browser-automation（浏览器自动化）

Browser Automation

Available Tools

• browser_act(instruction, starting_url?): Execute browser actions using natural language (click, type, scroll, select). Use starting_url to navigate to a page and act in a single call.
• browser_get_page_info(url?, text?, tables?, links?): Get page structure and DOM data (fast, no AI). Use url to navigate first; text=True for full text, tables=True for table data, links=True for all links.
• browser_manage_tabs(action, tab_index?, url?): Switch, close, or create browser tabs
• browser_save_screenshot(filename): Save current page screenshot to workspace

When to Use

Use browser automation when the task genuinely requires it:

• UI interactions: Filling forms, clicking buttons, navigating multi-step workflows
• Login-required pages: Accessing content behind authentication that APIs cannot reach
• Dynamic/JS-heavy pages: Content rendered client-side that plain HTTP requests can't capture
• Human-like browsing needed: Sites that block bots or require realistic interaction patterns
• Scraping structured data: When no API exists and the data must be extracted from rendered pages

Prefer web search or url_fetcher for general information lookup, news, or publicly accessible pages — browser automation is slower and heavier. Reserve it for tasks where simpler tools are insufficient.

Tool Selection

• browser_act: UI interactions (click, type, scroll, form fill). Use starting_url to open a page and act in one call.
• browser_get_page_info: Fast page structure check and optional content extraction (<300ms). Use url to navigate first.
• browser_manage_tabs: Switch/close/create tabs (view tabs via get_page_info)
• browser_save_screenshot: Save milestone screenshots (search results, confirmations, key data)

browser_act Best Practice

• Combine up to 3 predictable steps: "1. Type 'laptop' in search 2. Click search button 3. Click first result"
• Use starting_url when opening a fresh page: browser_act(instruction='Search for laptops', starting_url='https://amazon.com')
• On failure: check the screenshot to see current state, then retry from that point
• For visual creation (diagrams, drawings), prefer code/text input methods over mouse interactions

browser_get_page_info Best Practice

• Use url to navigate and inspect in one call: browser_get_page_info(url='https://example.com', tables=True)
• Use text=True to get full page text content (useful for reading article text)
• Use tables=True to extract structured table data from the page
• Use links=True to get all links on the page (up to 200)

UI Guidance (from tools-config)

Tool Selection:

• browser_act: UI interactions (click, type, scroll, form fill). Use starting_url to navigate and act in one call.
• browser_get_page_info: Fast DOM inspection (<300ms). Use url param to navigate first; text/tables/links params for content extraction.
• browser_manage_tabs: Switch, close, or create tabs.
• browser_save_screenshot: Save milestone screenshots to workspace for documents.

browser_act Best Practice:

• Combine up to 3 predictable steps: "1. Type 'laptop' in search 2. Click search button 3. Click first result"
• Use starting_url when opening a fresh page: browser_act(instruction='...', starting_url='https://...')
• On failure: check the screenshot to see current state, then retry from that point

browser_get_page_info Best Practice:

• Use url param to navigate and inspect in one call: browser_get_page_info(url='https://...', tables=True)
• Use text=True for full page text, tables=True for table data, links=True for all page links