nxm-webtomd

nxm-webtomd turns any web article into clean Markdown — right inside your AI agent.

Point it at a URL and it fetches the page, strips the navigation, ads, and boilerplate, and hands back the main article content as tidy Markdown (with optional YAML front‑matter). It speaks the Model Context Protocol (MCP) over stdio, so it plugs straight into agents like OpenCode, Claude Code, Cursor, and Kiro as a callable tool.

It is small, fast, and dependency‑light — a single self‑contained binary, no runtime downloads, nothing to configure beyond one line in your agent.

Fully open source (MIT / Apache‑2.0). Written in Rust.

Demo

From a noisy web page to clean, structured Markdown:

A real conversion, live in the terminal:

Why it exists

Agents are great at reasoning over text, but terrible at reading the raw web: a modern article page is 90% chrome — menus, cookie banners, share widgets, tracking scripts. Feeding all of that to a model wastes tokens and invites prompt‑injection from hidden page content.

nxm-webtomd does the dirty work first:

• 🧷 Safe by design. Strict fetch policy — request timeout, response size limit, SSRF protection (blocks requests to private/link‑local networks), and content‑type validation.
• 🧹 Clean extraction. A readability‑style engine keeps the article body — headings, paragraphs, lists, links, code blocks — and drops the rest.
• 🛡️ No active content. Raw HTML/JavaScript is never executed or propagated; all active content is stripped to mitigate prompt‑injection. Images are kept as ![](url) references only — no image bytes are downloaded.
• 🗂️ Structured output. Optional YAML front‑matter with title, canonical URL, byline, and publish date.

Everything runs locally. Only the article you ask for is fetched; nothing else leaves your machine.

1. Install

One command. It auto‑detects your system (macOS Apple Silicon or Linux x86_64), downloads the matching binary from the latest GitHub release, verifies its SHA‑256, and installs it to ~/.local/bin:

curl -fsSL https://raw.githubusercontent.com/dangranaz/nxm-webtomd/main/install.sh | sh

If ~/.local/bin is not on your PATH, add it:

export PATH="$HOME/.local/bin:$PATH"

Supported platforms: macOS arm64 (Apple Silicon) and Linux x86_64.

Requires a Rust toolchain (1.75+):

git clone https://github.com/dangranaz/nxm-webtomd.git
cd nxm-webtomd
cargo build --release
# binary at target/release/nxm-webtomd

2. Configure it in your agent

nxm-webtomd is a standard MCP server over stdio, so you configure it like any other MCP tool — the agent manages the process lifecycle.

Here is an example for OpenCode — add it to your opencode.json (global) or opencode.jsonc under the mcp key:

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "nxm-webtomd": {
      "type": "local",
      "command": ["nxm-webtomd"],
      "enabled": true
    }
  }
}

The configuration follows the same pattern in other agents (Claude Code, Cursor, Kiro…): a local MCP server whose command is nxm-webtomd over stdio. For example, in a mcpServers‑style config:

{
  "mcpServers": {
    "nxm-webtomd": {
      "command": "nxm-webtomd",
      "args": [],
      "env": {}
    }
  }
}

Then just ask, in a prompt:

"use nxm-webtomd to fetch https://example.com/article and save it as Markdown to ./article.md"

3. The tool

The server exposes a single, focused MCP tool:

• Omit output_path → the Markdown comes back inline in the tool response.
• Provide output_path → the Markdown is written to that file and the tool reports how many bytes it wrote.

4. Real output

Fetching a dense technical article and converting it produces Markdown like this (front‑matter + body):

---
title: "Hexo Labs Open-Sources SIA: A Self-Improving Agent That Updates Both the Harness and the Model Weights"
canonical_url: "https://www.marktechpost.com/2026/05/29/hexo-labs-open-sources-sia-.../"
byline: "Asif Razzaq"
published_time: "2026-05-29T07:28:37+00:00"
---
Most AI agents stop improving once a human stops tuning them. The model is fixed.
The scaffold around it is fixed. Hexo Labs wants to move both at once...

## What is SIA (Self-Improving AI)

SIA splits a task-specific agent into two parts. The first is the harness...

| Task | Initial | Prev. SOTA | SIA-H | SIA-W+H |
| ---- | ------- | ---------- | ----- | ------- |
| LawBench (top-1 acc) | 13.5% | 45.0% | 50.0% | 70.1% |

Tables, headings, links, and code blocks survive the trip; navigation, ads, and scripts do not.

5. Manual smoke test

The server reads one JSON‑RPC object per line on stdin and replies on stdout:

printf '%s\n%s\n' \
  '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{}}' \
  '{"jsonrpc":"2.0","id":2,"method":"tools/list","params":{}}' \
  | nxm-webtomd

Call the tool directly:

{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{
  "name":"web_to_md",
  "arguments":{"url":"https://example.com/article","output_path":"./article.md"}}}

⭐ Support the project

If nxm-webtomd saves you time or tokens, please give the repository a star and share it — it is the simplest way to help the project reach other developers. Feedback and suggestions are welcome via issues.

License

Licensed under either of

• Apache License, Version 2.0 (LICENSE-APACHE)
• MIT license (LICENSE-MIT)

at your option.