agent-rdp（Agent RDP 搜索）

Windows Remote Desktop Control with agent-rdp

Quick start

agent-rdp connect --host <ip> -u <user> -p <pass> --enable-win-automation
agent-rdp automate snapshot -i              # See interactive elements
agent-rdp automate click "@e5"              # Click button by ref
agent-rdp automate fill "@e7" "Hello"       # Type into field
agent-rdp disconnect

Core workflow

1. Connect with automation: agent-rdp connect --host <ip> -u <user> -p <pass> --enable-win-automation
2. Snapshot: agent-rdp automate snapshot -i (get accessibility tree with refs)
3. Act: agent-rdp automate click @e5 or agent-rdp automate fill @e7 "text"
4. Repeat: snapshot → act → snapshot → act...

Troubleshooting

Element not in snapshot with -i

Try without -i flag - some elements aren't marked as interactive but are still actionable:

agent-rdp automate snapshot              # Full tree, no filtering
agent-rdp automate snapshot -d 5         # Limit depth if too large

Element not in accessibility tree at all

Some UI elements (WebView content, certain dialogs, toast notifications) don't appear in the accessibility tree. Use OCR as a last resort:

1. Take screenshot to identify what you need: agent-rdp screenshot -o screen.png
2. Use locate to find coordinates: agent-rdp locate "Button Text"
3. Click using returned coordinates: agent-rdp mouse click <x> <y>

Commands

Connection

agent-rdp connect --host 192.168.1.100 -u Admin -p secret
agent-rdp connect --host 192.168.1.100 -u Admin --password-stdin  # Read password from stdin
agent-rdp connect --host 192.168.1.100 --width 1920 --height 1080
agent-rdp connect --host 192.168.1.100 --drive /tmp/share:Share   # Map local directory
agent-rdp disconnect

Screenshot

agent-rdp screenshot                      # Save to ./screenshot.png
agent-rdp screenshot -o desktop.png       # Save to specific file
agent-rdp screenshot --format jpeg        # JPEG format

Mouse

agent-rdp mouse click 500 300             # Left click at (500, 300)
agent-rdp mouse right-click 500 300       # Right click
agent-rdp mouse double-click 500 300      # Double click
agent-rdp mouse move 100 200              # Move cursor
agent-rdp mouse drag 100 100 500 500      # Drag from (100,100) to (500,500)

Keyboard

agent-rdp keyboard type "Hello World"     # Type text (supports Unicode)
agent-rdp keyboard press "ctrl+c"         # Key combination
agent-rdp keyboard press "alt+tab"        # Switch windows
agent-rdp keyboard press "ctrl+shift+esc" # Task manager
agent-rdp keyboard press "win+r"          # Run dialog
agent-rdp keyboard press enter            # Single key (use press, not key)
agent-rdp keyboard press escape
agent-rdp keyboard press f5

Scroll

agent-rdp scroll up --amount 3            # Scroll up 3 notches
agent-rdp scroll down --amount 5          # Scroll down 5 notches
agent-rdp scroll left
agent-rdp scroll right

Clipboard

agent-rdp clipboard set "Text to paste"   # Set clipboard (paste on Windows)
agent-rdp clipboard get                   # Get clipboard (after copy on Windows)

Drive mapping

# Map at connect time
agent-rdp connect --host <ip> -u <user> -p <pass> --drive /local/path:DriveName

# List mapped drives
agent-rdp drive list

Session management

agent-rdp session list                    # List active sessions
agent-rdp session info                    # Current session info
agent-rdp --session work connect ...      # Named session
agent-rdp --session work screenshot       # Use named session

Wait

agent-rdp wait 2000                       # Wait 2 seconds

Locate (OCR)

agent-rdp locate "Cancel"                 # Find lines containing "Cancel"
agent-rdp locate "Save*" --pattern        # Glob pattern matching
agent-rdp locate --all                    # Get all text on screen
agent-rdp locate "OK" --json              # JSON output with coordinates

Returns text lines with bounding boxes and center coordinates for clicking:

Found 1 line(s) containing 'Cancel':
  'Cancel' at (650, 420) size 45x14 - center: (672, 427)

To click the first match: agent-rdp mouse click 672 427

UI Automation

# Connect with automation enabled
agent-rdp connect --host 192.168.1.100 -u Admin -p secret --enable-win-automation

# Snapshot - get accessibility tree (refs always included)
agent-rdp automate snapshot                # Full desktop tree
agent-rdp automate snapshot -i             # Interactive elements only
agent-rdp automate snapshot -c             # Compact (remove empty elements)
agent-rdp automate snapshot -d 5           # Limit depth to 5 levels
agent-rdp automate snapshot -s "~*Notepad*"# Scope to a window/element
agent-rdp automate snapshot -f             # Start from focused element
agent-rdp automate snapshot -i -c -d 3     # Combine options

# Pattern-based element operations (use selectors: @eN, #automationId, .className, or name)
agent-rdp automate click "#SaveButton"    # Click button
agent-rdp automate click "@e5"            # Click by ref number
agent-rdp automate click "@e5" -d         # Double-click (for file list items)
agent-rdp automate select "@e10"          # Select item (SelectionItemPattern)
agent-rdp automate select "@e5" --item "Option 1"  # Select item by name in container
agent-rdp automate toggle "@e7"           # Toggle checkbox (TogglePattern)
agent-rdp automate toggle "@e7" --state on  # Set specific state
agent-rdp automate expand "@e3"           # Expand menu/tree (ExpandCollapsePattern)
agent-rdp automate collapse "@e3"         # Collapse menu/tree
agent-rdp automate context-menu "@e5"     # Open context menu (Shift+F10)
agent-rdp automate focus <selector>       # Focus element
agent-rdp automate get <selector>         # Get element properties

# Text input
agent-rdp automate fill <selector> "text" # Clear and fill text (ValuePattern)
agent-rdp automate clear <selector>       # Just clear

# Scrolling
agent-rdp automate scroll <selector> --direction down --amount 3

# Window operations
agent-rdp automate window list
agent-rdp automate window focus "~*Notepad*"
agent-rdp automate window maximize
agent-rdp automate window minimize
agent-rdp automate window restore
agent-rdp automate window close "~*Notepad*"

# Run commands/apps (best way to open apps)
agent-rdp automate run "notepad.exe"                                        # Open Notepad
agent-rdp automate run "Start-Process ms-settings:" --wait                  # Open Settings
agent-rdp automate run "calc.exe"                                           # Open Calculator
agent-rdp automate run "Get-Process" --wait --process-timeout 5000          # With 5s timeout

# Wait for element
agent-rdp automate wait-for <selector> --timeout 5000
agent-rdp automate wait-for <selector> --state visible

# Status
agent-rdp automate status

Selector syntax:

• @e5 or @5 - Reference number from snapshot (e prefix recommended)
• #SaveButton - Automation ID
• .Edit - Win32 class name
• ~*pattern* - Name with wildcard
• File - Element name (exact match)

Snapshot output format:

- Window "Notepad" [ref=e1, id=Notepad]
  - MenuBar "Application" [ref=e2]
    - MenuItem "File" [ref=e3]
  - Edit "Text Editor" [ref=e5, value="Hello"]

JSON output

Add --json for machine-readable output:

agent-rdp --json clipboard get
agent-rdp --json session info
agent-rdp --json automate snapshot

Example: Open PowerShell and run command

agent-rdp connect --host 192.168.1.100 -u Admin -p secret
agent-rdp wait 3000                       # Wait for desktop
agent-rdp keyboard press "win+r"          # Open Run dialog
agent-rdp wait 1000
agent-rdp keyboard type "powershell"
agent-rdp keyboard press enter
agent-rdp wait 2000                       # Wait for PowerShell
agent-rdp keyboard type "Get-Process"
agent-rdp keyboard press enter
agent-rdp screenshot --output result.png
agent-rdp disconnect

Example: File transfer via mapped drive

# Connect with local directory mapped
agent-rdp connect --host 192.168.1.100 -u Admin -p secret --drive /tmp/transfer:Transfer

# On Windows, access files at \\tsclient\Transfer
agent-rdp keyboard press "win+r"
agent-rdp wait 500
agent-rdp keyboard type "\\\\tsclient\\Transfer"
agent-rdp keyboard press enter

Example: Automate Notepad with UI Automation

# Connect with automation enabled
agent-rdp connect --host 192.168.1.100 -u Admin -p secret --enable-win-automation

# Open Notepad
agent-rdp automate run "notepad.exe"
agent-rdp wait 2000

# Get accessibility snapshot (refs are always included)
agent-rdp automate snapshot -i            # Interactive elements only

# Type text into the edit control (use ref from snapshot)
agent-rdp automate fill "@e5" "Hello from automation!"

# Use File menu to save - expand menu, then invoke menu item
agent-rdp automate expand "File"          # Expand menu (ExpandCollapsePattern)
agent-rdp wait 500
agent-rdp automate click "Save As..."     # Click menu item

# Wait for Save dialog
agent-rdp automate wait-for "#FileNameControlHost" --timeout 5000

# Fill filename and save
agent-rdp automate fill "#FileNameControlHost" "test.txt"
agent-rdp automate click "#1"             # Click Save button

Environment variables

export AGENT_RDP_HOST=192.168.1.100
export AGENT_RDP_PORT=3389
export AGENT_RDP_USERNAME=Administrator
export AGENT_RDP_PASSWORD=secret
export AGENT_RDP_SESSION=default
agent-rdp connect    # Uses env vars for connection

Debugging with WebSocket streaming

# Enable streaming viewer on port 9224
agent-rdp --stream-port 9224 connect --host 192.168.1.100 -u Admin -p secret

# Open web viewer in browser
agent-rdp view --port 9224

# Or manually access WebSocket at ws://localhost:9224 (broadcasts JPEG frames)

Tips

Prefer automate fill over keyboard type when automation is enabled—it's lossless (no dropped characters) and faster.

Opening applications

Use automate run to launch apps directly:

agent-rdp automate run "notepad.exe"
agent-rdp automate run "calc.exe"
agent-rdp automate run "Start-Process ms-settings:" --wait   # Settings
agent-rdp automate run "explorer.exe C:\\"                   # File Explorer

Limitations

IMPORTANT: Read these limitations carefully before attempting automation tasks.

UI Automation cannot access WebViews

• The Windows Start menu search, Edge browser content, Electron app content, and other WebView-based UIs are NOT accessible via automate snapshot.
• Workaround: Use Win+R (Run dialog) or automate run to launch programs directly instead of navigating through the Start menu.

UI Automation cannot handle UAC dialogs

• User Account Control elevation prompts run on a secure desktop isolated from UI Automation.
• UAC dialogs will NOT appear in automate snapshot output.
• Workaround: Use locate command (OCR) to find button text and mouse click to interact. This is unreliable but may work for simple Yes/No dialogs.

OCR (locate) is not highly reliable

• The locate command uses OCR which can misread characters, miss text entirely, or return imprecise coordinates.
• Use it as a last resort when UI Automation cannot access elements.
• Always verify coordinates before clicking critical buttons.

DO NOT estimate coordinates from screenshots (Claude only)

• Claude models in non-computer-use mode (like Claude Code) are very bad at pixel counting.
• Do NOT look at a screenshot and try to guess coordinates - the estimates will likely be wrong.
• Note: Gemini models are generally good at pixel coordinate estimation.
• If you need vision-based coordinate detection with Claude, the user must implement a harness using Claude's Computer Use Tool.

Recommended workflow when UI Automation fails

1. First, always try automate snapshot (with and without -i flag)
2. If element not found, try locate "text" to find via OCR
3. Use coordinates from locate output with mouse click
4. Never estimate coordinates by looking at screenshots