BrowserGPT

This project allows you to control your browser using natural language. It integrates OpenAI's GPT-4 with the Playwright library, enabling seamless browser navigation. GPT-4 generates code snippets, which Playwright executes to carry out specified tasks.

Demo

Installation

Install the required packages:

npm install

Create a .env file in the project root directory and add the following line:

OPENAI_API_KEY=your_openai_api_key

Replace your_openai_api_key with your actual OpenAI API key.

Run the script:

npm run start

Options:

Usage: npm run start -- [options]

Options:
  -a, --autogpt                          run with autogpt (default: false)
  -m, --model <model>                    openai model to use (default: "gpt-4-1106-preview")
  -o, --outputFilePath <outputFilePath>  path to store test code
  -u, --url <url>                        url to start on (default: "https://www.google.com")
  -v, --viewport <viewport>              viewport size to use (default: "1280,720")
  -h, --help                             display help for command

Usage

The script opens a browser window.

In the terminal, you'll be prompted to enter a task.

Type your task using natural language (e.g., "Generate an interesting phrase and type it into Google") and press Enter.

GPT-4 can recognize buttons and text on the page and will navigate the browser to complete the specified task.

To stop the script, press Ctrl + C in the terminal.

Examples

Here are some example tasks you can input:

• go to hn
• click on the abc article
• enter [email protected] into the email box. John and Doe in the first and last name boxes respectively
• generate a spicy comment on what xyz said and put it in the comment box

With autogpt enabled, you can also input more complex tasks like:

• go to hn and click on the first article
• use bing and find the abc article

Limitations

This script serves as a demonstration of GPT-4 and Playwright integration, and may not perform flawlessly for every task or website. Generated code snippets could fail to execute, or the model might not comprehend specific inputs. Consider providing a more detailed task description or rephrasing your input in these situations. Some websites might be too large to fit in the prompt for smaller models like base gpt-4, hence we default to gpt-4-1106-preview with 125k tokens.

License

This project is licensed under the MIT License. See the LICENSE file for details.