AI Studio to OpenAI API Adapter

This project provides a lightweight, self-hosted adapter that exposes a mock OpenAI-compatible API ( /v1/chat/completions and /v1/models) and translates the requests into automated UI interactions with Google's AI Studio, powered by Playwright.

Features

• OpenAI API Compatibility: Emulates the v1/chat/completions endpoint for both streaming and non-streaming responses.
• Model Support: Exposes a /v1/models endpoint listing available Gemini models (gemini-2.5-pro, gemini-2.5-flash).
• Cross-Platform: Works on both Windows and macOS, with automatic detection of Chrome's default installation path.
• Highly Configurable:
  • API Request: Supports max_tokens, top_p, and temperature on a per-request basis.
  • Application Settings: Configure browser automation, model capabilities (e.g., web search, code execution), and logging via appsettings.json or environment variables.
• Multi-Account Ready: Designed to cycle through multiple Google accounts to distribute usage.
• Structured Logging: Uses Serilog for structured, configurable, and easily searchable logs.
• UI Automation: Leverages the power of Playwright to interact with the AI Studio web interface reliably.

How It Works

The adapter starts a local Kestrel web server that listens for incoming OpenAI API requests. When a /v1/chat/completions request is received, it performs the following steps:

1. Launches or connects to a Chrome instance with remote debugging enabled.
2. Navigates to Google AI Studio in a new page, cycling through configured user accounts.
3. Selects the requested model (e.g., gemini-2.5-pro).
4. Configures advanced model settings based on the API request (temperature, top_p, etc.) and application configuration (EnableWebSearch, etc.).
5. Pastes the user's prompt into the input area.
6. Clicks the "Run" button and waits for the model to generate a response.
7. Copies the markdown-formatted response from the UI.
8. Transforms the response into the OpenAI ChatCompletion or ChatCompletionStream format and sends it back to the client.

Prerequisites

• .NET 8.0 SDK
• Google Chrome browser

Quick Start

1. Configure the Application

Edit the AIStudio2OpenAI/appsettings.json file to match your setup.

{
  "ChromeAutomation": {
    "ExecutablePath": null,
    // your Chrome.exe executable path
    "UserDataDir": null,
    // a empty folder to store chrome data
    "DebuggingPort": 9222,
    "MaxAccounts": 1
    // maximum number of Google accounts you logged in
  },
  "Gemini": {
    "SetMaxThinkingTokens": true,
    "EnableCodeExecution": false,
    "EnableWebSearch": false
  },
  "Kestrel": {
    "Endpoints": {
      "Http": {
        "Url": "http://localhost:3060"
      }
    }
  }
}

2. Launch Chrome with Remote Debugging

You must launch a dedicated Chrome instance for the adapter to connect to. Close all other Chrome instances first.

On Windows (Command Prompt):

"C:\Program Files\Google\Chrome\Application\chrome.exe" --remote-debugging-port=9222 --user-data-dir="C:\Users\%USERNAME%\Documents\ChromeAgent"

On macOS (Terminal):

/Applications/Google\ Chrome.app/Contents/MacOS/Google\ Chrome --remote-debugging-port=9222 --user-data-dir="$HOME/Documents/ChromeAgent"

In the newly opened Chrome window, log in to your Google Account(s). If using multiple accounts, log them all in.

3. Run the Adapter

Once Chrome is running and you are logged in, open a separate terminal and run the application:

cd AIStudio2OpenAI
dotnet run

The application will prompt you to press [Enter] to confirm Chrome is ready.

4. Use the API

You can now send requests to http://localhost:3060 using any OpenAI-compatible client.

Example using curl:

curl http://localhost:3060/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gemini-1.5-pro",
    "messages": [
      {
        "role": "user",
        "content": "Write a short story about a robot who learns to paint."
      }
    ],
    "temperature": 0.7,
    "max_tokens": 256
  }'

License

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