Gemini CodeAssist Proxy

This local server provides OpenAI (/openai) and Anthropic (/anthropic) compatible endpoints through Gemini CodeAssist (Gemini CLI).

• If you have used Gemini CLI before, it will utilize existing Gemini CLI credentials.
• If you have NOT used Gemini CLI before, you will be prompted to log in to Gemini CLI App through browser.

Gemini 3.0 Pro + GitHub Copilot

Gemini 3.0 Pro paired with GitHub Copilot has proven to be an exceptionally capable combination for AI-assisted development. The model excels at code generation, understanding complex codebases, and providing contextually relevant suggestions.

For this reason, special care has been taken to ensure gemini-cli-proxy is fully compatible with GitHub Copilot, including:

• Proper handling of Gemini's thinking/reasoning tokens
• Server-side caching of thought signatures for multi-turn conversations
• Streaming optimizations for responsive UI feedback

Why

Gemini 3 is cracked. So is GHCP now (I would not have said this in the past). Gemini Ultra is now worth the $ but Gemini 3 + Copilot is too good to give up for it.

Gemini CLI also offers a generous free tier. As of 2025-09-01, the free tier offers 60 requests/min and 1,000 requests/day.

Quick Start

npx gemini-cli-proxy

The server will start on http://localhost:3000

• OpenAI compatible endpoint: http://localhost:3000/openai
• Anthropic compatible endpoint: http://localhost:3000/anthropic

Usage

npx gemini-cli-proxy [options]

Options:

• -p, --port <port> - Server port (default: 3000)
• -g, --google-cloud-project <project> - Google Cloud project ID if you have paid/enterprise tier (default: GOOGLE_CLOUD_PROJECT env variable)
• -l, --log-level <level> - Log level: error, warn, info, debug (default: info)
• --disable-browser-auth - Disables browser auth flow and uses code based auth (default: false)
• --disable-google-search - Disables native Google Search tool (default: false)
• --disable-auto-model-switch - Disables auto model switching in case of rate limiting (default: false)

If you have NOT used Gemini CLI before, you will be prompted to log in to Gemini CLI App through browser. Credentials will be saved in the folder (~/.gemini/oauth_creds.json) used by Gemini CLI.

gemini-3-pro-preview is the default model when you request a model other than gemini-3-pro-preview or gemini-2.5-flash

Use with GitHub Copilot

Requirements: VS Code Insiders is required to use custom OpenAI-compatible endpoints with GitHub Copilot.

Setup Instructions

1. Install VS Code Insiders

   Download from code.visualstudio.com/insiders

2. Start the proxy server

   npx gemini-cli-proxy -p 8084

3. Configure GitHub Copilot in VS Code Insiders

   Open VS Code Insiders settings (JSON) and add:

   {
     "github.copilot.chat.models": [
       {
         "vendor": "copilot",
         "family": "gemini-3-pro",
         "id": "gemini-3-pro-preview",
         "name": "Gemini 3.0 Pro (via gemini-cli-proxy)",
         "version": "gemini-3-pro-preview",
         "capabilities": {
           "agents": true,
           "tokenCounting": false
         },
         "endpoint": {
           "url": "http://localhost:8084/openai/v1/chat/completions"
         }
       }
     ]
   }

or just use the Copilot Chat settings UI

4. Select the model in Copilot Chat

   Open GitHub Copilot Chat and select "Gemini 3.0 Pro (Preview) (via gemini-cli-proxy)" from the model dropdown under "Other Models."

Tips for GitHub Copilot Usage

• The proxy handles Gemini's thinking tokens automatically, displaying them in Copilot's reasoning UI
• For verbose debugging, use --log-level debug when starting the proxy
• If you experience rate limiting, the proxy will automatically switch to fallback models (unless --disable-auto-model-switch is set)

Use with Other Tools

Environment Variables

Most agentic tools rely on environment variables, you can export the following variables

export OPENAI_API_BASE=http://localhost:3000/openai
export OPENAI_API_KEY=ItDoesNotMatter
export ANTHROPIC_BASE_URL="http://localhost:3000/anthropic"
export ANTHROPIC_AUTH_TOKEN=ItDoesNotMatter

Claude Code

Add the following env fields to .claude/settings.json file

{
  "permissions": {
    ...
  },
  "env": {
    "ANTHROPIC_BASE_URL": "http://localhost:3000/anthropic",
    "ANTHROPIC_AUTH_TOKEN": "NotImportant",
    "ANTHROPIC_MODEL": "gemini-3-pro-preview"
  }
}

Zed

Add the following to the Zed config file

{
  "language_models": {
    "openai": {
      "api_url": "http://localhost:3000/openai",
      "available_models": [
        {
          "name": "gemini-2.5",
          "display_name": "localhost:gemini-2.5",
          "max_tokens": 131072
        }
      ]
    }
  }
}

Development

Scripts

• npm run dev - Start development server with hot reload
• npm run build - Build TypeScript to JavaScript
• npm start - Start production server
• npm run lint - Run ESLint
• npm test - Run tests

Project Structure

src/
├── auth/           # Google authentication logic
├── gemini/         # Gemini API client and mapping
├── routes/         # Express route handlers
├── types/          # TypeScript type definitions
└── utils/          # Utility functions

License

Apache-2.0