oagi-core 0.9.0

Official API of OpenAGI Foundation

OAGI Python SDK

Python SDK for the OAGI API - vision-based task automation.

Installation

# Recommended: All features (desktop automation + server)
pip install oagi

# Or install core only (minimal dependencies)
pip install oagi-core

# Or install with specific features
pip install oagi-core[desktop]  # Desktop automation support
pip install oagi-core[server]   # Server support

Requires Python >= 3.10

Installation Options

• oagi (Recommended): Metapackage that includes all features (desktop + server). Equivalent to oagi-core[desktop,server].
• oagi-core: Core SDK with minimal dependencies (httpx, pydantic). Suitable for server deployments or custom automation setups.
• oagi-core[desktop]: Adds pyautogui and pillow for desktop automation features like screenshot capture and GUI control.
• oagi-core[server]: Adds FastAPI and Socket.IO dependencies for running the real-time server for browser extensions.

Note: Features requiring desktop dependencies (like PILImage.from_screenshot(), PyautoguiActionHandler, ScreenshotMaker) will show helpful error messages if you try to use them without installing the desktop extra.

Quick Start

Set your API credentials:

export OAGI_API_KEY="your-api-key"
export OAGI_BASE_URL="https://api.oagi.com"  # or your server URL

Single-Step Analysis

Analyze a screenshot and get recommended actions:

from oagi import single_step

step = single_step(
    task_description="Click the submit button",
    screenshot="screenshot.png"  # or bytes, or Image object
)

print(f"Actions: {step.actions}")
print(f"Complete: {step.is_complete}")

Automated Task Execution

Run tasks automatically with screenshot capture and action execution:

from oagi import ShortTask, ScreenshotMaker, PyautoguiActionHandler

task = ShortTask()
completed = task.auto_mode(
    "Search weather on Google",
    max_steps=10,
    executor=PyautoguiActionHandler(),  # Executes mouse/keyboard actions
    image_provider=ScreenshotMaker(),    # Captures screenshots
)

Configure PyAutoGUI behavior with custom settings:

from oagi import PyautoguiActionHandler, PyautoguiConfig

# Customize action behavior
config = PyautoguiConfig(
    drag_duration=1.0,      # Slower drags for precision (default: 0.5)
    scroll_amount=50,       # Larger scroll steps (default: 30)
    wait_duration=2.0,      # Longer waits (default: 1.0)
    action_pause=0.2,       # More pause between actions (default: 0.1)
    hotkey_interval=0.1,    # Interval between keys in hotkey combinations (default: 0.1)
    capslock_mode="session" # Caps lock mode: 'session' or 'system' (default: 'session')
)

executor = PyautoguiActionHandler(config=config)
task.auto_mode("Complete form", executor=executor, image_provider=ScreenshotMaker())

Image Processing

Process and optimize images before sending to API:

from oagi import PILImage, ImageConfig

# Load and compress an image
image = PILImage.from_file("large_screenshot.png")
config = ImageConfig(
    format="JPEG",
    quality=85,
    width=1260,
    height=700
)
compressed = image.transform(config)

# Use with single_step
step = single_step("Click button", screenshot=compressed)

Async Support

Use async client for non-blocking operations and better concurrency:

import asyncio
from oagi import async_single_step, AsyncShortTask

async def main():
    # Single-step async analysis
    step = await async_single_step(
        "Find the search bar",
        screenshot="screenshot.png"
    )
    print(f"Found {len(step.actions)} actions")
    
    # Async task automation
    task = AsyncShortTask()
    async with task:
        await task.init_task("Complete the form")
        # ... continue with async operations

asyncio.run(main())

Examples

See the examples/ directory for more usage patterns:

• google_weather.py - Basic task execution with ShortTask
• single_step.py - Basic single-step inference
• screenshot_with_config.py - Image compression and optimization
• execute_task_auto.py - Automated task execution
• socketio_server_basic.py - Socket.IO server example
• socketio_client_example.py - Socket.IO client implementation

Socket.IO Server (Optional)

The SDK includes an optional Socket.IO server for real-time bidirectional communication with browser extensions or custom clients.

Installation

# Install with server support
pip install oagi  # Includes server features
# Or
pip install oagi-core[server]  # Core + server only

Running the Server

import uvicorn
from oagi.server import create_app, ServerConfig

# Create FastAPI app with Socket.IO
app = create_app()

# Run server
uvicorn.run(app, host="0.0.0.0", port=8000)

Or use the example script:

export OAGI_API_KEY="your-api-key"
python examples/socketio_server_basic.py

Server Features

• Dynamic namespaces: Each session gets its own namespace (/session/{session_id})
• Simplified events: Single init event from client with instruction
• Action execution: Emit individual actions (click, type, scroll, etc.) to client
• S3 integration: Server sends presigned URLs for direct screenshot uploads
• Session management: In-memory session storage with timeout cleanup
• REST API: Health checks and session management endpoints

Client Integration

Clients connect to a session namespace and handle action events:

import socketio

sio = socketio.AsyncClient()
namespace = "/session/my_session_id"

@sio.on("request_screenshot", namespace=namespace)
async def on_screenshot(data):
    # Upload screenshot to S3 using presigned URL
    return {"success": True}

@sio.on("click", namespace=namespace)
async def on_click(data):
    # Execute click at coordinates
    return {"success": True}

await sio.connect("http://localhost:8000", namespaces=[namespace])
await sio.emit("init", {"instruction": "Click the button"}, namespace=namespace)

See examples/socketio_client_example.py for a complete implementation.

Documentation

License

MIT