browse（browse 命令行）

Browse — Browser Automation for Agents

How it works

browse is a CLI that wraps Playwright behind a persistent daemon on a Unix socket. The daemon cold-starts in ~3s on first use, then every command runs in sub-200ms. Session state (cookies, localStorage, auth tokens) persists across commands within a session.

All output is plain text. Objects are JSON-stringified. Commands return non-zero on failure with an error message.

Important constraints:

• Commands are sequential — do not run multiple browse commands in parallel. The daemon handles one command at a time.
• Run browse help for the full command list, or browse help <command> for detailed usage and flags.

The ref system — read this first

Refs (@e1, @e2,...) are how you target elements. They replace CSS selectors for most interactions.

Rules:

1. Always browse snapshot before interacting. Refs only exist after a snapshot.
2. Refs are ephemeral. Every snapshot call regenerates them. Old refs are invalid.
3. Refs go stale after navigation. Any goto or click that changes the page invalidates refs. You'll get a clear error — just browse snapshot again.

Core interaction loop:

browse snapshot              # see what's on the page — get refs
browse fill @e3 "test"       # fill the search field
browse click @e4             # click a button
browse snapshot              # re-snapshot after the page changes

Workflow

The standard pattern for any browser task:

1. Navigate: browse goto <url>
2. Observe: browse snapshot for page structure (interactive elements with refs). Use browse snapshot -i to include structural elements (headings, text), or -f for the full accessibility tree.
3. Check for errors: browse console --level error after navigation.
4. Interact: browse fill @eN "value", browse click @eN, browse hover @eN, browse press Tab, browse select @eN "option", browse scroll @eN (scroll into view).

- Use browse press <key> for keyboard navigation (Tab, Escape, Enter, ArrowDown, Shift+Tab, etc.). Multiple keys: browse press Tab Tab Tab. - Use browse scroll down/up to page through content, browse scroll top/bottom to jump to extremes. - After clicks that trigger SPA navigation, use browse wait url /path, browse wait text "Expected", or browse wait visible.selector before snapshotting.

1. Verify: browse snapshot or browse screenshot after each interaction to confirm the result.
2. Repeat: Move through pages and flows.

For configured applications, browse healthcheck gives a quick pass/fail across key pages.

Key commands by category

Run browse help <command> for flags and detailed usage — don't guess at flags.

Observability quick notes

• Use browse status --metrics when CI or monitoring expects Prometheus-style output.
• Use browse status --json for structured health + daemon metadata in automation.
• Optional daemon env knobs:

- BROWSE_LOG_FORMAT=plain|json — controls daemon log output format (human-readable text vs structured JSON). - BROWSE_LOG_LEVEL=debug|info|warn|error — sets minimum log severity; lower levels (like debug) increase verbosity. - BROWSE_MAX_RSS_MB=<limit> — when RSS exceeds this limit, daemon triggers memory-pressure mitigation (buffer cleanup + deferred ref cleanup + optional GC). - BROWSE_SLOW_COMMAND_MS=<ms> — logs slow-command profiling details for commands that exceed this duration threshold.

Named sessions

Use named sessions to run multiple independent page groups:

browse session create worker-1               # shared context (same cookies/storage)
browse session create worker-2 --isolated    # isolated context (separate cookies/storage)
browse --session worker-1 goto https://a.com
browse --session worker-2 goto https://b.com
browse session list
browse session close worker-1

By default, sessions share the browser context. Use --isolated for fully separate cookies, storage, and permissions.

Authentication

Configured login (preferred — uses browse.config.json):

browse login --env staging

Manual login:

browse goto https://app.example.com/login
browse snapshot
browse fill @e1 "user@example.com"
browse fill @e2 "password123"
browse click @e3
browse snapshot        # verify redirect / dashboard loaded

Session reuse — save after login, load in future sessions:

browse auth-state save /tmp/auth.json
browse auth-state load /tmp/auth.json

Use browse wipe to clear all session data before switching accounts or at the end of a session.

Visual diff

Compare screenshots against a baseline to detect visual regressions:

browse screenshot current.png --diff baseline.png
browse screenshot current.png --diff baseline.png --threshold 5

Output includes similarity percentage, diff pixel count, and a path to the diff image (changed pixels highlighted in red).

Multi-browser

Browse defaults to Chromium. Use --browser to switch:

browse --browser firefox goto https://example.com
browse --browser webkit goto https://example.com
BROWSE_BROWSER=firefox browse goto https://example.com

Stealth features and CDP console capture are Chromium-only; Firefox/WebKit use standard Playwright.

Proxy

Route browser traffic through a proxy:

browse --proxy http://proxy:8080 goto https://example.com
BROWSE_PROXY=socks5://proxy:1080 browse goto https://example.com

Or configure in browse.config.json with "proxy": {"server": "http://proxy:8080", "bypass": "localhost", "username": "u", "password": "p"}.

Playwright passthrough

Pass any Playwright launch or context option via browse.config.json without waiting for explicit browse support:

{
  "playwright": {
    "launchOptions": { "locale": "fr-FR", "timezoneId": "Europe/Paris" },
    "contextOptions": { "colorScheme": "dark", "geolocation": { "latitude": 48.86, "longitude": 2.35 } }
  }
}

launchOptions are applied at browser startup; contextOptions are applied to isolated sessions and video recording contexts. Browse's own options (headless, viewport, stealth) take precedence on conflict.

Headed mode

Launch the browser visibly for debugging (set before the daemon starts):

BROWSE_HEADED=1 browse goto https://example.com

Timeout control

Any command accepts --timeout <ms> (default 30s). Use for slow pages:

browse goto https://slow-page.example.com --timeout 60000

Error recovery