pi-fetch

Web content extraction for pi. Fetch any URL as clean Markdown — no headless browser required.

Powered by Defuddle by Steph Ango (creator of Obsidian Web Clipper).

How it works

Registers a web_fetch tool that the LLM can call with any URL. Under the hood:

1. Native fetch retrieves the HTML (lightweight, no browser engine)
2. Defuddle extracts the main content, stripping navigation, ads, sidebars, cookie banners, and clutter
3. Returns clean Markdown with metadata (title, author, description, word count)

web_fetch https://www.npmjs.com/package/defuddle

# defuddle

**Author:** kepano · **Domain:** npmjs.com · **Words:** 342

> Get the main content of any page as Markdown.

---

Defuddle extracts the main content from web pages...

Output is automatically truncated to stay within pi's context limits.

Install

# As a pi package
pi install npm:@rjshrjndrn/pi-fetch

# Or test without installing
pi -e npm:@rjshrjndrn/pi-fetch

# Or test locally
pi -e ./extensions/index.ts

Usage

Once installed, just ask pi to fetch a URL:

fetch https://docs.example.com/getting-started

Or the LLM will use web_fetch automatically when it needs to read a webpage.

Limitations

• No JavaScript rendering — uses native fetch, not a browser. SPAs that require JS to render content will return empty or minimal results. For those, you'll still need a headless browser.
• Some sites block non-browser requests — sites with aggressive bot detection may reject the request.
• Output truncation — very large pages are truncated to 50KB / 2000 lines to protect context window.

Development

cd pi-fetch
npm install
npm test

Credits

• Defuddle by Steph Ango (kepano) — content extraction engine
• pi by Mario Zechner (badlogic) — agent framework

License

MIT