Agent browser

Agent Browser

Automate web browsers via the agent-browser CLI. Every browser task follows one pattern: navigate → snapshot → interact → re-snapshot.

When this applies

• User asks to open, scrape, test, or automate a website
• User asks to fill forms, click buttons, or extract data from web pages
• User asks to take screenshots or capture page content
• User asks to log into a site or automate an authenticated workflow

Core Workflow

Every browser task follows these steps:

1. agent-browser open <url> — navigate to the page
2. agent-browser wait --load networkidle — wait for page to fully load
3. agent-browser snapshot -i — get interactive elements as refs (@e1, @e2, ...)
4. Interact using refs — click @e1, fill @e2 "text", select @e3 "value"
5. Re-snapshot after any action that changes the page — refs are invalidated by navigation, form submissions, modals, and dynamic content

Essential Commands

Navigate

• open <url> — go to URL
• close — close browser (always close when done)

Snapshot

• snapshot -i — list interactive elements with refs (primary discovery method)
• snapshot -i -C — include cursor-interactive elements (divs with onclick)
• snapshot -s "#selector" — scope snapshot to a CSS selector

Interact (use @refs from snapshot)

• click @e1 — click element
• fill @e2 "text" — clear field and type text
• type @e2 "text" — type without clearing
• select @e1 "option" — select dropdown value
• check @e1 / uncheck @e1 — toggle checkbox
• press Enter — press a key
• scroll down 500 — scroll page (up/down/left/right + pixels)

Wait

• wait --load networkidle — wait for all network requests to finish
• wait @e1 — wait for element to appear
• wait --text "Welcome" — wait for text to appear
• wait --url "**/dashboard" — wait for URL pattern after redirect
• wait "#spinner" --state hidden — wait for element to disappear
• wait 2000 — wait fixed milliseconds (last resort)

Extract

• get text @e1 — get element text content
• get url — get current URL
• get title — get page title
• eval 'document.title' — run JavaScript (use --stdin for complex JS)

Capture

• screenshot — save screenshot to temp dir
• screenshot --full — full page screenshot
• screenshot --annotate — numbered labels on interactive elements (use for vision)
• pdf output.pdf — save page as PDF

Semantic Locators (when refs unavailable)

• find text "Sign In" click — find by visible text
• find label "Email" fill "[email protected]" — find by label
• find role button click --name "Submit" — find by ARIA role
• find placeholder "Search" type "query" — find by placeholder

Decision Rules

• If page has dynamic content or SPAs → always wait --load networkidle after open
• If you need element refs → snapshot -i (not screenshot)
• If you need visual layout info → screenshot --annotate (maps [N] labels to @eN refs)
• If you need to verify an action worked → diff snapshot after the action
• If multiple commands don't depend on each other's output → chain with &&
• If you need to read output before next step → run commands separately
• If refs stop working after an action → re-snapshot (refs invalidate on page change)
• If login is needed → use --session-name <name> to persist auth across runs
• If complex JS with nested quotes → use eval --stdin <<'EVALEOF' to avoid shell escaping

Auth Patterns

Session persistence (simplest)

agent-browser --session-name myapp open https://app.com/login
# ... fill credentials, click submit ...
agent-browser close
# Next run: auto-restored
agent-browser --session-name myapp open https://app.com/dashboard

State file (portable)

# After login:
agent-browser state save auth.json
# Future sessions:
agent-browser state load auth.json

Connect to user's browser (one-off)

agent-browser --auto-connect snapshot

Don't

• Use stale refs after page navigation (refs invalidate on DOM change — always re-snapshot)
• Run snapshot without -i flag (without -i you get the full tree, not actionable refs)
• Type passwords in plain commands visible in shell history (use --password-stdin or env vars)
• Forget to close when done (leaks browser processes)
• Use wait 5000 as primary wait strategy (use wait --load networkidle or wait @ref — fixed waits are fragile)
• Chain commands with && when you need intermediate output (snapshot refs must be read before interaction)