prava 0.1.29

Prava Python SDK for computer-use automation

Prava SDK

Prava is an API for Labor. We're a ML lab training computer-use models to do digital labor and our Controls API is a way to interact with them.

Installation

pip install prava  # or: npm install prava

Quick Start

from prava import ControlClient

client = ControlClient()

response = client.predict({
    "model": "prava-af-medium",
    "instruction": "Click the submit button",
    "image_url": "data:image/png;base64,..."
})

action = response["action"]
# {"kind": "left_click", "coordinate": {"x": 245, "y": 67}}

How It Works

Similar to Claude CUA and OpenAI's ChatGPT Agent, this API is intended to be used in a loop which starts off with a screenshot provided by the user, we then propose an action, the user executes it, we initiate a screenshot, and then predicts the next action. The loop keeps going until we stop.

Models

Model	Use Case
prava-af-medium	General automation
prava-quick-click	Simple, fast automation

Action Types

Action	Example
left_click	{"kind": "left_click", "coordinate": {"x": 245, "y": 67}}
type	{"kind": "type", "text": "[e-mail]"}
key	{"kind": "key", "keys": ["Enter"]}
scroll	{"kind": "scroll", "delta_y": -100}
wait	{"kind": "wait", "duration_ms": 1000}
stop	{"kind": "stop"}

Playwright Example

from playwright.sync_api import sync_playwright
from prava import ControlClient

client = ControlClient()

with sync_playwright() as p:
    browser = p.chromium.launch()
    page = browser.new_page()
    page.goto("https://www.google.com")

    previous_actions = []

    for step in range(5):
        screenshot = page.screenshot()
        image_url = client.screenshot_to_data_uri(screenshot)

        response = client.predict({
            "model": "prava-af-medium",
            "instruction": "Search for 'Prava AI'",
            "image_url": image_url,
            "previous_actions": previous_actions
        })

        action = response["action"]
        if action["kind"] == "stop":
            break

        if action["kind"] == "left_click":
            coord = action["coordinate"]
            page.mouse.click(coord["x"], coord["y"])
        elif action["kind"] == "type":
            page.keyboard.type(action["text"])

        previous_actions.append(action)

    browser.close()

PyAutoGUI Example

import pyautogui
from prava import ControlClient

client = ControlClient()

# Take desktop screenshot
screenshot = pyautogui.screenshot()
import io
img_bytes = io.BytesIO()
screenshot.save(img_bytes, format='PNG')
image_url = client.screenshot_to_data_uri(img_bytes.getvalue())

# Get action
response = client.predict({
    "model": "prava-af-medium",
    "instruction": "Open the calculator app",
    "image_url": image_url
})

# Execute action
action = response["action"]
if action["kind"] == "left_click":
    coord = action["coordinate"]
    pyautogui.click(coord["x"], coord["y"])

TypeScript

import { ControlClient } from 'prava';
import { chromium } from 'playwright';

const client = new ControlClient();

const browser = await chromium.launch();
const page = await browser.newPage();
await page.goto('https://www.google.com');

const screenshot = await page.screenshot();
const imageUrl = client.screenshotToDataUri(new Uint8Array(screenshot));

const response = await client.predict({
  model: 'prava-af-medium',
  instruction: 'Search for "Prava AI"',
  image_url: imageUrl
});

const action = response.action;
if (action?.kind === 'left_click' && action.coordinate) {
  await page.mouse.click(action.coordinate.x, action.coordinate.y);
}

await browser.close();

API Keys

from prava import ControlClient

# Pass API key directly
client = ControlClient(api_key="prava_sk_...")

# Or use environment variable
import os
os.environ["PRAVA_API_KEY"] = "prava_sk_..."
client = ControlClient()

Examples

• examples/playwright_example.py - Browser automation
• examples/pyautogui_example.py - Desktop automation