Claude computer use in 2026 — practical workflows you can actually build today

# Pull the official computer use Docker image
docker pull ghcr.io/anthropics/anthropic-quickstarts:computer-use-demo-latest

# Run with VNC access
docker run -it \
  -e ANTHROPIC_API_KEY=$ANTHROPIC_API_KEY \
  -v $HOME/.anthropic:/home/user/.anthropic \
  -p 5900:5900 \  # VNC port
  -p 8080:8080 \  # Web interface
  ghcr.io/anthropics/anthropic-quickstarts:computer-use-demo-latest

# On Hostinger VPS or any cloud server
docker run -d \
  -e ANTHROPIC_API_KEY=$ANTHROPIC_API_KEY \
  -p 5900:5900 \
  -p 8080:8080 \
  ghcr.io/anthropics/anthropic-quickstarts:computer-use-demo-latest

import anthropic
import base64

client = anthropic.Anthropic()

def take_screenshot() -> str:
    """Capture current desktop screenshot as base64. 
    In the Docker container, use the provided screenshot tool."""
    import subprocess
    result = subprocess.run(
        ["scrot", "-o", "/tmp/screenshot.png"],
        capture_output=True
    )
    with open("/tmp/screenshot.png", "rb") as f:
        return base64.standard_b64encode(f.read()).decode()

def run_computer_use_agent(task: str, max_steps: int = 25) -> str:
    messages = [{
        "role": "user",
        "content": task
    }]
    
    for step in range(max_steps):
        # Add current screenshot
        screenshot = take_screenshot()
        
        response = client.messages.create(
            model="claude-opus-4-7",  # Computer use requires Opus for reliability
            max_tokens=1024,
            tools=[
                {
                    "type": "computer_20250124",
                    "name": "computer",
                    "display_width_px": 1920,
                    "display_height_px": 1080,
                    "display_number": 1,
                }
            ],
            messages=messages + [{
                "role": "user",
                "content": [{
                    "type": "image",
                    "source": {
                        "type": "base64",
                        "media_type": "image/png",
                        "data": screenshot,
                    }
                }]
            }],
        )
        
        if response.stop_reason == "end_turn":
            return response.content[-1].text if response.content else "Task complete"
        
        # Execute tool actions
        for block in response.content:
            if block.type == "tool_use" and block.name == "computer":
                execute_computer_action(block.input)
        
        messages.append({"role": "assistant", "content": response.content})
    
    return "Reached step limit"

def execute_computer_action(action: dict):
    """Execute a mouse/keyboard action on the virtual desktop."""
    import pyautogui
    
    action_type = action.get("action")
    
    if action_type == "screenshot":
        pass  # Screenshot is taken at the start of each loop
    elif action_type == "mouse_move":
        pyautogui.moveTo(action["coordinate"][0], action["coordinate"][1])
    elif action_type == "left_click":
        pyautogui.click(action["coordinate"][0], action["coordinate"][1])
    elif action_type == "type":
        pyautogui.typewrite(action["text"], interval=0.05)
    elif action_type == "key":
        pyautogui.press(action["key"])
    elif action_type == "scroll":
        pyautogui.scroll(action.get("direction", "down") == "up" and 3 or -3)

# Prepare the data you want to submit
form_data = {
    "company_name": "MasterPrompting Technologies Pvt Ltd",
    "cin": "U72900KA2024PTC123456",
    "director_din": "12345678",
    "filing_type": "MGT-7",
    "financial_year": "2025-26",
}

task = f"""
Open Firefox and navigate to mca.gov.in/mcafoportal/login.do

Log in with credentials:
- Username: {MCA_USERNAME}
- Password: {MCA_PASSWORD}

After login, navigate to e-Filing → Annual Return (MGT-7).

Fill in the form with this data:
{form_data}

Take a screenshot before submitting. 
DO NOT click the final Submit button — stop at the review page and describe what you see.
"""

result = run_computer_use_agent(task)
print(result)  # Description of the review page for human to approve

checklist = """
Navigate to https://staging.yourapp.com/login

Check and report on each item:
1. Is the logo visible in the top-left corner?
2. Are there two input fields labeled "Email" and "Password"?
3. Is the "Sign In" button visible and enabled?
4. Is there a "Forgot password?" link?
5. Does the page have a dark mode toggle?
6. Are there any visible error messages or broken images?
7. Does clicking "Sign In" with empty fields show validation messages?

For each check: PASS, FAIL, or PARTIAL. Describe any failures specifically.
"""

result = run_computer_use_agent(checklist)
print(result)
# → "1. PASS - Logo visible top-left
#    2. PASS - Two fields present
#    3. PASS
#    4. PASS
#    5. FAIL - No dark mode toggle visible on this page
#    6. PASS - No errors
#    7. PASS - 'Email is required' validation shown"

task = """
Open Chrome and navigate to http://erp.internal.company.com/login
Log in with the provided credentials.

Navigate to: Procurement → Purchase Orders → All POs

For each purchase order in the list (up to 20):
1. Click on the PO number to open it
2. Extract: PO number, vendor name, total amount, status, date created
3. Go back to the list
4. Move to the next PO

Format the extracted data as a JSON array.
Stop when you've processed 20 POs or reach the end of the list.
"""

result = run_computer_use_agent(task, max_steps=50)
# Parse the JSON from result
import json, re
json_match = re.search(r'\[.*\]', result, re.DOTALL)
if json_match:
    po_data = json.loads(json_match.group())

import subprocess

browsers = ["google-chrome", "firefox"]
results = {}

for browser in browsers:
    subprocess.Popen([browser, "https://yourapp.com"])
    import time; time.sleep(3)
    
    task = f"""
    The {browser} browser should now be open with https://yourapp.com loaded.
    
    Take a screenshot and describe:
    1. Does the header render correctly (logo, navigation, dark mode toggle)?
    2. Are the fonts rendering (no missing glyphs)?
    3. Is the layout centered or broken?
    4. Rate the visual appearance: Good/Degraded/Broken
    
    Be specific about any issues you see.
    """
    
    results[browser] = run_computer_use_agent(task, max_steps=5)

# Compare results
for browser, result in results.items():
    print(f"\n{browser}:")
    print(result)

task = """
Label Studio is open in the browser with the Image Classification project.

For each unlabeled image shown:
1. Look at the image
2. Identify the main subject: cat, dog, bird, other
3. Click the appropriate label button
4. Click Next to move to the following image

Continue until you've labeled 50 images or the project shows "Complete".
Count: keep track of how many of each label you've applied.
"""

result = run_computer_use_agent(task, max_steps=200)
print(result)  # Summary: "Labeled 50 images: 23 cat, 15 dog, 8 bird, 4 other"