Delegate Browser Operations to Intelligent Systems 🧠

🌐 This framework provides the most streamlined pathway for augmenting your autonomous agents with direct control over web browser environments.

💡 Explore innovative applications and share your unique automation creations within our Community Forum – showcasing your work is highly encouraged!

🌩️ Eliminate setup friction! Access our instantly available, remotely hosted platform for immediate initiation of web automation procedures! Activate Cloud Instance.

Rapid Initiation Guide

Installation via pip (requires Python version 3.11 or higher):

bash pip install browser-use

Install necessary browser automation dependencies:

bash playwright install

Deploying your autonomous entity:

python from langchain_openai import ChatOpenAI from browser_use import Agent import asyncio from dotenv import load_dotenv load_dotenv()

async def execute_workflow(): orchestrator = Agent( task="Contrast the pricing tiers for gpt-4o versus DeepSeek-V3 across online vendors", llm=ChatOpenAI(model="gpt-4o"), ) outcome = await orchestrator.run() print(outcome)

asyncio.run(execute_workflow())

Populate your .env file with the required secret keys for your chosen large language model provider.

bash OPENAI_API_KEY=

For comprehensive details on configuration parameters, supported models, and advanced features, consult the Official Reference Materials 📚.

Experiencing the Interface via UI

You have the option to pilot the capabilities of this system using the dedicated web-ui repository

Alternatively, initiate the Gradio demonstration interface with these steps:

uv pip install gradio

bash python examples/ui/gradio_demo.py

Illustrative Functionality

Workflow Script: Populate a digital shopping cart with specified grocery items and finalize the transaction.

Instruction Set: Ingest the newest connection from LinkedIn and synchronize this contact into my Salesforce lead management system.

Instruction Set: Analyze my curriculum vitae, identify relevant Machine Learning roles, archive the listings to a local file, and subsequently commence applications within new browser tabs. Seek human intervention if required for ambiguous steps.'

https://github.com/user-attachments/assets/171fb4d6-0355-46f2-863e-edb04a828d04

Instruction Set: Compose a formal letter of gratitude in Google Docs addressed to my grandfather, detailing my appreciation for his support, and export the final document as a PDF file.

Instruction Set: Query the Hugging Face repository for models exclusively licensed under 'cc-by-sa-4.0', rank them based on accumulated endorsements (likes), and save the top five entries to a persistent file.

https://github.com/user-attachments/assets/de73ee39-432c-4b97-b4e8-939fd7f323b3

Additional Use Cases

Explore the comprehensive repository of operational examples within the examples directory or connect with the community on Discord to present your own automated solutions.

Project Vision

Define an objective for your computing system, and observe its successful completion.

Development Trajectory

Agent Core

• [ ] Enhance agent cognitive retention mechanisms (e.g., summarization, data compression, Retrieval-Augmented Generation).
• [ ] Augment strategic planning modules (incorporate website-specific contextual data).
• [ ] Minimize resource utilization (optimize system prompts, refine Document Object Model state representation).

DOM Interpretation Layer

• [ ] Improve parsing precision for intricate UI elements such as calendar widgets, hierarchical selection boxes, and proprietary components.
• [ ] Develop a more robust representation format for the current state of user interface components.

Task Re-execution Capability

• [ ] Implement Large Language Model-driven strategies as a recovery mechanism upon failure.
• [ ] Facilitate straightforward definition of standardized workflow blueprints where the LLM populates the variable details.
• [ ] Enable the agent to output the generated Playwright script upon task completion.

Data Curation

• [ ] Assemble standardized datasets featuring intricate, multi-step procedures.
• [ ] Establish comparative benchmarks across various foundational models for task execution proficiency.
• [ ] Engage in targeted fine-tuning of models specifically for domain-centric web interaction tasks.

User Interaction Paradigm

• [ ] Integrate mechanisms for Human-in-the-Loop verification during critical execution steps.
• [ ] Elevate the visual fidelity and smoothness of the dynamically generated outcome GIFs.
• [ ] Produce diverse demonstration pathways covering tutorials, professional application submissions, quality assurance routines, social media management, etc.

Collaboration Opportunities

We warmly welcome external contributions! Please submit issues for any discovered defects or proposed enhancements. To contribute to documentation upkeep, refer to the contents of the /docs subdirectory.

Local Development Setup

For in-depth knowledge regarding the library's internals, refer to the Local Setup Guide 📕.

Collaborative Commission

We are currently forming a specialized working group focused on establishing industry best practices for the UI/UX design paradigms governing browser agents. We are collectively investigating how refined software architecture can significantly boost the performance metrics of AI agents, granting these enterprises a substantial competitive advantage by positioning their current software ecosystem at the vanguard of the agent-centric era.

Kindly direct an electronic mail to Toby to formally apply for a position within this committee.

Citation Requirement

Should you incorporate Browser Use within academic research or development projects, we respectfully request the inclusion of the following BibTeX citation:

bibtex @software{browser_use2024, author = {Müller, Magnus and Žunič, Gregor}, title = {Browser Use: Enable AI to control your browser}, year = {2024}, publisher = {GitHub}, url = {https://github.com/browser-use/browser-use} }

[](https://x.com/gregpr07) [](https://x.com/mamagnus00)

WIKIPEDIA DIGEST: A web browser operating without a graphical presentation layer is termed a headless browser. These specialized browsers permit automated command over web page rendering and interaction via command-line interfaces or network protocols, mirroring the behavior of standard browsers. They are exceptionally valuable for quality assurance testing, as they accurately interpret HTML, including CSS layout, coloring, typography, and crucial JavaScript/AJAX execution—features often inaccessible through alternative verification methodologies. Native remote operation capabilities have been integrated into major browser engines (Chrome since version 59, Firefox since version 56), rendering previous solutions like PhantomJS largely obsolete.

== Primary Applications == Key domains leveraging headless browsers include:

• Automated validation procedures for contemporary web architectures (web testing).
• Programmatic capture of full-page screenshots.
• Execution of automated unit tests for JavaScript libraries.
• Systematic automation of user interactions with web interfaces.

=== Secondary Utilities === Headless environments are also instrumental for sophisticated web harvesting operations. Google, for instance, noted in 2009 that employing a headless agent could aid in indexing content generated dynamically via Ajax. Conversely, their utilization has faced scrutiny due to potential misuse, such as orchestrating Distributed Denial of Service (DDoS) attacks, inflating digital advertisement metrics, or automating unintended site interactions (e.g., brute-forcing credentials). Nevertheless, a 2018 traffic analysis indicated no discernible preference among malicious entities for headless versus standard browser traffic concerning security breaches like DDoS, SQL injection, or XSS.

== Deployment Methods == Because several leading browsers natively incorporate headless operation through specific interfaces, various software libraries have emerged to unify this control mechanism. Prominent examples are:

• Selenium WebDriver – A conformance-compliant implementation adhering to W3C WebDriver specifications.
• Playwright – A library for automating Chromium, Firefox, and WebKit environments.
• Puppeteer – A library tailored for controlling Chrome or Firefox instances.

=== Test Automation Integration === Numerous testing frameworks incorporate headless browser functionality into their validation toolkits. Examples include:

• Capybara, utilizing either WebKit or Headless Chrome to simulate user actions.
• Jasmine, which defaults to Selenium but supports WebKit or Headless Chrome for browser-based tests.
• Cypress, a dedicated frontend testing framework.
• QF-Test, a commercial tool for GUI-based program testing that supports headless execution.

=== Alternative Approaches == An alternative strategy involves using software that exposes browser-like Application Programming Interfaces (APIs). For example, Deno incorporates browser APIs natively. For Node.js environments, jsdom offers the most comprehensive feature set. While these alternatives generally handle core browser functionalities (HTML parsing, cookies, HTTP requests, limited JavaScript execution), they typically bypass the actual rendering pipeline, resulting in constrained DOM event support. They often achieve faster throughput than full rendering engines.