pywebtask 0.18.0

LLM-powered web automation library with autonomous agents

webtask

Easy to use LLM-powered web automation.

Installation

pip install pywebtask
playwright install chromium

Quick Start

from webtask import Webtask
from webtask.integrations.llm import Gemini
import os

wt = Webtask()

llm = Gemini(model="gemini-2.5-flash", api_key=os.getenv("GEMINI_API_KEY"))
agent = await wt.create_agent(llm=llm)

await agent.goto("https://practicesoftwaretesting.com/")

await agent.do("Add 2 Flat-Head Wood Screws to the cart")

verdict = await agent.verify("the cart contains 2 items")
if verdict:
    print("Success!")

Features

Simple or complex tasks - From single actions to multi-step workflows:

# Simple action
await agent.do("Click the login button")

# Complex multi-step task
await agent.do("Go to the product page, find the blue shirt, add it to cart, and proceed to checkout")

Stateful agents - Remember context across multiple tasks:

await agent.do("Go to https://practicesoftwaretesting.com/ and add 2 Flat-Head Wood Screws to the cart")
await agent.do("Add 5 Cross-head screws to the cart")
await agent.do("Go to the cart page and verify the items")

Verification - Simple boolean checks with natural language:

verdict = await agent.verify("the cart contains 7 items")
if verdict:
    print("Success!")

Structured output - Extract data with Pydantic schemas:

from pydantic import BaseModel

class ProductInfo(BaseModel):
    name: str
    price: float
    in_stock: bool

result = await agent.do("Extract product information", output_schema=ProductInfo)

print(f"{result.output.name}: ${result.output.price}")

Easy integration - Multiple ways to create agents:

# Create with new browser
agent = await wt.create_agent(llm=llm)

# Use existing browser
agent = await wt.create_agent_with_browser(llm=llm, browser=browser)

# Use existing context
agent = wt.create_agent_with_context(llm=llm, context=context)

# Use existing page
agent = wt.create_agent_with_page(llm=llm, page=page)

TODO

• Unlimited context - Compact conversation history with LLM for extended sessions
• Computer Use model integration - Pixel-based interaction with Claude Computer Use
• Mind2Web benchmark - Evaluation on Mind2Web dataset

Links

• Documentation
• Benchmarks
• MCP Server

License

MIT