Windows-Use is an AI agent that controls Windows at the GUI layer. It reads the screen via the Windows UI Automation API and uses any LLM to decide what to click, type, scroll, or run — no computer vision model required.
Give it a task in plain English. It handles the rest.
- Open, switch between, and resize application windows
- Click, type, scroll, drag, and use keyboard shortcuts
- Run PowerShell commands and read their output
- Scrape web pages via the browser accessibility tree
- Read and write files on the filesystem
- Manage Windows virtual desktops (create, rename, switch)
- Remember information across steps with persistent memory
- Speak and listen via STT/TTS (voice input and output)
Prerequisites: Python 3.10+, Windows 7/8/10/11
pip install windows-useOr with uv:
uv add windows-usePick any supported LLM provider and run a task:
from windows_use.providers.anthropic import ChatAnthropic
from windows_use.agent import Agent, Browser
llm = ChatAnthropic(model="claude-sonnet-4-5")
agent = Agent(llm=llm, browser=Browser.EDGE)
agent.invoke(task="Open Notepad and write a short poem about Windows")from windows_use.providers.openai import ChatOpenAI
from windows_use.agent import Agent, Browser
llm = ChatOpenAI(model="gpt-4o")
agent = Agent(llm=llm, browser=Browser.CHROME)
agent.invoke(task="Search for the weather in New York on Google")from windows_use.providers.google import ChatGoogle
from windows_use.agent import Agent, Browser
llm = ChatGoogle(model="gemini-2.5-flash")
agent = Agent(llm=llm, browser=Browser.EDGE)
agent.invoke(task=input("Enter a task: "))from windows_use.providers.ollama import ChatOllama
from windows_use.agent import Agent, Browser
llm = ChatOllama(model="qwen3-vl:235b-cloud")
agent = Agent(llm=llm, use_vision=False)
agent.invoke(task=input("Enter a task: "))import asyncio
from windows_use.providers.anthropic import ChatAnthropic
from windows_use.agent import Agent
async def main():
llm = ChatAnthropic(model="claude-sonnet-4-5")
agent = Agent(llm=llm)
result = await agent.ainvoke(task="Take a screenshot and describe the desktop")
print(result.content)
asyncio.run(main())Run the interactive agent directly from your terminal:
windows-useOptions:
--model, -m LLM model to use
--provider, -p LLM provider
--max-steps Max steps per task (default: 200)
--debug, -d Enable debug logging
In-session commands:
| Command | Description |
|---|---|
\llm |
Switch provider or model |
\key |
Change API key |
\speech |
Configure STT/TTS |
\voice |
Record voice input |
\clear |
Clear the screen |
\quit |
Exit |
| Provider | Import |
|---|---|
| Anthropic | from windows_use.providers.anthropic import ChatAnthropic |
| OpenAI | from windows_use.providers.openai import ChatOpenAI |
from windows_use.providers.google import ChatGoogle |
|
| Groq | from windows_use.providers.groq import ChatGroq |
| Ollama | from windows_use.providers.ollama import ChatOllama |
| Mistral | from windows_use.providers.mistral import ChatMistral |
| Cerebras | from windows_use.providers.cerebras import ChatCerebras |
| DeepSeek | from windows_use.providers.deepseek import ChatDeepSeek |
| Azure OpenAI | from windows_use.providers.azure_openai import ChatAzureOpenAI |
| Open Router | from windows_use.providers.open_router import ChatOpenRouter |
| LiteLLM | from windows_use.providers.litellm import ChatLiteLLM |
| NVIDIA | from windows_use.providers.nvidia import ChatNvidia |
| vLLM | from windows_use.providers.vllm import ChatVLLM |
Agent(
llm=llm, # LLM instance (required)
mode="normal", # "normal" (full context) or "flash" (lightweight, faster)
browser=Browser.EDGE, # Browser.EDGE | Browser.CHROME | Browser.FIREFOX
use_vision=False, # Send screenshots to the LLM
use_annotation=False, # Annotate UI elements on screenshots
use_accessibility=True, # Use the Windows accessibility tree
auto_minimize=False, # Minimize active window before the agent starts
max_steps=25, # Max number of steps before giving up
max_consecutive_failures=3, # Abort after N consecutive tool failures
instructions=[], # Extra system instructions
secrets={}, # Key-value secrets passed to the agent context
log_to_console=True, # Print steps to the console
log_to_file=False, # Write steps to a log file
event_subscriber=None, # Custom event listener (see Events section)
experimental=False, # Enable experimental tools (file, memory, multi-select)
)Tip: Use claude-haiku-4-*, claude-sonnet-4-*, or claude-opus-4-* for best results.
The agent has access to these tools automatically — no configuration needed.
Core Tools:
| Tool | Description |
|---|---|
click_tool |
Left, right, middle click or hover at coordinates |
type_tool |
Type text into any input field |
scroll_tool |
Scroll vertically or horizontally |
move_tool |
Move mouse or drag-and-drop |
shortcut_tool |
Press keyboard shortcuts (e.g. ctrl+c, alt+tab) |
app_tool |
Launch, switch, or resize application windows |
shell_tool |
Run PowerShell commands |
scrape_tool |
Extract text content from web pages |
desktop_tool |
Create, rename, switch Windows virtual desktops |
wait_tool |
Pause execution for N seconds |
done_tool |
Return the final answer to the user |
Experimental Tools (enable with experimental=True):
| Tool | Description |
|---|---|
file_tool |
Read, write, list, move, copy, delete files |
memory_tool |
Persist information across steps in markdown files |
multi_select_tool |
Ctrl+click multiple elements at once |
multi_edit_tool |
Fill multiple form fields in one action |
Observe every step the agent takes with the event system:
from windows_use.agent import Agent, AgentEvent, EventType, BaseEventSubscriber
class MySubscriber(BaseEventSubscriber):
def invoke(self, event: AgentEvent):
if event.type == EventType.TOOL_CALL:
print(f"Tool: {event.data['tool_name']}")
elif event.type == EventType.DONE:
print(f"Done: {event.data['answer']}")
agent = Agent(llm=llm, event_subscriber=MySubscriber())Or use a plain callable:
def on_event(event: AgentEvent):
print(f"{event.type.value}: {event.data}")
agent = Agent(llm=llm, event_subscriber=on_event)Event types: THOUGHT · TOOL_CALL · TOOL_RESULT · DONE · ERROR
Windows-Use supports voice input and spoken output via multiple providers.
STT (Speech-to-Text): OpenAI Whisper · Google · Groq · ElevenLabs · Deepgram
TTS (Text-to-Speech): OpenAI · Google · Groq · ElevenLabs · Deepgram
from windows_use.providers.openai import ChatOpenAI, STTOpenAI, TTSOpenAI
from windows_use.speech import STT, TTS
llm = ChatOpenAI(model="gpt-4o")
stt = STT(provider=STTOpenAI())
tts = TTS(provider=TTSOpenAI())
task = stt.invoke() # Record and transcribe voice input
agent = Agent(llm=llm)
result = agent.invoke(task=task)
tts.invoke(result.content) # Speak the response aloudThe agent can manage Windows virtual desktops natively:
from windows_use.vdm.core import create_desktop, switch_desktop, remove_desktop
create_desktop("Work")
switch_desktop("Work")
remove_desktop("Work")Supported on Windows 10 (build 17763+) and all Windows 11 versions.
This agent can:
- Operate your computer on behalf of the user
- Modify files and system settings
- Make irreversible changes to your system
The project provides NO sandbox or isolation layer. For your safety:
- ✅ Use a Virtual Machine (VirtualBox, VMware, Hyper-V)
- ✅ Use Windows Sandbox (Windows 10/11 Pro/Enterprise)
- ✅ Use a dedicated test machine
📖 Read the full Security Policy before deployment.
Windows-Use includes lightweight, privacy-friendly telemetry to help improve reliability and understand real-world usage.
Disable it at any time:
ANONYMIZED_TELEMETRY=falseOr in code:
import os
os.environ["ANONYMIZED_TELEMETRY"] = "false"MIT — see LICENSE.
Contributions are welcome! See CONTRIBUTING for the development workflow.
Made with ❤️ by Jeomon George
@software{
author = {George, Jeomon},
title = {Windows-Use: Enable AI to control Windows OS},
year = {2025},
publisher = {GitHub},
url = {https://github.com/CursorTouch/Windows-Use}
}
