Core Usage

When what you want is the data on a page rather than to act on it, don't dump the whole HTML into your context. chrome-use offers a full ladder of extraction techniques from light to heavy: a structured snapshot, precise reads by ref, custom-shaped eval, community site adapters, all the way up to the one-shot extract --schema. This page helps you pick the cheapest one for each kind of page.

Pick the cheapest tool first

Driving a real browser is heavy work. chrome-use is only worth starting up when you genuinely need a logged-in session, interaction, JS rendering, or an anti-bot page. To read a piece of public, static text, don't open a browser at all. Use the table below to decide whether to open a browser at all.

Structured snapshot: snapshot --json

To make a model reason about page content, the best input is a structured snapshot. snapshot -i --json outputs the interactive elements together with their refs as a single line of parseable JSON — compact and stable, ready to feed straight into downstream logic. For how to read the snapshot itself, see Reading Pages.

# structured snapshot — best for letting AI reason over page content
$ chrome-use snapshot -i --json > page.json

Precise extraction: get text / get attr

Once you already have an @ref from a snapshot, reading by ref is the least effort: get text pulls the visible text, get attr pulls a given attribute (such as a link's href). Refs are reassigned on every snapshot, so snapshot first, then read.

# targeted extraction by ref
$ chrome-use snapshot -i
$ chrome-use get text @e5
$ chrome-use get attr @e10 href

Any shape: extract with eval

For data that needs a custom shape, like tables and lists, a snippet of JavaScript is the most direct: querySelectorAll in the page and map it into the array of objects you want. eval runs in the page's MAIN world, so it reads the real DOM directly.

# feed the script to eval --stdin via a heredoc
$ cat <<'EOF' | chrome-use eval --stdin
const rows = document.querySelectorAll("table tbody tr");
Array.from(rows).map(r => ({
  name: r.cells[0].innerText,
  price: r.cells[1].innerText,
}));
EOF

MAIN-world state persists across calls

Because eval runs in the page's MAIN world and state persists across calls, a top-level const x/let x/var x in one call collides with the next (SyntaxError: Identifier 'x' has already been declared). The fixes: use non-colliding variable names, assign to window.x, or wrap the whole snippet in an IIFE (() => { const x = …; return x; })().

Site adapters: the cheapest path to "structured data from site X"

A site adapter is a small piece of community-written JS that hits a site's own JSON API from your logged-in tab and returns clean structured data — no clicking, no scraping, no screenshots, much cheaper than snapshot+click. Prefer one when it exists.

Full site-adapters guide →

One shot: extract --schema

When there's no community adapter and you don't feel like writing eval, use extract --schema: give it a {rows, fields} shape and it extracts tables, search results, feed rows, and dashboards into clean JSON in a single call — no N rounds of find/get, and no dumping the whole page's HTML into your context to blow out the window.

# rows = selector for the repeating container; fields = what to pull per row
# a field is either shorthand CSS (text by default) or { "sel": "…", "get": "text|@attr|html|value" }
$ chrome-use extract --schema '{"rows":".product","fields":{"name":".name","price":".price","href":{"sel":"a","get":"@href"}}}'

# with rows, you get a JSON array — one object per matched container
[
  {"name": "Widget A", "price": "$12.00", "href": "/p/a"},
  {"name": "Widget B", "price": "$8.50", "href": "/p/b"}
]

How to choose: a cheat card

The same page of data often has several routes; roughly from cheap to expensive, this is the order. The full "which extraction tool" decision table (with a "why" column) is maintained on the Site Adapters page.