Antigravity Proxy

Use any AI provider with Antigravity 2.0 — NVIDIA, OpenRouter, OpenAI, Groq, Anthropic, Google Gemini, OpenCode Zen, or 9+ local inference solutions.

The proxy intercepts Antigravity's Google Gemini API calls and translates them to the target provider's format. It supports:

• Multi-provider failover with retry + exponential backoff
• Per-model, per-provider routing via a matrix UI
• Reasoning effort control
• Real-time hot-reload of all config
• Advanced tool normalization and capability detection

Quick Start

Get the Repository

git clone https://github.com/12errh/antigravity-proxy.git
cd antigravity-proxy

All launcher scripts and source code are inside the cloned directory.

Prerequisites

• Antigravity 2.0 Desktop (Windows, macOS, or Linux)
• Node.js 20+
• Administrator/root privileges (port 443; can use 8443 without admin)
• At least one API key or local model server

Platform Support

TypeScript code is cross-platform. Only launcher scripts and TLS cert trust are platform-specific.

Launch

Windows (PowerShell as Administrator)

.\start.ps1

macOS / Linux

chmod +x start.sh
./start.sh

Without admin (port 8443)

./start.sh --port 8443

Then open http://localhost:4000 in your browser to configure providers, model mappings, and view live stats.

🎯 Key Features

🚀 How It Works

graph TD
    A[Antigravity 2.0 Desktop] --> B[TLS 443]
    B --> C[Proxy]
    C --> D[Cloud APIs]
    C --> E[Local APIs]
    C --> F[Dashboard]

    D --> G[OpenRouter]
    D --> H[NVIDIA]
    D --> I[OpenAI]
    D --> J[Groq]
    D --> K[Anthropic]
    D --> L[Google Gemini]
    D --> M[Zen]

    E --> N[Ollama]
    E --> O[vLLM]
    E --> P[LM Studio]
    E --> Q[llama.cpp]

    F --> R[Live Logs]
    F --> S[Request History]
    F --> T[Config Editor]
    F --> U[Model Matrix]
    F --> V[Provider Priority]

Per Request Pipeline

1. TLS Intercept — Receives Gemini API call on port 443
2. Context Stripping — Removes bulk context, injects compact agent-context.md reference
3. Tool Normalization — Resolves aliases, coerces types, fills defaults
4. Model Resolution — Routes through priority order with retry + backoff
5. Capability Detection — Auto-detects reasoning/vision/tool support
6. Reasoning Effort — Applies reasoning_effort if configured
7. Translation — Converts to target provider's API format
8. Streaming — Returns SSE response wrapped in Gemini format
9. Logging — Records to SQLite for cost tracking, history, sessions

📊 Supported Providers

Cloud APIs

Local Inference

📋 Models Tab — Matrix

The Models tab shows a matrix: one row per Antigravity model alias, one column per provider.

graph TD
    A[Model Alias] --> B[Default]
    A --> C[OpenRouter]
    A --> D[Zen]

    E[claude-sonnet-4-6] --> F[claude-sonnet-4-5]
    E --> G[anthropic/claude-3-5-sonnet-20241022]
    E --> H[claude-sonnet-4-5]

    I[gemini-2.5-flash] --> J[gemini-2.5-flash]
    I --> K[google/gemini-2.0-flash-exp]
    I --> L[deepseek-v4-flash]

    M[gpt-4o] --> N[openai/gpt-4o]
    M --> O[openai/gpt-4o]

Features:

• Double-click any cell to open live model picker
• Column toggles to hide unused providers
• Quick-add buttons (+ Claude, + Gemini, + GPT)
• Empty cells use Default value
• Provider logos in filled cells

🧠 Model Options — Reasoning Effort

Some models support reasoning_effort to control thinking depth:

Auto-detection: The Model Options tab detects supported models and lets you set levels per model. Settings persist in proxy/reasoning-effort.json without restart.

⚙️ Config Tab — API Keys

All keys are hot-reloaded from .env — no restart needed.

🏗️ Architecture

graph TD
    %% Root level
    A[agent-context.md] --> B[start.ps1]
    B --> C[README.md]
    C --> D[CHANGELOG.md]
    C --> E[docs/]
    C --> F[proxy/]

    %% docs/ directory
    E --> G[SETUP.md]
    E --> H[CONFIGURATION.md]
    E --> I[DEVELOPER.md]
    E --> J[antigravity-v2-analysis.md]

    %% proxy/ directory
    F --> K[src/]

    %% src/ directory - main components
    K --> L[adapters/]
    K --> M[adapter.ts]
    K --> N[provider-plugin.ts]
    K --> O[provider-registry.ts]
    K --> P[plugins/]
    K --> Q[tool-capabilities.ts]
    K --> R[tool-normalizer.ts]
    K --> S[model-capabilities.ts]
    K --> T[router.ts]
    K --> U[models.ts]
    K --> V[index.ts]
    K --> W[engine.ts]
    K --> X[mapper.ts]
    K --> Y[config.ts]
    K --> Z[provider-cache.ts]

    %% adapters subdirectory
    L --> A1[openai.ts]
    L --> A2[anthropic.ts]
    L --> A3[google.ts]
    L --> A4[groq.ts]
    L --> A5[zen.ts]
    L --> A6[nvidia.ts]
    L --> A7[types.ts]

    %% plugins subdirectory
    P --> B1[builtin-plugins.ts]

    %% dashboard subdirectory
    F --> C1[dashboard/]
    C1 --> C2[index.html]
    C1 --> C3[login.html]

    %% Other files in proxy/
    F --> D1[models.json]
    F --> D2[pricing.json]
    F --> D3[reasoning-effort.json]
    F --> D4[blocklist.json]
    F --> D5[.env]
    F --> D6[.env.example]
    F --> D7[certs/]
    F --> D8[logs/]

    %% Note: actual configuration for certs and logs are nested

---

## 🔧 Environment Variables

```env
# Provider priority (first = primary)
PROVIDER_PRIORITY=openrouter,nvidia,zen

# API keys
OPENROUTER_API_KEY=sk-or-v1-...
NVIDIA_API_KEY=nvapi-...
OPENAI_API_KEY=sk-...
GROQ_API_KEY=gsk_...
ANTHROPIC_API_KEY=sk-ant-...
GOOGLE_API_KEY=AIza...
OPENCODE_API_KEY=sk-...

# Ports
PROXY_PORT=443          # TLS port Antigravity connects to
API_PORT=4000           # Dashboard + REST API port

# Retry & failover
PROXY_RETRIES=3         # Attempts per provider
PROXY_BACKOFF_MS=100    # Initial backoff (doubles each retry)

# Rate limiting (0 = unlimited)
RATE_LIMIT_GLOBAL=0
RATE_LIMIT_PROVIDER=0
RATE_LIMIT_WINDOW_MS=60000

# Logging
LOG_LEVEL=info          # debug | info | warn | error
LOG_MAX_SIZE_MB=10      # Rotate when reached
LOG_MAX_FILES=5         # Keep N rotated files
LOG_MAX_AGE_DAYS=30     # Delete files older than

# Dashboard auth (optional)
DASHBOARD_USER=admin
DASHBOARD_PASSWORD=your_password

# Failover webhook (optional)
FAILOVER_WEBHOOK_URL=https://hooks.example.com/webhook

# Workspace context hardening
WORKSPACE_CONTEXT_ENVELOPE=strict   # off | loose | strict

# Inline context mode
CONTEXT_STRIP_MODE=passthrough       # passthrough (default) | strip

💡 Why Passthrough is the Default

We originally built an external context system (agent-context.md + antigravity-context.ts) to teach external models about Antigravity's tools. But after implementation, we discovered the external context grew to ~28K tokens — the exact same size as the native Antigravity context that Gemini receives.

Since both consume identical tokens, there's no benefit to stripping the native context and injecting our own. Passthrough simply forwards what Antigravity already sends — simpler, no injection overhead, and external models understand it natively (validated: 18/20 tools working with MiMo-v2.5).

Future goal: Compress the external context to deliver the same native quality with fewer tokens, reducing the ~28K token overhead while maintaining full tool coverage.

📚 Quick Links

• Setup Guide
• Configuration Reference
• Developer Guide ← For adding providers & plugin development
• Antigravity v2 Protocol Analysis
• CHANGELOG
• Contributing
• Security Policy

⚡ Development Quick Start

cd proxy
npm install            # Install dependencies
npm run build          # Compile TypeScript → dist/
npm run typecheck      # Type-check only
npm test               # Run tests
npm run dev            # Watch mode

New Providers: See docs/DEVELOPER.md for plugin architecture guide.

Run tests:

# All tests
npm test

# Filter by component
npx tsx test/run.ts plugin-architecture
npx tsx test/run.ts tool-translation
npx tsx test/run.ts model-discovery
npx tsx test/run.ts provider-adapters

📝 License

MIT — see LICENSE.