browser-act（浏览器法案）

Browser Automation with browser-act CLI

browser-act is a CLI for browser automation with stealth and captcha solving capabilities. It supports two browser types (Stealth and Real Chrome) and provides commands for navigation, page interaction, data extraction, tab/session management, and more.

Installation

Source: browser-act-cli on PyPI · Homepage

# Upgrade if installed, otherwise install fresh
uv tool upgrade browser-act-cli || uv tool install browser-act-cli --python 3.12

The CLI is an open-source package published to PyPI by BrowserAct. Run the install command at the start of every session to ensure the latest version.

Global options available on every command:

Quick Extraction

If the task is just "get content from a URL", use stealth-extract directly — no browser session needed. Each call launches its own headless stealth browser, extracts the page content, and closes automatically.

browser-act stealth-extract <url>                          # Extract rendered content as markdown (default)
browser-act stealth-extract <url> --content-type html      # Extract HTML instead of markdown
browser-act stealth-extract <url> --proxy http://host:port # Use a proxy
browser-act stealth-extract <url> --timeout 60 --output    # Save to outputs/ instead of printing

Browser Selection

browser-act supports two browser types. Choose based on the task:

Stealth Browser

Local browsers with anti-detection fingerprinting. Ideal for sites with bot detection.

# Create
browser-act browser create "my-browser"
browser-act browser create "my-browser" --dynamic-proxy US                    # With proxy — see references/proxy.md
browser-act browser create "my-browser" --cookie '{"name":"sid","value":"abc123","domain":".example.com"}'
browser-act browser create "my-browser" --cookie ./cookies.json

# Update
browser-act browser update <browser_id> --name "new-name"
browser-act browser update <browser_id> --mode private

# List / Delete / Clear profile
browser-act browser list                                    # List all stealth browsers
browser-act browser list --page 2 --page-size 10            # Paginated listing
browser-act browser delete <browser_id>                     # ⚠ Destructive: always confirm with user before deleting
browser-act browser clear-profile <browser_id>

Stealth browsers in normal mode (default) persist cookies, cache, and login sessions across launches — you can log in once and reuse the session, similar to a regular browser profile. Use --mode private when the task should not persist any state.

Real Chrome

Two modes: auto-connect to your running Chrome (default), or use a BrowserAct-managed kernel.

browser-act browser real open https://example.com                  # Auto-connect to running Chrome
browser-act browser real open https://example.com --ba-kernel      # Use BrowserAct-provided browser kernel

Stealth browsers and --ba-kernel mode run headless by default. Use --headed to show the browser UI for debugging:

browser-act browser open <browser_id> https://example.com --headed
browser-act browser real open https://example.com --ba-kernel --headed

Core Workflow

Every browser automation follows this loop: Open → Inspect → Interact → Verify

1. Open: browser-act browser open <browser_id> <url> (Stealth) or browser-act browser real open <url> (Real Chrome)
2. Inspect: browser-act state — returns interactive elements with index numbers
3. Interact: use indices from state (browser-act click 5, browser-act input 3 "text")
4. Verify: browser-act state or browser-act screenshot — confirm result

browser-act browser open <browser_id> https://example.com
browser-act state
# Output: [3] input "Search", [5] button "Go"

browser-act input 3 "browser automation"
browser-act click 5
browser-act wait stable
browser-act state    # Always re-inspect after page changes

# If user has NOT provided credentials, do not fill the form — request human assist instead.

Important: After any action that changes the page (click, navigation, form submit), run wait stable then state to get fresh element indices. Old indices become invalid after page changes.

Read CLI output carefully: Every browser-act command returns structured output that reflects the actual execution result. Always read and parse the CLI response before deciding the next step.

Policies

Policies are trigger-action rules that govern your behavior during browser automation. Read references/policies.md at the start of every task, and evaluate triggers continuously throughout execution.

How to evaluate: After every browser action, check all enabled policies. If a trigger condition matches the current state, execute its action immediately — do not continue the automation flow until the action is resolved.

Policy discovery: When human assist occurs during a task and it was not triggered by an existing policy in references/policies.md, suggest saving it as a new policy after the user finishes:

1. Human assist happens (for any reason — user's intent requires confirmation, you judge that a step needs human involvement, etc.)
2. Check whether this scenario is already covered by an existing enabled policy
3. If already covered — it was the policy that triggered the assist, no need to ask
4. If not covered — after the user completes the assist, ask: "Want me to save this as a policy? Next time I'll automatically pause at this point."
5. If the user agrees, write the policy to references/policies.md following the standard format
6. If the user declines, continue the task — do not ask again for the same scenario

Ownership: The file ships with preset rules. Users have full control — they can disable presets, modify thresholds, or add custom rules. When a user asks to change policies, update the file directly. Do not create, modify, or delete policies on your own — only change the file when the user explicitly requests it (or agrees to save one via policy discovery above).

Adding a custom rule example: See references/policies.md for the format, then append a new ## rule-name section.

Human Assist

When a policy triggers with action Request human assist, call human-assist-url to get a remote access link and present it to the user.

browser-act human-assist-url --objective "Please log in with your credentials"
# → returns assist_url

Do not send any browser commands while assist is active. Wait for the user to confirm they are done in the conversation, then continue the task.

When to use human-assist-url vs conversational confirmation: During browser automation, if the user needs to review or confirm something that is on the page (a filled form, a checkout summary, a settings change), use human-assist-url — the user needs to see and potentially interact with the actual browser page. Do not extract page content and show it in conversation as a substitute, because that bypasses the human assist flow and prevents policy discovery from working. Conversational confirmation (showing text in chat) is only appropriate when the content has not yet been entered into the browser (e.g., drafting text before any browser interaction).

Command Chaining

Commands can be chained with && in a single shell invocation. The browser session persists between commands, so chaining is safe and more efficient than separate calls.

# Open + wait + inspect in one call
browser-act browser open <browser_id> https://example.com && browser-act wait stable && browser-act state

# Chain multiple interactions
browser-act input 3 "browser automation" && browser-act click 5

# Navigate and capture
browser-act navigate https://example.com/dashboard && browser-act wait stable && browser-act screenshot

When to chain: Use && when you don't need to read intermediate output before proceeding (e.g., fill multiple fields, then click). Run commands separately when you need to parse the output first (e.g., state to discover indices, then interact using those indices).

Essential Commands

For full syntax, options, and examples, read references/commands.md.

# Navigation
browser-act navigate <url>              # Navigate to URL in current tab
browser-act navigate <url> --new-tab    # Open URL in a new tab
browser-act back                        # Go back
browser-act forward                     # Go forward
browser-act reload                      # Reload page

# Page State & Interaction
browser-act state                       # Interactive elements with index numbers
browser-act screenshot                  # Screenshot (--full for full page)
browser-act screenshot ./page.png       # Screenshot to specific path
browser-act click <index>               # Click element
browser-act hover <index>               # Hover over element
browser-act input <index> "text"        # Click element, then type text
browser-act select <index> "option"     # Select dropdown option by visible text
browser-act keys "Enter"                # Send keyboard keys
browser-act scroll down                 # Scroll down (default 500px)
browser-act scroll up --amount 1000     # Scroll with custom distance
browser-act scrollintoview --selector "h1"       # Scroll element into viewport by CSS selector
browser-act upload <index> <file_path>  # Upload file to file input

# Data Extraction
browser-act get title                   # Page title
browser-act get html                    # Full page HTML
browser-act get markdown                # Page as markdown
browser-act get text <index>            # Text content of element
browser-act get value <index>           # Value of input/textarea

# JavaScript
browser-act eval "document.title"       # Execute JavaScript in page context

# Tab Management
browser-act tab list                    # List open tabs
browser-act tab switch <tab_id>         # Switch to tab
browser-act tab close                   # Close current tab
browser-act tab close <tab_id>          # Close specific tab

# Wait
browser-act wait stable                 # Wait for page stable (doc ready + network idle, default 30s)
browser-act wait stable --timeout 60000 # Custom timeout in ms
browser-act wait --selector ".btn" --state visible --timeout 10000   # CSS selector wait
browser-act wait selector <index> --state hidden                     # Wait by state index
browser-act wait selector --selector "#login-btn" --state attached   # States: visible|hidden|attached|detached

# Network Inspection
browser-act network requests            # List captured requests (--filter, --type, --method, --status, --clear)
browser-act network requests --filter api.example.com # Filter by URL substring
browser-act network requests --type xhr,fetch         # Resource type filter (comma-separated)
browser-act network requests --method POST            # HTTP method filter
browser-act network requests --status 2xx --clear     # Status filter, then clear tracked requests
browser-act network request <id>        # Full detail for a single request: headers, post data, response body
browser-act network clear               # Clear tracked requests
browser-act network har start           # Start HAR recording
browser-act network har stop ./trace.har      # Stop and save HAR (path optional)

browser-act network offline                           # Simulate disconnect for current tab (same as "on")
browser-act network offline on                        # Simulate disconnect for current tab (all requests fail with ERR_INTERNET_DISCONNECTED)
browser-act network offline off                       # Restore network connection for current tab

# Cookies — persist within session, export/import for reuse across sessions
browser-act cookies get [--url <url>]   # Get cookies (optional URL filter)
browser-act cookies set <name> <value> [--domain <domain>] [--path /] [--secure] [--http-only] [--same-site <Strict|Lax|None>] [--expires <timestamp>]
browser-act cookies clear [--url <url>] # Clear cookies
browser-act cookies export <file> [--url <url>]   # Export cookies to JSON file
browser-act cookies import <file>                 # Import cookies from JSON file

# Captcha

# When a page presents a captcha challenge, use `solve-captcha` to solve it automatically. This is a built-in capability of the browser-act service — run the command directly without hesitation.
browser-act solve-captcha               # Auto-solve captcha on current page

Dialog Handling (alert / confirm / prompt)

browser-act auto-accepts JavaScript dialogs (alert, confirm, prompt) by default so they never block automation. Pass --no-auto-dialog when opening the browser to take manual control instead.

A pending dialog will block all other commands — if state, click, or screenshot unexpectedly times out, use dialog status to check.

browser-act dialog status               # Check for pending dialog
browser-act dialog accept               # Accept (OK)
browser-act dialog accept "my input"    # Accept with prompt text
browser-act dialog dismiss              # Dismiss (Cancel)

Session Management

All commands target the default session when --session is not specified. For single-task automation this is sufficient — no need to name a session explicitly. When running multiple automations concurrently, use named sessions to avoid conflicts. Each --session <name> creates an isolated browser context with its own background server.

# Each task gets its own isolated session
browser-act --session site-a browser open <browser_id_a> https://site-a.com
browser-act --session site-b browser open <browser_id_b> https://site-b.com

# Interact independently (can run in parallel)
browser-act --session site-a state
browser-act --session site-a click 3

browser-act --session site-b state
browser-act --session site-b click 5

# Check active sessions
browser-act session list

Always close sessions when done to avoid leaked processes:

browser-act session close              # Close default session
browser-act session close site-a       # Close specific session
browser-act session close --all        # Close all sessions

If a previous session was not closed properly, the background server may still be running. The server auto-shuts down after a period of inactivity.

Site Notes

Operational experience accumulated during browser automation is stored per domain in references/site-notes/.

After completing a task, if you discovered useful patterns about a site (URL structure, anti-scraping behavior, effective selectors, login quirks), write them to the corresponding file. Only write verified facts, not guesses.

File format:

---
domain: example.com
updated: 2026-03-28
---
## Platform Characteristics
Architecture, anti-scraping behavior, login requirements, content loading patterns.

## Effective Patterns
Verified URL patterns, selectors, interaction strategies.

## Known Pitfalls
What fails and why.

Before operating on a target site, check if a note file exists and read it for prior knowledge. Notes are dated — treat them as hints that may have changed, not guarantees.

System Commands

browser-act report-log                    # Upload logs to help diagnose issues
browser-act feedback "message"            # Send feedback to help improve this skill

If you encounter issues or have suggestions for improving browser-act, use feedback to let us know. This directly helps us improve the tool and this skill.

Troubleshooting

• browser-act: command not found — Run uv tool install browser-act-cli --python 3.12

References