Browse

When to use

Use when you want Claude to actually look at a page instead of guessing: opening a URL, testing a site, taking a screenshot, dogfooding a feature, or verifying a deployment. It triggers on phrases like “open in browser”, “test the site”, or “take a screenshot”.

Requires a one-time build (~30 seconds, needs bun) the first time the skill runs. The skill checks for the binary and asks before building.

Capabilities

Accessibility snapshots with refs — snapshot -i prints an indented tree of interactive elements tagged @e1, @e2, … Those refs then work as selectors in any command (click @e3, fill @e4 "value"). -C also catches non-ARIA clickables — divs with cursor:pointer, onclick, or tabindex.

Before/after diffs — snapshot -D stores a baseline and prints a unified diff on the next call, so you see exactly what an interaction changed.

Annotated and responsive screenshots — snapshot -a -o out.png overlays red boxes and ref labels for bug reports. responsive captures mobile, tablet, and desktop in one shot. prettyscreenshot adds clutter removal, scroll positioning, and element hiding; viewport --scale 2 gives retina output.

State assertions — is visible|hidden|enabled|disabled|checked|editable|focused on any selector or ref, plus console, network, and perf for JS errors, failed requests, and load timings.

Forms, dialogs, and uploads — fill, select, type, press, upload, and dialog-accept / dialog-dismiss cover full flows. Cookies, headers, storage, and user agent are all settable.

Headed mode and proxies — --headed opens a visible Chromium (auto-spawning Xvfb on headless Linux) for sites that block headless browsers; --proxy supports HTTP and authenticated SOCKS5.

User handoff — handoff opens a visible Chrome at the current page when Claude hits a CAPTCHA, MFA, or an OAuth flow it can’t complete. All state survives; resume re-snapshots and hands control back.

Offline rendering — goto file://… or load-html renders local HTML with zero network, so it doubles as a rasterizer for diagrams, cards, and og-images without installing a second Chromium.

Also included: page text and HTML extraction, structured data (JSON-LD, Open Graph, Twitter Cards), link and form dumps, media scraping, MHTML archiving, PDF export, CSS cascade inspection with live style edits and undo, tab management, iframe switching, and cross-environment text diffs.

How it works

1. The first command auto-starts a persistent Chromium daemon (~3s).
2. Every command after that reuses the same browser — roughly 100ms each.
3. Cookies, tabs, and login sessions persist between calls, so multi-step flows work without re-authenticating.
4. Output from the page (text, HTML, console, snapshots) is wrapped in untrusted-content markers so page content is never treated as instructions.
5. --headed and --proxy only apply to a fresh daemon; a mismatched config fails loudly rather than silently restarting and dropping your session.

One shared Chromium per machine — no per-project Puppeteer install.

Invoke

/browse:browse

Credits

This plugin is a standalone extraction of the browse skill from gstack by Garry Tan, MIT licensed. All browser engine code is his work; this plugin only packages it for the Claude Code marketplace.