Stagehand Browser Cli

Browser Automation

Automate browser interactions using Stagehand CLI with Claude.

First: Environment Selection (Local vs Remote)

The skill automatically selects between local and remote browser environments:

• If Browserbase API keys exist (BROWSERBASE_API_KEY and BROWSERBASE_PROJECT_ID in .env file): Uses remote Browserbase environment
• If no Browserbase API keys: Falls back to local Chrome browser
• No user prompting: The selection happens automatically based on available configuration

Setup (First Time Only)

Check setup.json in this directory. If setupComplete: false:

npm install    # Install dependencies
npm link       # Create global 'browser' command

Commands

All commands work identically in both modes:

browser navigate <url>                    # Go to URL
browser act "<action>"                    # Natural language action
browser extract "<instruction>" ['{}']    # Extract data (optional schema)
browser observe "<query>"                 # Discover elements
browser screenshot                        # Take screenshot
browser close                             # Close browser

Quick Example

browser navigate https://example.com
browser act "click the Sign In button"
browser extract "get the page title"
browser close

Mode Comparison

Best Practices

1. Always navigate first before interacting 2. View screenshots after each command to verify 3. Be specific in action descriptions 4. Close browser when done

Troubleshooting

• Chrome not found: Install Chrome or use Browserbase mode
• Action fails: Use browser observe to discover available elements
• Browserbase fails: Verify API key and project ID are set

For detailed examples, see EXAMPLES.md. For API reference, see REFERENCE.md.

Browser Automation Examples

This document provides detailed examples of common browser automation tasks using the CLI tool.

Example 1: Extract Product Information from E-commerce

User request: "Go to example.com/product/123 and extract the product details"

Workflow:

1. Navigate to the product page:

   browser navigate https://example.com/product/123

2. Extract product data with schema:

   browser extract "Extract the product information" '{"productName": "string", "price": "number", "currency": "string", "inStock": "boolean", "rating": "number", "reviewCount": "number"}'

3. Close the browser:

   browser close

Expected result: JSON object with product details that can be analyzed or stored.

---

Example 2: Fill Out and Submit a Contact Form

User request: "Fill out the contact form on example.com with my information"

Workflow:

1. Navigate to contact page:

   browser navigate https://example.com/contact

2. Act: Fill in name field:

   browser act "Fill in the name field with 'John Doe'"

3. Act: Fill in email field:

   browser act "Fill in the email field with 'john.doe@example.com'"

4. Act: Fill in message field:

   browser act "Fill in the message field with 'I would like to inquire about your services'"

5. Act: Submit the form:

   browser act "Click the Submit button"

6. Screenshot to capture confirmation:

   browser screenshot

7. Close the browser:

   browser close

---

Example 3: Research and Summarize News Articles

User request: "Check the latest tech news on techcrunch.com and summarize the top stories"

Workflow:

1. Navigate to news site:

   browser navigate https://techcrunch.com

2. Extract article headlines and summaries:

   browser extract "Extract the top 5 article headlines and their summaries" '{"headlines": "string", "summary": "string", "author": "string", "publishedDate": "string"}'

3. Close the browser:

   browser close

4. Analyze and summarize the extracted data using Claude's text analysis capabilities.

---

Example 4: Login and Navigate Authenticated Area

User request: "Log into example.com and navigate to my dashboard"

Workflow:

1. Navigate to login page:

   browser navigate https://example.com/login

2. Act: Fill in username:

   browser act "Fill in the username field with 'myusername'"

3. Act: Fill in password:

   browser act "Fill in the password field with 'mypassword'"

4. Act: Click login button:

   browser act "Click the Login button"

5. Act: Wait for page load:

   browser act "Wait for the page to fully load"

6. Navigate to dashboard:

   browser navigate https://example.com/dashboard

7. Screenshot the dashboard:

   browser screenshot

8. Close the browser:

   browser close

Note: This example uses Chrome's user profile (.chrome-profile/) which may preserve session cookies between runs.

---

Example 5: Search and Collect Results

User request: "Search Google for 'best TypeScript practices' and get the top 5 results"

Workflow:

1. Navigate to Google:

   browser navigate https://www.google.com

2. Act: Perform search:

   browser act "Type 'best TypeScript practices' in the search box and press Enter"

3. Act: Wait for results:

   browser act "Wait for search results to load"

4. Extract search results:

   browser extract "Extract the top 5 search results" '{"title": "string", "url": "string", "snippet": "string"}'

5. Close the browser:

   browser close

---

Example 6: Download a File

User request: "Download the PDF file from example.com/documents/report.pdf"

Workflow:

1. Navigate to the file URL:

   browser navigate https://example.com/documents/report.pdf

2. Act: Wait for download to start:

   browser act "Wait for 5 seconds for the download to complete"

3. Close the browser:

   browser close

Note: Files are automatically downloaded to ./agent/downloads/ directory due to CDP configuration.

---

Example 7: Debugging a Page Issue

User request: "Check why the submit button isn't working on example.com/form"

Workflow:

1. Navigate to the form page:

   browser navigate https://example.com/form

2. Screenshot initial state:

   browser screenshot

3. Observe available elements:

   browser observe "Find all buttons and their states"

4. Observe form fields:

   browser observe "Find all form input fields and their required status"

5. Act: Try filling required fields:

   browser act "Fill in all required fields with test data"

6. Screenshot after filling:

   browser screenshot

7. Observe button state again:

   browser observe "Check if the submit button is now enabled"

8. Close the browser:

   browser close

Analyze the screenshots and observations to determine the issue.

---

Example 8: Multi-Page Data Collection

User request: "Extract product information from the first 3 pages of results on example.com/products"

Workflow:

1. Navigate to products page:

   browser navigate https://example.com/products

2. Extract products from page 1:

   browser extract "Extract all products on this page" '{"name": "string", "price": "number", "imageUrl": "string"}'

3. Act: Click next page:

   browser act "Click the Next Page button"

4. Extract products from page 2:

   browser extract "Extract all products on this page" '{"name": "string", "price": "number", "imageUrl": "string"}'

5. Act: Click next page:

   browser act "Click the Next Page button"

6. Extract products from page 3:

   browser extract "Extract all products on this page" '{"name": "string", "price": "number", "imageUrl": "string"}'

7. Close the browser:

   browser close

Combine and process all extracted data.

---

Tips for Success

• Be specific with natural language: "Click the blue Submit button in the footer" is better than "click submit". This is extremely important because there's much ambiguity in many websites.
• Wait when needed: After navigation or actions that trigger page changes, explicitly wait
• Use observe for discovery: When unsure what elements exist, use observe first
• Take screenshots for debugging: Visual confirmation helps understand what the browser sees
• Handle errors gracefully: If an action fails, try breaking it into smaller steps
• Clean up resources: Always close the browser when done to free up system resources

Browser Automation CLI Reference

This document provides detailed technical reference for the CLI browser automation tool.

Architecture Overview

The browser automation system consists of:

• Stagehand: TypeScript library wrapping Playwright for AI-driven browser control. Uses AI model to find and interact with the right elements, so be specific
• Chrome CDP: Chrome DevTools Protocol connection on port 9222
• CLI Tool: Command-line interface in src/cli.ts for browser automation
• Local Chrome: Chrome browser launched with remote debugging enabled

File Locations

• Chrome Profile: .chrome-profile/ - Persistent browser profile directory
• Screenshots: ./agent/browser_screenshots/ - Screenshot output directory
• Downloads: ./agent/downloads/ - File download directory

CLI Command Reference

navigate

Navigate to a URL in the browser.

Usage:

browser navigate <url>

Parameters:

• url (string, required): The URL to navigate to. Must include protocol (http:// or https://)

Returns: JSON output:

{
  "success": true,
  "message": "Successfully navigated to <url>",
  "screenshot": "/path/to/screenshot.png"
}

Implementation Details:

• Uses Playwright's page.goto() under the hood
• Waits for network idle and DOM content loaded
• Automatically takes a screenshot after navigation
• Supports HTTPS upgrade for HTTP URLs

Example:

browser navigate https://example.com

Error Handling:

• Invalid URLs return error with success: false
• Network timeouts return timeout error
• SSL certificate errors may fail navigation

---

act

Perform an action on the page using natural language.

Usage:

browser act "<action>"

Parameters:

• action (string, required): Natural language description of the action to perform

Returns: JSON output:

{
  "success": true,
  "message": "Successfully performed action: <action>",
  "screenshot": "/path/to/screenshot.png"
}

Note: Without specificity it might succeed on the wrong element!

Implementation Details:

• Uses Stagehand's page.act() which leverages Claude Haiku 4.5
• AI model interprets natural language and executes corresponding browser actions
• Supports: clicking, typing, selecting, scrolling, waiting, hovering, and more
• Automatically handles element location and interaction
• Automatically takes a screenshot after the action

Natural Language Examples:

browser act "Click the login button"
browser act "Fill in email field with test@example.com"
browser act "Scroll to the bottom of the page"
browser act "Select 'California' from the state dropdown"
browser act "Hover over the menu icon"
browser act "Wait for 3 seconds"
browser act "Press the Enter key"
browser act "Double-click the file icon"

Best Practices:

• Be specific about which element to interact with
• Include visual descriptors ("button next to the form", "top menu", "form at bottom")
• For ambiguous elements, mention nearby context
• Break complex actions into multiple simple actions

Error Handling:

• Element not found errors indicate selector couldn't be resolved
• Timeout errors occur when action takes too long
• Action not possible errors indicate element state prevents action
• All errors return JSON with success: false

---

extract

Extract structured data from the current page using a schema.

Usage:

browser extract "<instruction>" '{"field": "type"}'

Parameters:

• instruction (string, required): Natural language description of what to extract
• schema (JSON string, required): Schema definition mapping field names to types

Schema Types:

• "string": Text content
• "number": Numeric values (integers or floats)
• "boolean": True/false values

Returns: JSON output:

{
  "success": true,
  "data": {
    "field1": "value",
    "field2": 123
  }
}

Implementation Details:

• Uses Stagehand's page.extract() with Zod schema validation
• AI model (Claude Haiku 4.5) identifies relevant page elements
• Automatically handles pagination and dynamic content
• Validates extracted data against schema

Schema Example:

browser extract "Extract the product information" '{"productName": "string", "price": "number", "inStock": "boolean", "description": "string", "rating": "number"}'

Complex Extraction Example:

browser extract "Extract all items from the shopping cart" '{"itemName": "string", "quantity": "number", "unitPrice": "number", "totalPrice": "number", "imageUrl": "string"}'

Best Practices:

• Use clear, descriptive field names
• Match schema types to expected data types
• Provide specific extraction instructions
• Handle missing data by checking result properties

Error Handling:

• Schema validation errors indicate type mismatch
• Extraction failures occur when data not found on page
• Timeout errors for pages that take too long to analyze
• All errors return JSON with success: false

---

observe

Discover available actions on the page.

Usage:

browser observe "<query>"

Parameters:

• query (string, required): Natural language query to discover elements

Returns: JSON output:

{
  "success": true,
  "data": [
    {
      "selector": "button.submit-btn",
      "text": "Submit Form",
      "type": "button",
      "visible": true,
      "enabled": true
    }
  ]
}

Implementation Details:

• Uses Stagehand's page.observe() to scan page elements
• Returns actionable elements matching the query
• Provides element properties, states, and available actions

Query Examples:

browser observe "Find all buttons"
browser observe "Find clickable links in the navigation"
browser observe "Find form input fields"
browser observe "Find all submit buttons"
browser observe "Find elements with text 'Login'"
browser observe "Find all images"

Use Cases:

• Page exploration and discovery
• Debugging action failures
• Understanding page structure
• Finding dynamic element selectors

Error Handling:

• Empty array returned when no elements match
• Timeout for pages that take too long to scan
• All errors return JSON with success: false

---

screenshot

Take a screenshot of the current page.

Usage:

browser screenshot

Parameters: None

Returns: JSON output:

{
  "success": true,
  "screenshot": "/path/to/screenshot.png"
}

Implementation Details:

• Captures full viewport at current scroll position
• Saves as PNG format with timestamp in filename
• Automatically resizes images larger than 2000x2000 pixels using Sharp
• Uses lossless PNG compression

Screenshot Path Format:

./agent/browser_screenshots/screenshot-YYYY-MM-DDTHH-MM-SS-mmmZ.png

Example:

browser screenshot

Image Processing:

• Original resolution preserved if ≤ 2000x2000
• Larger images resized to fit within 2000x2000 while maintaining aspect ratio
• Uses Sharp library for high-quality image processing

Best Practices:

• Take screenshots before and after important actions
• Use for visual debugging and verification
• Screenshot after navigation to confirm page loaded
• Capture error states for troubleshooting

Error Handling:

• Directory creation errors if screenshots folder can't be created
• CDP errors if Chrome DevTools Protocol connection fails
• File write errors if disk space insufficient
• All errors return JSON with success: false

---

close

Close the browser and cleanup resources.

Usage:

browser close

Parameters: None

Returns: JSON output:

{
  "success": true,
  "message": "Browser closed"
}

Implementation Details:

• Calls stagehand.close() to clean up Playwright resources
• Kills Chrome process if it was started by the CLI tool
• Clears internal state variables
• Does NOT delete .chrome-profile/ directory (preserved for reuse)

Resource Cleanup:

• Closes all browser tabs and windows
• Terminates Chrome process (only if started by this tool)
• Releases CDP connection
• Clears Stagehand instance

Best Practices:

• Always call at the end of browser automation tasks
• Call even if errors occurred during automation
• Don't call mid-workflow unless explicitly needed

Error Handling:

• Continues cleanup even if some steps fail
• Safe to call multiple times
• Gracefully handles already-closed browser
• All errors return JSON with success: false

---

Configuration Details

Stagehand Initialization

The Stagehand instance is configured in src/cli.ts with:

new Stagehand({
  env: "LOCAL",
  verbose: 0,
  enableCaching: true,
  model: "anthropic/claude-haiku-4-5-20251001",
  localBrowserLaunchOptions: {
    cdpUrl: wsUrl,
  },
})

Configuration Options:

• env: "LOCAL": Uses local Chrome instead of remote browser
• verbose: 0: Minimal logging output
• enableCaching: true: Caches page analysis for better performance
• modelName: Claude Haiku 4.5 for AI-driven actions and extraction
• cdpUrl: Chrome DevTools Protocol endpoint

Chrome Launch Arguments

Chrome is launched by src/cli.ts with:

--remote-debugging-port=9222
--user-data-dir=.chrome-profile
--window-position=-9999,-9999
--window-size=1280,720

Arguments:

• --remote-debugging-port: Enables CDP on port 9222
• --user-data-dir: Persistent profile directory for session/cookie persistence
• --window-position: Launches minimized off-screen
• --window-size: Default window size

Download Configuration

Downloads are configured via CDP:

await client.send("Browser.setDownloadBehavior", {
  behavior: "allow",
  downloadPath: "./agent/downloads",
  eventsEnabled: true,
})

Behavior:

• Downloads start automatically (no dialog)
• Files saved to ./agent/downloads/
• Download events can be monitored via CDP

---

Error Messages Reference

Common Errors

"Could not find local Chrome installation"

• Cause: Chrome/Chromium not installed or not in standard locations
• Solution: Install Chrome from https://www.google.com/chrome/

"Chrome failed to start with remote debugging on port 9222"

• Cause: Port 9222 already in use or Chrome can't bind to port
• Solution: Close other Chrome instances or change CDP port

"Browser failed to become ready within timeout"

• Cause: Chrome launched but page context not ready
• Solution: Check Chrome version compatibility, restart system

"Error performing action: element not found"

• Cause: Natural language description didn't match any page element
• Solution: Use more specific description or use observe to find elements

"Error extracting data: schema validation failed"

• Cause: Extracted data type doesn't match schema
• Solution: Verify schema types match actual page data

"Error taking screenshot: directory not writable"

• Cause: Insufficient permissions for screenshots directory
• Solution: Check file permissions on ./agent/browser_screenshots/

---

Performance Considerations

Caching

Stagehand caches page analysis to improve performance on repeated actions. Cache is maintained for:

• Element selectors
• Page structure analysis
• Vision model results

Timeouts

Default timeouts:

• Navigation: 30 seconds
• Action execution: 30 seconds
• Extraction: 60 seconds
• CDP connection: 15 seconds (50 retries × 300ms)

Resource Usage

Browser automation consumes:

• Memory: ~200-500MB for Chrome process
• CPU: Variable based on page complexity
• Disk: ~50-200MB for Chrome profile
• Network: Depends on pages visited

---

Security Considerations

Credential Handling

• Browser uses persistent profile (.chrome-profile/)
• Saved passwords and cookies persist between sessions
• Consider using isolated profiles for sensitive operations

Download Safety

• Downloads automatically saved to ./agent/downloads/
• No file type restrictions enforced
• Verify downloaded file integrity before use

Network Access

• Browser has full network access
• Respects system proxy settings
• Can access localhost and internal networks

---

Debugging Tips

Enable Verbose Logging

Edit src/cli.ts and change verbose level in Stagehand configuration:

// Change verbose: 0 to verbose: 1 or 2
verbose: 2,  // Maximum verbosity

View Chrome Console

Connect to Chrome DevTools manually: 1. Open Chrome 2. Navigate to chrome://inspect 3. Click "inspect" under Remote Target

Check CDP Connection

Test CDP endpoint:

curl http://localhost:9222/json/version

Monitor Browser Process

Check Chrome process:

ps aux | grep chrome

View Screenshots

Screenshots provide visual debugging:

ls -lh ./agent/browser_screenshots/
open ./agent/browser_screenshots/screenshot-*.png

Test CLI Commands

Test individual commands:

browser navigate https://example.com
browser screenshot
browser close

---

Version Information

• Stagehand: Uses @browserbasehq/stagehand package v2.5.2+
• Model: Claude Haiku 4.5 (claude-haiku-4-5-20251001) for browser actions
• CLI Tool: TypeScript CLI in src/cli.ts
• Agent SDK: @anthropic-ai/claude-agent-sdk for conversation framework
• Browser: Local Chrome/Chromium installation

For updates and changelog, see the main project repository.