cull▌

jq for HTML.

Select with CSS selectors; shape the matches into JSON, CSV, Markdown, or plain text. One small static binary, built for pipes.

Install ▶ Try it in your browser GitHub →

$ curl -s https://news.ycombinator.com | cull '.athing' -j '{rank: .rank | num, title: .titleline a, url: .titleline a @href}'
{"rank":1,"title":"…","url":"https://…"}
{"rank":2,"title":"…","url":"https://…"}
…

pup and htmlq proved that HTML belongs in shell pipelines — but they stop at printing raw HTML or bare text. cull goes where the data actually needs to go:

-j shaped JSON

A tiny jq-style template turns each match into a clean JSON object (NDJSON, or --array). No other selector CLI does this.

--table

HTML tables straight to CSV or NDJSON — colspan/rowspan expanded, Wikipedia-style multi-row headers merged.

--md

Page (or any selection) to readable Markdown: headings, links, lists, code, tables. Ideal for feeding pages to an LLM.

URLs as input

cull h1 -t https://… fetches for you; relative links resolve against the page URL automatically.

Any encoding

Shift_JIS, KOI8-R, windows-1251 … BOM, header, and <meta charset> honored, browser-style.

grep ergonomics

-c, -l, -1, exit codes, multiple files & globs, clean SIGPIPE. Feels like the tools you know.

Install

brew

curl | sh

scoop

nix

docker

cargo binstall

cargo

brew install rashida-thorne/cull/cull

curl -fsSL https://raw.githubusercontent.com/rashida-thorne/cull/main/scripts/install.sh | sh

scoop bucket add cull https://github.com/rashida-thorne/scoop-cull
scoop install cull

nix run github:rashida-thorne/cull

docker run --rm -i ghcr.io/rashida-thorne/cull '.title' -t < page.html

cargo binstall cull

cargo install cull

Prebuilt binaries (Linux x86_64/arm64 fully static, macOS, Windows) on the releases page. Shell completions: cull --completions zsh · man page: cull --man.

Demo

Show me

# Scrape a Wikipedia table into DuckDB-ready CSV
$ cull --table 'https://en.wikipedia.org/wiki/List_of_ISO_639_language_codes' > codes.csv

# Structured NDJSON from any listing page — missing values become null, rows stay rectangular
$ cull '.post' -j '{title: h2, url: a @href, score: .pts | num, tags: [.tag]}' blog.html
{"title":"Hello, world","url":"/posts/hello","score":128,"tags":["rust","intro"]}

# Strip the boilerplate, feed the article to an LLM
$ cull article --md -r 'nav, footer, script, style' "$URL" | llm "summarize this"

# All links on a page, absolutized
$ cull a -a href https://example.com

# grep across snapshots: which pages leak the generator tag?
$ cull 'meta[name=generator]' -l snapshots/*.html
snapshots/blog.html

More live-verified recipes in the cookbook. Coming from pup or htmlq? Side-by-side migration tables, flag for flag. There are also guides to HTML → Markdown conversion (whole pages, selected regions, LLM pipelines) and HTML table → CSV extraction (colspan/rowspan, stacked headers, sqlite pipelines), and RSS/XML → JSON parsing (feeds, Atom, sitemaps, namespaces).

vs. pup and htmlq

	pup	htmlq	cull
CSS selectors → HTML / text / attr	✓	✓	✓
`:has()`, `:is()`, `:where()`	error	panic	✓
Shaped JSON (jq-style templates)	—	—	✓
Tables → CSV / NDJSON	—	—	✓
Page → Markdown	—	—	✓
Fetch URLs directly	—	—	✓
Non-UTF-8 pages decoded correctly	—	—	✓
Multiple inputs / globs	—	—	✓
XML: RSS/Atom feeds, sitemaps, SVG	mangled	mangled	✓ auto-detected
Interactive selector TUI (live preview)	—	—	✓ `-I`
Remove nodes first (`-r`)	—	✓	✓
Colorized / pretty HTML output	✓	—	✓
Pretty-print that doesn't change rendering	adds spaces	drops spaces	✓ faithful
Whole-doc output keeps `<!DOCTYPE>`	dropped	dropped	✓
Actively maintained	—	—	✓

Also 1.2–1.6× faster on typical pages, with a reproducible benchmark script.