modbender Visual RPA Desktop Automation
Auto-execute all steps without waiting for user confirmation between steps.
Desktop automation via screen capture + Qwen vision model (Qwen-VL). No DOM or accessibility API needed.
How it works
- Capture screen -> thumbnail rough positioning
- Full-resolution crop -> precise coordinate refinement
- Execute mouse/keyboard action -> screenshot verification
- Compound instructions automatically decomposed into atomic steps
Usage
Use exec tool to run commands. Script path: $env:TAXBOT_ROOT/skills/visual-rpa/scripts/visual_rpa.py
Requires DASHSCOPE_API_KEY environment variable to be set.
Single task
python "$env:TAXBOT_ROOT/skills/visual-rpa/scripts/visual_rpa.py" --mode task --task "click to open WeChat"
Compound task (auto-decomposed)
python "$env:TAXBOT_ROOT/skills/visual-rpa/scripts/visual_rpa.py" --mode task --task "open WeChat, open File Transfer chat, type hello in input box, click send"
Multi-step task (manually specified)
python "$env:TAXBOT_ROOT/skills/visual-rpa/scripts/visual_rpa.py" --mode task --task "click Chrome browser" "type baidu.com in address bar and press enter" "type weather in search box" "click search button"
Skip verification (faster)
python "$env:TAXBOT_ROOT/skills/visual-rpa/scripts/visual_rpa.py" --mode task --no-verify --task "click to open Calculator"
Parameters
| Parameter | Description |
|---|---|
--mode task |
Batch task mode (required) |
--mode interactive |
Interactive mode (default) |
--task "step1" "step2" |
Task instructions, supports multiple |
--no-verify |
Skip post-action verification |
--model MODEL |
Vision model name (default: qwen-vl-max-latest) |
--api-key KEY |
API Key (defaults to DASHSCOPE_API_KEY env var) |
Supported actions
| Action | Example instructions |
|---|---|
| Click | "click start menu", "click Chrome icon" |
| Double click | "double click Recycle Bin on desktop" |
| Right click | "right click on desktop blank area" |
| Type text | "type weather in search box", "type hello in input box" |
| Hotkey | "press Ctrl+C" |
| Scroll | "scroll down the page" |
| Wait | "wait for page to load" |
Instruction tips
- Be specific: "click WeChat icon on taskbar" is better than "open WeChat"
- Instructions can be in Chinese or English, the model understands both
- Complex operations can be written as compound instructions, system auto-decomposes
- For text input: say "type XXX in YYY", system auto-detects as input action
Output format
[OK] Step 0: click to open WeChat
click @ (375,1591)
[OK] Step 1: click File Transfer Assistant in WeChat
click @ (154,97)
[FAIL] Step 2: type hello in input box
type @ (300,1364)
2/3 succeeded
- OK = action succeeded and verified
- FAIL = action failed or verification failed, auto-retries up to 3 times
Common scenarios
Send WeChat message
python "$env:TAXBOT_ROOT/skills/visual-rpa/scripts/visual_rpa.py" --mode task --task "open WeChat, open File Transfer Assistant chat, type hello in input box, click send"
Open app and navigate
python "$env:TAXBOT_ROOT/skills/visual-rpa/scripts/visual_rpa.py" --mode task --task "click Chrome browser" "type https://www.baidu.com in address bar and press enter"
Desktop operations
python "$env:TAXBOT_ROOT/skills/visual-rpa/scripts/visual_rpa.py" --mode task --task "right click on desktop blank area" "click New Folder"
Notes
- Each step takes 3-8 seconds (screenshot + API calls + verification)
- Chinese text input uses clipboard paste, will overwrite current clipboard
- Only operates on primary screen
- Logs and screenshots saved in
./rpa_logs/directory for debugging