Furst/furst-read

furst-read

Fetch a page, throw away everything that is not the article, and render what is left as a small self-contained HTML document — no scripts, no stylesheets, no webfonts, no third-party requests.

This is the part of the toolset that does the real work. Choosing a lighter browser engine buys 2–3×. Not loading 3MB of JavaScript buys considerably more.

Use

furst-read <url>              # render and open in a browser (default)
furst-read --html <url>       # minimal HTML to stdout
furst-read --text <url>       # plain text to stdout
furst-read --out page.html <url>
furst-read --no-images <url>
curl -s <url> | furst-read --stdin --text <url>

The default writes to $XDG_CACHE_HOME/furst-read/<hash>.html and opens it in $FURST_BROWSER, $BROWSER, or the first light browser on $PATH. Because the file is named by a hash of the URL, revisiting a page reuses it instead of littering.

Slots straight into furst:

[[rule]]
name = "reader"
hosts = ["wikipedia.org", "news.ycombinator.com", "*.substack.com"]
run = ["furst-read", "{url}"]

How it finds the article

Readability's approach, reimplemented:

1. Score p, pre, td and blockquote by how much prose they hold — character count and comma count, which boilerplate has neither of.
2. Push each score up to the ancestors: parent takes it whole, grandparent half, and it thins out from there.
3. Weight by class and id. post-content gains, sidebar loses.
4. Multiply by 1 − linkDensity, which is what separates a nav block from a paragraph.
5. Take the winner, then pull in sibling nodes that also read like body copy, because articles are routinely split across several divs.

Then serialise against a tag whitelist. Wrapper elements contribute their children and no tag of their own, which is how the div soup disappears.

Details that matter in practice

• Redirect stubs. A <meta http-equiv="refresh"> is followed, including when it is tucked inside <noscript>, which is how most static site generators write them.
• Lazy images. A src holding a placeholder falls back to data-src and friends, so the images are the real ones.
• srcset takes the first candidate, not the last — on old hardware the smallest variant is the one you want.
• Character sets. Content-Type first, then <meta charset>, then UTF-8. Latin-1 pages are still common enough to matter.
• Links are made absolute against the post-redirect URL, so the output works from a file:// path.
• Scripts and styles are never counted as text, so a page cannot inflate its own score with an inline bundle.

Limits

• Listing pages and forums have no single article to find. The extractor will return whatever scored highest, which may be very little.
• Anything that renders client-side arrives empty; there is no JavaScript here and that is the point.
• Paywalls and interstitials are not defeated, only rendered plainly.