdrip/pages/... | | mdrip/node | fetchManyToStore(urls, options?) | Promise<FetchResult[]> | Fetch many URLs and persist successful results | | mdrip/node | listStoredPages(cwd?) | Promise<PageEntry[]> | List tracked snapshots from mdrip/sources.json |

FetchMarkdownOptions supports: timeoutMs, userAgent, htmlFallback, fetchImpl, tokenModel, tokenEncoding. Default token encoding is o200k_base (recommended for GPT-4o/4.1/5-family style counting). StoreFetchOptions extends that with cwd.

Workers / Edge / In-memory

import { fetchMarkdown } from "mdrip";

const page = await fetchMarkdown("https://example.com/docs");

console.log(page.markdown);       // clean markdown content
console.log(page.markdownTokens); // estimated token count
console.log(page.source);         // "cloudflare-markdown" or "html-fallback"

Node.js (fetch and store to disk)

import { fetchToStore, listStoredPages } from "mdrip/node";

const result = await fetchToStore("https://example.com/docs", {
  cwd: process.cwd(),
});

if (result.success) {
  console.log(`Saved to ${result.path}`);
}

const pages = await listStoredPages(process.cwd());

Remote MCP + HTTP API

mdrip is available as a remote service at mdrip.createmcp.dev with MCP transports and a direct JSON API.

| Endpoint | Transport | Use case | |---|---|---| | /mcp | Streamable HTTP MCP | Recommended for MCP clients | | /sse | SSE MCP | Legacy MCP client compatibility | | /api | JSON over HTTP | Direct non-MCP integration |

MCP tools

fetch_markdown:

• Inputs: url (required), timeout_ms (optional), html_fallback (optional)
• Output: markdown + metadata (resolvedUrl, status, contentType, source, markdownTokens, contentSignal)

batch_fetch_markdown:

• Inputs: urls (required array, 1-10), timeout_ms (optional), html_fallback (optional)
• Output: one result per URL, with success/error details

HTTP API (/api)

GET /api expects query params:

• url (required)
• timeout (optional ms)
• html_fallback (optional true/false)

curl "https://mdrip.createmcp.dev/api?url=https://example.com&timeout=30000&html_fallback=true"

POST /api supports both single and batch bodies:

{ "url": "https://example.com", "timeout_ms": 30000, "html_fallback": true }

{
  "urls": ["https://example.com", "https://example.com/docs"],
  "timeout_ms": 30000,
  "html_fallback": true
}

Single responses return one fetch result object. Batch responses return { "results": [...] } with success: true|false per URL.

Claude Desktop

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "mdrip": {
      "command": "npx",
      "args": ["mcp-remote", "https://mdrip.createmcp.dev/mcp"]
    }
  }
}

Claude Code

claude mcp add mdrip-remote --transport sse https://mdrip.createmcp.dev/sse

Cloudflare AI Playground

Enter mdrip.createmcp.dev/sse at playground.ai.cloudflare.com.

OpenClaw Integration

Option 1: Dedicated OpenClaw skill (recommended)

Install skills from this repo in your OpenClaw workspace:

cd ~/.openclaw/workspace
npx skills add charl-kruger/mdrip

Enable the OpenClaw-focused skill in ~/.openclaw/openclaw.json:

{
  "skills": {
    "entries": {
      "mdrip-openclaw": {
        "enabled": true
      }
    }
  }
}

If you want OpenClaw to load skills directly from this local repository:

{
  "skills": {
    "load": {
      "extraDirs": ["<absolute-path-to-fetchmd>/skills"]
    }
  }
}

Option 2: Direct CLI usage in OpenClaw workflows

# In-memory markdown only (best for tool pipelines)
mdrip https://example.com/docs --raw

# Persist reusable snapshots in workspace
mdrip https://example.com/docs https://example.com/api
mdrip list --json

File modifications

On first run, mdrip can optionally update:

• .gitignore — adds mdrip/
• tsconfig.json — excludes mdrip/
• AGENTS.md — adds a section pointing agents to your snapshots

Choice is stored in mdrip/settings.json. Use --modify or --modify=false to skip the prompt.

--raw mode bypasses this entirely.

Output structure

mdrip/
├── settings.json
├── sources.json
└── pages/
    └── example.com/
        └── docs/
            └── getting-started/
                └── index.md

Benchmark

Measured on February 13, 2026 (values vary as pages change):

| Page | Mode | Chars saved | Tokens saved | |------|------|------------:|-------------:| | blog.cloudflare.com/markdown-for-agents | cloudflare-markdown | 94.3% | 96.2% | | developers.cloudflare.com/.../markdown-for-agents | cloudflare-markdown | 95.5% | 97.3% | | en.wikipedia.org/wiki/Markdown | html-fallback | 73.8% | 76.4% | | github.com/cloudflare/skills | html-fallback | 96.7% | 98.0% | | Average | | 90.1% | 92.0% |

pnpm build && pnpm benchmark

Token counts use mdrip's tokenizer-based estimator (default encoding: o200k_base).

AI Skills

This repo includes an AI-consumable skills catalog in skills/, following the agentskills format.

• mdrip: general-purpose mdrip skill (CLI, package APIs, remote MCP/API)
• mdrip-openclaw: OpenClaw-focused skill and config/workflow reference

npx skills add charl-kruger/mdrip

Requirements

• Node.js 18+

Author

Charl Kruger

License

Apache-2.0