PyBA

Tell the AI what to do once. Get a Python script you can run forever.

PyBA uses LLMs to autonomously navigate any website, then exports the session as a standalone Playwright script - no API costs on repeat runs.

PyPI • Documentation • OpenHub

The Problem with AI Browser Agents

Every AI browser agent has the same issue: you pay for every single run.

• Run it 100 times? Pay for 100 LLM calls.
• Same task every day? Pay every day.
• The AI figures out the same clicks over and over.

PyBA is different. Let the AI figure it out once, then export a deterministic script you own forever.

from pyba import Engine

engine = Engine(openai_api_key="sk-...")

# Step 1: AI navigates autonomously
engine.sync_run(
    prompt="Go to Hacker News, click the top story, extract all comments"
)

# Step 2: Export as a standalone Playwright script
engine.generate_code(output_path="hacker_news_scraper.py")

Now run python hacker_news_scraper.py forever. No AI. No API costs. Just Playwright.

Installation

pip install py-browser-automation

What Can You Do?

Automate Repetitive Browser Tasks

engine.sync_run(
    prompt="Login to my bank, download this month's statement as PDF",
    automated_login_sites=["swissbank"]
)
engine.generate_code("download_statement.py")

OSINT & Reconnaissance

from pyba import DFS

dfs = DFS(openai_api_key="sk-...")
dfs.sync_run(
    prompt="Find all social media accounts linked to username 'targetuser123'"
)

Structured Data Extraction

from pydantic import BaseModel

class Product(BaseModel):
    name: str
    price: float
    rating: float

engine.sync_run(
    prompt="Scrape all products from the first 3 pages",
    extraction_format=Product
)
# Data is extracted DURING navigation, stored in your database

Authenticated Workflows

engine.sync_run(
    prompt="Go to my Instagram DMs and message john Paula 'Running 10 mins late'",
    automated_login_sites=["instagram"]
)
# Credentials come from env vars - never exposed to the LLM

Four Exploration Modes

from pyba import Engine, Step, DFS, BFS

# Normal mode (default)
engine = Engine(openai_api_key="...")

# Step-by-step interactive mode
step = Step(openai_api_key="...")

# Deep-first exploration
dfs = DFS(openai_api_key="...")

# Breadth-first discovery
bfs = BFS(openai_api_key="...")

Interactive Step-by-Step Automation

from pyba import Step

step = Step(openai_api_key="sk-...")

await step.start()
await step.step("Go to google.com and search for 'playwright python'")
await step.step("Click the first result")
output = await step.step("Extract the installation instructions")
await step.stop()

Key Features

Code Generation

Export any successful run as a standalone Python script. Run it forever without AI.

Trace Files

Every run generates a Playwright trace.zip — replay exactly what happened in Trace Viewer.

Stealth Mode

Anti-fingerprinting, random mouse movements, human-like delays. Bypass common bot detection.

Multi-Provider

Works with OpenAI, Google VertexAI, or Gemini.

Database Logging

Store every action in SQLite, PostgreSQL, or MySQL. Audit trails and replay capability.

Platform Logins

Built-in login handlers for Instagram, Gmail, Facebook. Credentials stay in env vars.

Quick Examples

Extract YouTube Video Metadata

engine.sync_run(
    prompt="Go to this YouTube video and extract: title, view count, like count, channel name, upload date"
)

Fill a Multi-Page Form

engine.sync_run(
    prompt="Fill out the job application: Name='John Doe', Email='john@email.com', upload resume from ~/resume.pdf, submit"
)
engine.generate_code("job_application.py")  # Replay anytime

Research a Company

dfs = DFS(openai_api_key="...")
dfs.sync_run(
    prompt="Find the leadership team, recent news, and funding history for Acme Corp"
)

Configuration

from pyba import Engine, Database

# With database logging
db = Database(engine="sqlite", name="runs.db")

engine = Engine(
    openai_api_key="sk-...",
    headless=False,           # Watch it work
    enable_tracing=True,      # Generate trace.zip
    max_depth=20,             # Max actions per run
    database=db               # Log everything
)

See full configuration options in the docs.

Origin

PyBA was built for automated intelligence and OSINT — replicating everything a human analyst can do in a browser, but with reproducibility and speed.

If you're doing security research, competitive intelligence, or just automating tedious browser tasks, this is for you.

Status

v0.3.0 - Active development. First stable release: December 18, 2025.

Breaking changes may occur. Pin your version in production.

If PyBA saved you time, consider giving it a ⭐