Antigravity for Copilot

Expose Google Antigravity models to VS Code via Copilot's official BYOK (Bring Your Own Key) interface - Manages CLIProxyAPI server lifecycle and configures custom language models using VS Code's supported extension APIs.

⚠️ Disclaimer: This extension uses VS Code's official Language Model API for custom model configuration. It does not modify GitHub Copilot internals, intercept Copilot traffic, or patch any Copilot files. This project is unofficial and not affiliated with GitHub, Microsoft, Google, or Anthropic.

🌟 Features

• One-Click Server Management: Start/stop CLIProxyAPI directly from VS Code
• Automatic Configuration: Creates default config.yaml if missing
• BYOK Model Registration: Registers Antigravity models using VS Code's official Language Model API
• Status Bar Integration: Quick visual status and server controls
• Sidebar Dashboard: Monitor server status, view available models, and manage settings
• Auto-Start Support: Configure the server to start automatically with VS Code
• Authentication Launcher: Launches Antigravity's authentication flow via CLIProxyAPI
• Rate Limiting: Built-in rate limiter to prevent 429 errors with thinking models
• Optional Throttling Proxy: Local proxy that queues BYOK requests and can clamp max output tokens to reduce upstream 429s

🤖 Available Models (10)

Model	Description	Capabilities
Claude Sonnet 4.5	Latest Claude model	Tools
Claude Sonnet 4.5 (Thinking)	Extended thinking mode	Tools, Thinking
Claude Opus 4.5 (Thinking)	Most powerful Claude	Tools, Thinking
Gemini 2.5 Flash	Fast Gemini model	Tools
Gemini 2.5 Flash Lite	Lightweight Gemini	Tools
Gemini 3 Pro (Preview)	Latest Gemini Pro	Tools
Gemini 3 Flash (Preview)	Latest Gemini Flash	Tools
Gemini 3 Pro Image (Preview)	Gemini with vision	Tools, Vision
Gemini 2.5 Computer Use	Computer interaction	Tools, Vision
gpt-oss-120b-medium	Open source model	Basic

📦 Prerequisites

1. VS Code Insiders (required for custom models support)

   winget install --id Microsoft.VisualStudioCode.Insiders
   

2. GitHub Copilot Pro subscription

3. GitHub Copilot Extensions (pre-release versions)

   code-insiders --install-extension github.copilot --pre-release
   code-insiders --install-extension github.copilot-chat --pre-release
   

4. CLIProxyAPI installed in %USERPROFILE%\CLIProxyAPI\

   The extension can automatically download and install the latest version for you when you attempt to start the server. Alternatively, you can install it manually:

   $zipPath = "$env:TEMP\CLIProxyAPI.zip"
   $extractPath = "$env:USERPROFILE\CLIProxyAPI"
   # Download latest from GitHub
   Invoke-WebRequest -Uri "https://github.com/router-for-me/CLIProxyAPI/releases/latest/download/CLIProxyAPI_windows_amd64.zip" -OutFile $zipPath
   Expand-Archive -Path $zipPath -DestinationPath $extractPath -Force
   Remove-Item $zipPath
   

Model Enablement

After configuring models, you must manually enable them in VS Code:

1. Open Copilot Chat (Ctrl+Alt+I)
2. Click the model picker dropdown
3. Select "Manage Models..."
4. Enable these recommended models:
   • Gemini 3 Pro (Preview)
   • Gemini 3 Flash (Preview)
   • Claude Opus 4.5 (Thinking)
5. Click the 👁️ eye icon next to each model to enable it

Note: Models can only be enabled manually through the VS Code UI. Programmatic enablement is not supported by the current VS Code BYOK API.

5. CLIProxyAPI Configuration: (Optional) The extension will automatically create a default config.yaml if one doesn't exist.
   • Default location: %USERPROFILE%\CLIProxyAPI\config.yaml
   • Default content:

     port: 8317
     host: "127.0.0.1"
     auth-dir: "C:\\Users\\<USERNAME>\\.cli-proxy-api"
     providers:
       antigravity:
         enabled: true
     

🚀 Quick Start

Step 1: Install the Extension

1. Download the .vsix file
2. Open VS Code Insiders
3. Press Ctrl+Shift+P → "Extensions: Install from VSIX..."
4. Select the downloaded file

Step 2: Authenticate

1. Click the Antigravity icon in the Activity Bar
2. Click "Login to Antigravity" button
3. Follow the authentication flow in the terminal (Server will stop temporarily during login)

Step 3: Start Server

1. Click "Start Server" in the sidebar
2. Wait for the server to start (status will turn green)

Step 4: Configure Models

1. Click "Configure Models" button
2. Reload VS Code when prompted
3. Open Copilot Chat (Ctrl+Alt+I)
4. Click the model picker dropdown → "Manage Models..."
5. Find the Antigravity models and click the eye icon to enable them
6. The models will now appear in the model picker dropdown

⚙️ Configuration

Open VS Code Settings (Ctrl+,) and search for antigravityCopilot:

Server Settings

Setting	Default	Description
server.enabled	false	Enable server on startup
server.autoStart	false	Auto-start server with VS Code
server.executablePath	(auto)	Path to cli-proxy-api.exe
server.port	8317	Starting port (auto-increments if in use)
server.host	127.0.0.1	Server host
autoConfigureCopilot	true	Auto-configure models on startup
showNotifications	true	Show notifications

Rate Limiting Settings

Rate limiting provides a safety net for thinking models. The primary 429 mitigation is now aggressive retries matching Antigravity IDE's behavior.

Setting	Default	Description
rateLimit.enabled	true	Enable rate limiting
rateLimit.cooldownMs	5000	Base cooldown between requests
rateLimit.showNotifications	true	Show notifications when blocked

Exponential Backoff

When consecutive 429 errors occur, the rate limiter automatically applies exponential backoff:

• Each consecutive 429 doubles the effective cooldown (up to 5× the base)
• The backoff resets after a successful request
• Check current backoff status via Command Palette → "Antigravity: Rate Limit Status"

This prevents hammering the upstream server when quota is exhausted.

Proxy Settings

The optional throttling proxy queues requests to prevent upstream 429 errors with thinking models. Thinking models use reduced token limits (maxInputTokens: 32000, maxOutputTokens: 2048) to minimize quota burn.

Setting	Default	Description
proxy.enabled	true	Enable the local throttling proxy
proxy.host	127.0.0.1	Proxy bind host
proxy.port	8420	Starting port for proxy (auto-increments if in use)
proxy.rewriteMaxTokens	true	Clamp output tokens to reduce long generations
proxy.maxTokensThinking	1024	Max output tokens for Thinking models
proxy.maxTokensStandard	4096	Max output tokens for standard models
proxy.logRequests	true	Log request metadata (model, status, duration)
proxy.transformThinking	true	Transform streaming responses for thinking display
proxy.thinkingTransformMode	none	Transform mode: none, annotate, enhanced, or claude
proxy.thinkingTimeoutMs	60000	Timeout for Thinking requests (abort long runs)
proxy.requestTimeoutMs	120000	Timeout for standard requests
proxy.truncateToolOutput	true	Truncate very large tool outputs (e.g., git diff)
proxy.maxToolOutputChars	12000	Max chars kept per tool output after truncation
proxy.toolOutputHeadChars	6000	Chars kept from start of tool output
proxy.toolOutputTailChars	2000	Chars kept from end of tool output
proxy.maxRequestBodyBytes	10485760	Max request body size; returns 413 if exceeded
proxy.thinkingConcurrency	1	Max concurrent requests for Thinking models
proxy.standardConcurrency	3	Max concurrent requests for standard models
proxy.maxRetries	3	Retry attempts for 429 errors (0 to disable)
proxy.retryBaseDelayMs	1000	Base delay before first retry (exponential backoff)

Thinking Transform Modes

The proxy can transform streaming responses from Thinking models to help clients display reasoning content:

• none: Direct passthrough without any transformation (most compatible)
• annotate (default): Adds minimal _is_thinking markers to delta objects
• enhanced: Adds comprehensive thinking block markers in OpenAI format
• claude: Full conversion to Anthropic/Claude streaming format (experimental)

Timeouts (prevent long-thinking quota burn)

Long “thinking” runs can consume a lot of quota (even if you didn’t request a huge visible answer). The proxy can abort long-running requests:

• antigravityCopilot.proxy.thinkingTimeoutMs (default: 60s)
• antigravityCopilot.proxy.requestTimeoutMs (default: 120s)

Tool output truncation (reduce context size)

Copilot Chat tool calls like git diff can produce very large outputs that get embedded into subsequent requests. This increases prompt size and can trigger upstream RESOURCE_EXHAUSTED.

When enabled, the proxy truncates only messages with role: "tool" that exceed your configured limits:

• antigravityCopilot.proxy.truncateToolOutput: true
• antigravityCopilot.proxy.maxToolOutputChars: 12000
• antigravityCopilot.proxy.toolOutputHeadChars: 6000
• antigravityCopilot.proxy.toolOutputTailChars: 2000

If you want to hard-limit request size regardless, set:

• antigravityCopilot.proxy.maxRequestBodyBytes (default: 10MB)

Concurrency queue (prevent request bursts)

Copilot Chat can fire multiple requests per prompt (tools, retries, follow-ups). For resource-intensive Thinking models, this can trip upstream quota even if you only clicked once.

The proxy uses a semaphore-based concurrency queue with separate limits for thinking vs standard models:

• antigravityCopilot.proxy.thinkingConcurrency: 1 (keep low to avoid exhaustion)
• antigravityCopilot.proxy.standardConcurrency: 3

Excess requests queue until a slot opens. Thinking requests have lower priority than standard requests.

Retry with exponential backoff (Antigravity IDE-style)

When 429 or RESOURCE_EXHAUSTED errors occur, the proxy automatically retries with aggressive short-delay retries matching Antigravity IDE's approach:

• antigravityCopilot.proxy.maxRetries: 5 (set to 0 to disable)
• antigravityCopilot.proxy.retryBaseDelayMs: 100 (almost immediate first retry)

Retry delays: ~200ms → ~400ms → ~800ms → ~1.6s → ~3.2s. Most 429 errors resolve within 2-3 retries.

Example settings.json

{
  "antigravityCopilot.server.autoStart": true,
  "antigravityCopilot.autoConfigureCopilot": true,
  "antigravityCopilot.showNotifications": true,
  "antigravityCopilot.rateLimit.enabled": true,
  "antigravityCopilot.rateLimit.cooldownMs": 5000,
  "antigravityCopilot.proxy.enabled": true,
  "antigravityCopilot.proxy.thinkingConcurrency": 1,
  "antigravityCopilot.proxy.standardConcurrency": 3,
  "antigravityCopilot.proxy.maxRetries": 5,
  "antigravityCopilot.proxy.retryBaseDelayMs": 100,
  "antigravityCopilot.proxy.thinkingTimeoutMs": 60000,
  "antigravityCopilot.proxy.requestTimeoutMs": 120000,
  "antigravityCopilot.proxy.truncateToolOutput": true,
  "antigravityCopilot.proxy.maxToolOutputChars": 12000,
  "antigravityCopilot.proxy.toolOutputHeadChars": 6000,
  "antigravityCopilot.proxy.toolOutputTailChars": 2000,
  "antigravityCopilot.proxy.maxRequestBodyBytes": 10485760
}

🎮 Commands

Access commands via Command Palette (Ctrl+Shift+P):

• Antigravity: Start Server - Start the CLIProxyAPI server
• Antigravity: Stop Server - Stop the server
• Antigravity: Restart Server - Restart the server
• Antigravity: Login to Antigravity - Authenticate with Google
• Antigravity: Configure Models - Add models to Copilot Chat
• Antigravity: Show Server Controls - Open quick controls menu
• Antigravity: Rate Limit Status - View and manage rate limiter status

❓ Troubleshooting

Server won't start

• Verify CLIProxyAPI is installed at %USERPROFILE%\CLIProxyAPI\cli-proxy-api.exe
• Check if port 8317 is already in use: netstat -ano | findstr :8317
• Review logs: Click "Show Logs" in the dashboard

Models not appearing in Copilot Chat

• Ensure you're using VS Code Insiders (not stable VS Code)
• Ensure Copilot extensions are pre-release versions
• Click "Configure Models" and reload VS Code
• Check if Custom OpenAI feature is available (gradual rollout)

Authentication failed

• Run Antigravity: Login to Antigravity command
• Follow the browser authentication flow
• Check auth files in %USERPROFILE%\.cli-proxy-api\

Rate limit (429) errors

429 RESOURCE_EXHAUSTED indicates a quota or concurrency limit on the server side — not a syntax error. Common causes:

• Too many concurrent requests (especially with Thinking models)
• Hitting model-provider GPU/compute quota
• Very large context causing expensive internal passes
• Repeated long "thinking" requests consuming extra backend resources

Quick fixes:

• Enable rate limiting: antigravityCopilot.rateLimit.enabled: true
• Increase cooldown if errors persist: antigravityCopilot.rateLimit.cooldownMs: 30000
• Enable tool output truncation to reduce context size
• Check rate limiter status via Command Palette → "Antigravity: Rate Limit Status"
• Reset the rate limiter from the sidebar dashboard if needed

The rate limiter applies exponential backoff automatically after consecutive 429s.

If 429 happens repeatedly with Thinking models

Copilot Chat can send multiple requests per prompt (tools, retries, follow-ups). For resource-intensive Thinking models, this can trip upstream quota/throttling even if you only clicked once.

This extension includes an optional local throttling proxy that queues requests before they reach CLIProxyAPI. It does not modify Copilot internals; it simply changes the BYOK endpoint URL Copilot uses.

1. Enable the proxy:
   • antigravityCopilot.proxy.enabled: true
2. Re-run Antigravity: Configure Models, then reload VS Code.
3. Use a longer cooldown (start with 30–60s):
   • antigravityCopilot.rateLimit.cooldownMs: 60000

If you still see RESOURCE_EXHAUSTED immediately, your Antigravity account/model quota may be exhausted; switch to a lighter model or wait for quota reset.

Diagnosing 429s (proxy request logging)

If you want to understand why 429s happen (bursting, large requested outputs, etc.), you can enable proxy request logging.

• Setting: antigravityCopilot.proxy.logRequests: true
• What it logs: request metadata only (endpoint, model, token limits, status code, duration)
• What it does not log: your prompt text or chat content

Open the Antigravity output channel to view [PROXY ...] log lines.

🛠️ Building from Source

Prerequisites

1. Node.js (v18 or later)

   winget install OpenJS.NodeJS.LTS
   

2. VS Code Extension Manager (vsce)

   npm install -g @vscode/vsce
   

Build Steps

1. Clone the repository

   git clone https://github.com/punal100/antigravity-copilot.git
   cd antigravity-copilot
   

2. Install dependencies

   npm install
   

3. Compile TypeScript

   npm run compile
   

4. Package the extension

   # Using npm script
   npm run package
   
   # Or directly with vsce
   vsce package
   

   This creates a .vsix file (e.g., antigravity-copilot-1.5.3.vsix) in the project root.

Development

• Watch mode (auto-recompile on changes):

  npm run watch
  

• Lint the code:

  npm run lint
  

Installing the Built Extension

1. Open VS Code Insiders
2. Press Ctrl+Shift+P → "Extensions: Install from VSIX..."
3. Select the generated .vsix file

🔗 Resources

• Repository
• Issues
• License
• Microsoft (Copilot BYOK docs)
• CLIProxyAPI GitHub

📝 License

MIT License

🏗️ Architecture

This extension:

1. Manages CLIProxyAPI: A local OpenAI-compatible proxy server that launches and manages Antigravity authentication via CLIProxyAPI
2. Registers Models via BYOK: Uses VS Code/Copilot's BYOK setting github.copilot.chat.customOAIModels to register custom OpenAI-compatible endpoints
3. Displays Status: Provides a sidebar UI for server management and status monitoring

No Copilot internals are modified. The extension only uses documented VS Code APIs and settings.

Compliance Boundary

This extension explicitly does NOT:

• ❌ Modify GitHub Copilot internals or files
• ❌ Host or redistribute any AI models
• ❌ Collect, store, or transmit user credentials
• ❌ Intercept or proxy GitHub Copilot’s own service traffic
• ❌ Provide access to Antigravity (users must obtain access independently)
• ❌ Connect to any internal/private services

Notes:

• ✅ If you enable antigravityCopilot.proxy.enabled, the extension runs an optional local throttling proxy only for the BYOK endpoint you configured (Copilot → your local endpoint). This is used to queue requests and reduce upstream 429s.

Credits

• Punal Manalan - Author and maintainer
• CLIProxyAPI - The proxy server powering this extension

---

⚠️ Legal Notice

This extension requires CLIProxyAPI and a Google account with Antigravity access.

• This project does not provide access to Antigravity — users must obtain access independently
• This project is unofficial and not affiliated with GitHub, Microsoft, Google, Anthropic, or OpenAI
• Users are responsible for ensuring their use complies with all applicable terms of service
• The authors assume no liability for any misuse or ToS violations
• Antigravity access may be subject to eligibility requirements or usage policies set by Google