🚀 Antigravity Copilot

Use Google Gemini, Anthropic Claude, and other powerful AI models directly in GitHub Copilot Chat — for free, via the Antigravity service.

A VS Code extension that bridges the Antigravity AI service with GitHub Copilot's Bring Your Own Key (BYOK) model system. It runs a local proxy, auto-configures Copilot, and gives you access to the latest frontier models without writing any API keys.

✨ Features

🏗️ Architecture

VS Code Copilot Chat
        │
        │  POST /v1/chat/completions
        │  (tools may contain $comment, enumDescriptions — Gemini-incompatible)
        ▼
┌─────────────────────────────────────────┐
│     Antigravity Proxy  (port 8420)      │
│                                         │
│  • sanitizeToolSchemas()                │  ← strips $comment etc.
│  • truncateToolOutput()                 │  ← reduces prompt size
│  • clamp max_tokens                     │  ← prevents quota blowout
│  • rate limiter + concurrency queue     │  ← 429 protection
│  • retry with exponential backoff       │
│  • ThinkingStreamTransformer            │  ← renders <thinking> blocks
└─────────────────────────────────────────┘
        │
        │  Clean request
        ▼
┌─────────────────────────────────────────┐
│     CLIProxyAPI  (port 8317)            │
│     (local binary managed by extension) │
│  • /v1/models  — dynamic model list     │
│  • /v1/chat/completions → routes to:    │
│       Google Gemini API                 │
│       Anthropic Claude API              │
│       OpenAI-compatible APIs            │
└─────────────────────────────────────────┘

📦 Source Files

🤖 Dynamic Model Auto-Discovery

There are no hardcoded model configurations or static lists in the extension. Instead, it is 100% future-proof:

1. Auto-Discovery: On start, the extension calls GET /v1/models on the running CLIProxyAPI to dynamically fetch all available models.
2. Capability Inference: It automatically determines model features and limits based on its ID:
   • Tools: Enabled (true) for all models automatically.
   • Vision: Enabled for Claude/Gemini/vision models; disabled (false) for GPT-OSS models.
   • Thinking: Enabled (true) for Claude 4.6 and reasoning models (IDs containing thinking, sonnet-4-6, opus-4-6).
   • Context Window: Standard models get 128K inputs and 16K outputs. Thinking models get a conservative 32K inputs and 2K outputs to prevent massive generations from instantly exhausting rate-limit quotas.

This means if Antigravity adds new models in the future, they will immediately show up and work in Copilot without needing an extension update!

🛠️ Development Setup

Prerequisites

• Node.js 18+
• VS Code 1.90+
• GitHub Copilot Chat extension installed

Install & Build

git clone https://github.com/Akshar2325/antigravity-copilot-enhanced.git
cd antigravity-copilot
npm install
npm run compile

Run in Development Mode

1. Open the project in VS Code
2. Press F5 → opens a new Extension Development Host window
3. In the new window, the Antigravity status bar item appears bottom-left
4. Click it → Start Server (or set antigravityCopilot.server.autoStart: true)
5. Models auto-register in Copilot Chat

Watch Mode (auto-rebuild on save)

npm run watch

Package as VSIX

npm run package
# Output: antigravity-copilot-x.y.z.vsix

Install VSIX

# Via CLI
code --install-extension antigravity-copilot-1.6.0.vsix

# Via UI
# Extensions panel → ··· → Install from VSIX...

⚙️ Configuration

Open VS Code Settings and search for antigravityCopilot:

🔑 How Models Are Added to Copilot

The extension writes to VS Code's chatLanguageModels.json (BYOK config file):

• Windows: %APPDATA%\Code\User\chatLanguageModels.json
• macOS: ~/Library/Application Support/Code/User/chatLanguageModels.json
• Linux: ~/.config/Code/User/chatLanguageModels.json

Each model entry looks like:

{
  "name": "Antigravity",
  "vendor": "customendpoint",
  "models": [
    {
      "id": "gemini-3-flash-agent",
      "name": "Antigravity: Gemini 3 Flash Agent",
      "url": "http://127.0.0.1:8420/v1",
      "toolCalling": true,
      "vision": true,
      "thinking": false,
      "maxInputTokens": 1048576,
      "maxOutputTokens": 8192
    }
  ]
}

The url always points to the proxy (port 8420), not CLIProxyAPI directly — this ensures schema sanitization and rate limiting are always applied.

🐛 Troubleshooting

$comment / enumDescriptions 400 error from Gemini

Root cause: GitHub Copilot injects JSON Schema meta-keywords ($comment, enumDescriptions) into tool definitions. Gemini's API rejects these.

Fix: Already handled automatically by sanitizeToolSchemas() in the proxy. Make sure antigravityCopilot.proxy.enabled is true.

Models not appearing in Copilot

1. Ensure the Antigravity server is running (status bar shows $(broadcast) Antigravity: ON)
2. Run command: Antigravity: Configure Copilot Models
3. Reload VS Code window (Ctrl+Shift+P → Developer: Reload Window)
4. Open Copilot Chat → model picker → Manage Models → search for "Antigravity"

429 / Rate limit errors

Thinking models (Claude Opus Thinking) have strict upstream rate limits. The proxy automatically:

• Queues concurrent requests (1 at a time for thinking models)
• Retries with exponential backoff (up to 5 retries)
• Shows a notification with the cooldown period

Reduce context sent by enabling tool output truncation (on by default).

📄 License

MIT — see LICENSE