| name | browse |
| description | Browser automation CLI for AI agents. Use when the user needs to interact with a web page or web app in a real browser, including navigating pages, filling forms, clicking buttons, taking screenshots, extracting page data, testing rendered UI, or running browser-based health checks. Do not use for code review, API-only checks, generic QA, or general web research unless browser interaction is required.
|
| allowed-tools | Bash(browse:*) |
| compatibility | Requires the browse binary. Do not auto-install; ask before installing and prefer a pinned, verifiable release artifact. |
Browse — Browser Automation for Agents
Setup
Check if browse is installed:
browse version
If not installed, ask the user before installing. Prefer a pinned release artifact with a published checksum or signature. Do not pipe a mutable remote script directly into bash. If the user approves a convenience installer, download it to a temporary file, show the source URL, inspect or verify it where possible, then execute the local file.
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:
- Always
browse snapshot before interacting. Refs only exist after a snapshot.
- Refs are ephemeral. Every
snapshot call regenerates them. Old refs are invalid.
- 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:
- Navigate:
browse goto <url>
- 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.
- Check for errors:
browse console --level error after navigation.
- 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.
- Verify:
browse snapshot or browse screenshot after each interaction to confirm the result.
- Repeat: Move through pages and flows.
For configured applications, browse healthcheck gives a quick pass/fail across key pages.
Key commands by category
| Category | Commands |
|---|
| Navigate | goto <url>, url, back, forward, reload [--hard], text, version, quit, wipe |
| Observe | snapshot, screenshot, console, network |
| Interact | click @eN, hover @eN [--duration ms], press <key> [key ...], fill @eN "value", select @eN "option", upload @eN <file> [file ...], attr @eN [attribute], scroll down/up/top/bottom/@eN/x y |
| Wait | wait url <str>, wait text <str>, wait visible <sel>, wait hidden <sel>, wait network-idle, wait <ms> |
| Viewport | viewport, goto --viewport/--device/--preset |
| Evaluate | eval <expr> (in-page JS), page-eval <expr> (Playwright page API) |
| Auth | login --env <name>, auth-state save/load <path> |
| Tabs | tab list/new/switch/close |
| Assert | assert visible/text-contains/url-contains/... |
| Accessibility | a11y (full page), a11y @eN (element), a11y --standard wcag2aa, a11y --json |
| Flows | flow list, flow <name> --var key=value, healthcheck |
Run browse help <command> for flags and detailed usage — don't guess at flags.
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.
Timeout control
Any command accepts --timeout <ms> (default 30s). Use for slow pages:
browse goto https://slow-page.example.com --timeout 60000
Error recovery
| Error | Fix |
|---|
"element is outside of the viewport" | Run browse scroll @eN to scroll it into view, then retry |
"Refs are stale" / "Unknown ref" | Run browse snapshot to refresh refs |
"Daemon connection lost" | Re-run the command — CLI auto-restarts the daemon |
"Command timed out after Nms" | Use --timeout 60000, or check the URL |
"Daemon crashed and recovery failed" | Run browse quit, then retry |
"Unknown command" for a valid command | Stale daemon — run browse quit, then retry |
"Unknown flag" | Check browse help <cmd> for valid flags |
| Login fails | Check env vars, verify login URL, browse screenshot to see the page |