feat(providers): add Antigravity provider for unified Claude/Gemini access

Implements a new provider that uses Google's Antigravity unified gateway API to access Claude, Gemini, and other models through a single OAuth2-authenticated endpoint. Features: - OAuth2 token management with automatic refresh - Multi-account rotation for rate limit distribution - Support for Claude Opus/Sonnet 4.5 (with/without thinking) - Support for Gemini 2.5/3 models (Pro/Flash variants) - Thinking mode support with configurable tokens - Image processing support - Dual quota pool tracking (antigravity vs gemini-cli) - Gemini-style API request format Authentication: - Reads from ANTIGRAVITY_REFRESH_TOKEN env var (priority) - Falls back to ~/.config/opencode/antigravity-accounts.json - Automatic token refresh with retry logic - Rate limit tracking per account and quota pool Files added: - providers/antigravity.py - Main provider implementation - providers/antigravity_auth.py - OAuth token manager - providers/registries/antigravity.py - Model registry - conf/antigravity_models.json - Model definitions (11 models) - docs/antigravity_provider.md - Setup and usage docs - tests/test_antigravity_provider.py - Unit tests (14 pass) Integration: - Added to provider priority order after ZEN - Registered in server.py with auto-detection - ToS warning logged on first use
2026-02-01 17:55:26 +01:00
parent c71a535f16
commit 5add230d4c
10 changed files with 1612 additions and 1 deletions
--- a/conf/antigravity_models.json
+++ b/conf/antigravity_models.json
@@ -0,0 +1,140 @@
+{
+  "$schema": "https://json-schema.org/draft/2020-12/schema",
+  "description": "Antigravity model configurations for the PAL MCP server",
+  "models": [
+    {
+      "model_name": "claude-opus-4-5-thinking",
+      "friendly_name": "Claude Opus 4.5 Thinking (Antigravity)",
+      "aliases": ["antigravity-claude-opus-4-5-thinking", "ag-opus-thinking"],
+      "context_window": 200000,
+      "max_output_tokens": 64000,
+      "supports_extended_thinking": true,
+      "max_thinking_tokens": 32768,
+      "supports_images": true,
+      "intelligence_score": 20,
+      "temperature_constraint": "range"
+    },
+    {
+      "model_name": "claude-sonnet-4-5-thinking",
+      "friendly_name": "Claude Sonnet 4.5 Thinking (Antigravity)",
+      "aliases": ["antigravity-claude-sonnet-4-5-thinking", "ag-sonnet-thinking"],
+      "context_window": 200000,
+      "max_output_tokens": 64000,
+      "supports_extended_thinking": true,
+      "max_thinking_tokens": 32768,
+      "supports_images": true,
+      "intelligence_score": 19,
+      "temperature_constraint": "range"
+    },
+    {
+      "model_name": "claude-sonnet-4-5",
+      "friendly_name": "Claude Sonnet 4.5 (Antigravity)",
+      "aliases": ["antigravity-claude-sonnet-4-5", "ag-sonnet"],
+      "context_window": 200000,
+      "max_output_tokens": 64000,
+      "supports_extended_thinking": false,
+      "supports_images": true,
+      "intelligence_score": 18,
+      "temperature_constraint": "range"
+    },
+    {
+      "model_name": "gemini-3-pro-high",
+      "friendly_name": "Gemini 3 Pro High (Antigravity)",
+      "aliases": ["antigravity-gemini-3-pro", "ag-gemini-3-pro"],
+      "context_window": 1048576,
+      "max_output_tokens": 65535,
+      "supports_extended_thinking": true,
+      "max_thinking_tokens": 65535,
+      "supports_images": true,
+      "intelligence_score": 20,
+      "temperature_constraint": "range"
+    },
+    {
+      "model_name": "gemini-3-pro-low",
+      "friendly_name": "Gemini 3 Pro Low (Antigravity)",
+      "aliases": ["ag-gemini-3-pro-low"],
+      "context_window": 1048576,
+      "max_output_tokens": 65535,
+      "supports_extended_thinking": true,
+      "max_thinking_tokens": 16384,
+      "supports_images": true,
+      "intelligence_score": 19,
+      "temperature_constraint": "range"
+    },
+    {
+      "model_name": "gemini-3-flash",
+      "friendly_name": "Gemini 3 Flash (Antigravity)",
+      "aliases": ["antigravity-gemini-3-flash", "ag-gemini-3-flash"],
+      "context_window": 1048576,
+      "max_output_tokens": 65536,
+      "supports_extended_thinking": true,
+      "max_thinking_tokens": 32768,
+      "supports_images": true,
+      "intelligence_score": 17,
+      "temperature_constraint": "range"
+    },
+    {
+      "model_name": "gemini-3-flash-preview",
+      "friendly_name": "Gemini 3 Flash Preview (Gemini CLI)",
+      "aliases": ["ag-gemini-3-flash-preview"],
+      "context_window": 1048576,
+      "max_output_tokens": 65536,
+      "supports_extended_thinking": true,
+      "max_thinking_tokens": 32768,
+      "supports_images": true,
+      "intelligence_score": 17,
+      "temperature_constraint": "range",
+      "quota_pool": "gemini-cli"
+    },
+    {
+      "model_name": "gemini-3-pro-preview",
+      "friendly_name": "Gemini 3 Pro Preview (Gemini CLI)",
+      "aliases": ["ag-gemini-3-pro-preview"],
+      "context_window": 1048576,
+      "max_output_tokens": 65535,
+      "supports_extended_thinking": true,
+      "max_thinking_tokens": 65535,
+      "supports_images": true,
+      "intelligence_score": 19,
+      "temperature_constraint": "range",
+      "quota_pool": "gemini-cli"
+    },
+    {
+      "model_name": "gemini-2.5-pro",
+      "friendly_name": "Gemini 2.5 Pro (Gemini CLI)",
+      "aliases": ["ag-gemini-2.5-pro"],
+      "context_window": 1048576,
+      "max_output_tokens": 65536,
+      "supports_extended_thinking": true,
+      "max_thinking_tokens": 32768,
+      "supports_images": true,
+      "intelligence_score": 18,
+      "temperature_constraint": "range",
+      "quota_pool": "gemini-cli"
+    },
+    {
+      "model_name": "gemini-2.5-flash",
+      "friendly_name": "Gemini 2.5 Flash (Gemini CLI)",
+      "aliases": ["ag-gemini-2.5-flash"],
+      "context_window": 1048576,
+      "max_output_tokens": 65536,
+      "supports_extended_thinking": true,
+      "max_thinking_tokens": 16384,
+      "supports_images": true,
+      "intelligence_score": 16,
+      "temperature_constraint": "range",
+      "quota_pool": "gemini-cli"
+    },
+    {
+      "model_name": "gpt-oss-120b-medium",
+      "friendly_name": "GPT-OSS 120B Medium (Antigravity)",
+      "aliases": ["ag-gpt-oss"],
+      "context_window": 128000,
+      "max_output_tokens": 16384,
+      "supports_extended_thinking": false,
+      "supports_images": false,
+      "intelligence_score": 15,
+      "temperature_constraint": "range"
+    }
+  ]
+}
--- a/docs/antigravity_provider.md
+++ b/docs/antigravity_provider.md
@@ -0,0 +1,189 @@
+# Antigravity Provider
+
+The Antigravity provider enables access to Claude, Gemini, and other models through Google's unified gateway API. This provides an alternative way to access high-quality models using Google OAuth2 authentication instead of individual API keys.
+
+## Important Warning
+
+> **Terms of Service Risk**: Using the Antigravity provider may violate Google's Terms of Service. Users have reported account bans when using this approach.
+>
+> **High-Risk Scenarios**:
+> - Fresh Google accounts have a very high chance of getting banned
+> - New accounts with Pro/Ultra subscriptions are frequently flagged
+>
+> **By using this provider, you acknowledge and accept all associated risks.**
+
+## Available Models
+
+### Antigravity Quota Pool
+
+These models use the Antigravity unified gateway quota:
+
+| Model ID | Description | Thinking Support |
+|----------|-------------|------------------|
+| `claude-opus-4-5-thinking` | Claude Opus 4.5 with extended thinking | Yes (up to 32K tokens) |
+| `claude-sonnet-4-5-thinking` | Claude Sonnet 4.5 with extended thinking | Yes (up to 32K tokens) |
+| `claude-sonnet-4-5` | Claude Sonnet 4.5 standard | No |
+| `gemini-3-pro-high` | Gemini 3 Pro with high thinking budget | Yes (up to 65K tokens) |
+| `gemini-3-pro-low` | Gemini 3 Pro with low thinking budget | Yes (up to 16K tokens) |
+| `gemini-3-flash` | Gemini 3 Flash | Yes (up to 32K tokens) |
+| `gpt-oss-120b-medium` | GPT-OSS 120B | No |
+
+### Gemini CLI Quota Pool
+
+These models use a separate Gemini CLI quota:
+
+| Model ID | Description | Thinking Support |
+|----------|-------------|------------------|
+| `gemini-3-flash-preview` | Gemini 3 Flash Preview | Yes |
+| `gemini-3-pro-preview` | Gemini 3 Pro Preview | Yes |
+| `gemini-2.5-pro` | Gemini 2.5 Pro | Yes |
+| `gemini-2.5-flash` | Gemini 2.5 Flash | Yes |
+
+## Setup
+
+### Option 1: OpenCode Integration (Recommended)
+
+If you use OpenCode with the `opencode-antigravity-auth` plugin, your accounts are automatically detected from:
+
+```
+~/.config/opencode/antigravity-accounts.json
+```
+
+No additional configuration is needed. The PAL MCP server will automatically discover and use these accounts.
+
+### Option 2: Manual Refresh Token
+
+Set the refresh token directly via environment variable:
+
+```bash
+export ANTIGRAVITY_REFRESH_TOKEN="your_oauth_refresh_token"
+export ANTIGRAVITY_PROJECT_ID="your_gcp_project_id"  # Optional
+```
+
+### Option 3: Custom Accounts File
+
+Create `~/.config/opencode/antigravity-accounts.json` manually:
+
+```json
+{
+  "version": 3,
+  "accounts": [
+    {
+      "email": "your.email@gmail.com",
+      "refreshToken": "1//your_refresh_token_here",
+      "projectId": "your-gcp-project-id",
+      "enabled": true
+    }
+  ],
+  "activeIndex": 0
+}
+```
+
+## Environment Variables
+
+| Variable | Required | Description |
+|----------|----------|-------------|
+| `ANTIGRAVITY_REFRESH_TOKEN` | No* | OAuth2 refresh token (overrides accounts file) |
+| `ANTIGRAVITY_PROJECT_ID` | No | Google Cloud project ID (defaults to `rising-fact-p41fc`) |
+| `ANTIGRAVITY_BASE_URL` | No | API endpoint (defaults to production `cloudcode-pa.googleapis.com`) |
+
+*Either `ANTIGRAVITY_REFRESH_TOKEN` or a valid accounts file at `~/.config/opencode/antigravity-accounts.json` is required.
+
+## Features
+
+### Multi-Account Rotation
+
+When multiple accounts are configured, the provider automatically:
+- Rotates between accounts when one is rate-limited
+- Tracks per-model rate limits separately
+- Preserves account preferences for cache efficiency
+
+### Extended Thinking
+
+For models that support thinking (Claude and Gemini 3), you can control the thinking budget:
+
+```python
+# Thinking modes: minimal, low, medium, high, max
+response = provider.generate_content(
+    prompt="Explain quantum computing",
+    model_name="claude-opus-4-5-thinking",
+    thinking_mode="high"  # Uses 67% of max thinking budget
+)
+```
+
+| Mode | Budget Percentage |
+|------|-------------------|
+| minimal | 0.5% |
+| low | 8% |
+| medium | 33% (default) |
+| high | 67% |
+| max | 100% |
+
+### Image Support
+
+Models with vision capability support image inputs:
+
+```python
+response = provider.generate_content(
+    prompt="What's in this image?",
+    model_name="gemini-3-flash",
+    images=["/path/to/image.png"]
+)
+```
+
+## Troubleshooting
+
+### "No Antigravity accounts configured"
+
+Ensure either:
+1. `ANTIGRAVITY_REFRESH_TOKEN` environment variable is set, or
+2. `~/.config/opencode/antigravity-accounts.json` exists with valid accounts
+
+### "All Antigravity accounts are rate limited"
+
+The provider has exhausted all available accounts. Either:
+1. Wait for rate limits to reset (usually 60 seconds to a few minutes)
+2. Add more accounts to the accounts file
+
+### "Token refresh failed"
+
+Your refresh token may have expired. Re-authenticate:
+1. If using OpenCode: `opencode auth login`
+2. If manual: Obtain a new refresh token via OAuth2 flow
+
+### 403 Permission Denied
+
+Your Google Cloud project may not have the required APIs enabled:
+1. Go to [Google Cloud Console](https://console.cloud.google.com/)
+2. Enable the **Gemini for Google Cloud API** (`cloudaicompanion.googleapis.com`)
+
+## API Format
+
+The Antigravity API uses Gemini-style request format internally:
+
+```json
+{
+  "project": "your-project-id",
+  "model": "claude-opus-4-5-thinking",
+  "request": {
+    "contents": [
+      {"role": "user", "parts": [{"text": "Your prompt"}]}
+    ],
+    "generationConfig": {
+      "temperature": 0.7,
+      "maxOutputTokens": 8192,
+      "thinkingConfig": {
+        "thinkingBudget": 8000,
+        "includeThoughts": true
+      }
+    }
+  }
+}
+```
+
+This transformation is handled automatically by the provider.
+
+## Related Resources
+
+- [opencode-antigravity-auth](https://github.com/NoeFabris/opencode-antigravity-auth) - OpenCode plugin for Antigravity authentication
+- [Antigravity API Specification](https://github.com/NoeFabris/opencode-antigravity-auth/blob/dev/docs/ANTIGRAVITY_API_SPEC.md) - Detailed API documentation
--- a/providers/init.py
+++ b/providers/init.py
@@ -1,5 +1,6 @@
 """Model provider abstractions for supporting multiple AI providers."""

+from .antigravity import AntigravityProvider
 from .azure_openai import AzureOpenAIProvider
 from .base import ModelProvider
 from .gemini import GeminiModelProvider
@@ -15,6 +16,7 @@ __all__ = [
    "ModelResponse",
    "ModelCapabilities",
    "ModelProviderRegistry",
+    "AntigravityProvider",
    "AzureOpenAIProvider",
    "GeminiModelProvider",
    "OpenAIModelProvider",
--- a/providers/antigravity.py
+++ b/providers/antigravity.py
@@ -0,0 +1,576 @@
+"""Antigravity model provider implementation.
+
+Antigravity is Google's unified gateway API for accessing multiple AI models
+(Claude, Gemini, GPT-OSS) through a single Gemini-style interface.
+
+WARNING: Using this provider may violate Google's Terms of Service.
+See docs/antigravity_provider.md for important information about risks.
+"""
+
+from __future__ import annotations
+
+import base64
+import json
+import logging
+import re
+import uuid
+from typing import TYPE_CHECKING, Any, ClassVar
+
+import httpx
+
+from utils.env import get_env
+from utils.image_utils import validate_image
+
+from .antigravity_auth import AntigravityTokenManager
+from .base import ModelProvider
+from .registries.antigravity import AntigravityModelRegistry
+from .shared import ModelCapabilities, ModelResponse, ProviderType
+
+if TYPE_CHECKING:
+    from tools.models import ToolModelCategory
+
+logger = logging.getLogger(__name__)
+
+# Antigravity API endpoints
+PRODUCTION_ENDPOINT = "https://cloudcode-pa.googleapis.com"
+DAILY_ENDPOINT = "https://daily-cloudcode-pa.sandbox.googleapis.com"
+
+# API path for content generation
+GENERATE_CONTENT_PATH = "/v1internal:generateContent"
+STREAM_GENERATE_CONTENT_PATH = "/v1internal:streamGenerateContent"
+
+# ToS warning message
+TOS_WARNING = """
+================================================================================
+ANTIGRAVITY PROVIDER WARNING
+
+Using the Antigravity provider may violate Google's Terms of Service.
+Users have reported account bans when using this approach.
+
+HIGH-RISK SCENARIOS:
+- Fresh Google accounts have a very high chance of getting banned
+- New accounts with Pro/Ultra subscriptions are frequently flagged
+
+By using this provider, you acknowledge and accept all associated risks.
+See docs/antigravity_provider.md for more information.
+================================================================================
+"""
+
+
+class AntigravityProvider(ModelProvider):
+    """Provider for Google's Antigravity unified gateway API.
+
+    Antigravity provides access to Claude, Gemini, and other models through a
+    single Gemini-style API interface. This provider handles:
+
+    * OAuth2 authentication with refresh token management
+    * Multi-account rotation on rate limits
+    * Request transformation to Gemini-style format
+    * Extended thinking support for compatible models
+    * Image processing for vision requests
+
+    Configuration:
+        ANTIGRAVITY_REFRESH_TOKEN: OAuth2 refresh token (env var override)
+        ANTIGRAVITY_PROJECT_ID: Google Cloud project ID
+        ANTIGRAVITY_BASE_URL: Custom endpoint URL (defaults to production)
+    """
+
+    # Model registry for capability metadata
+    _registry: ClassVar[AntigravityModelRegistry | None] = None
+    MODEL_CAPABILITIES: ClassVar[dict[str, ModelCapabilities]] = {}
+
+    # Thinking mode budgets (percentage of max_thinking_tokens)
+    THINKING_BUDGETS = {
+        "minimal": 0.005,
+        "low": 0.08,
+        "medium": 0.33,
+        "high": 0.67,
+        "max": 1.0,
+    }
+
+    # Track whether warning has been shown this session
+    _warning_shown: ClassVar[bool] = False
+
+    def __init__(self, api_key: str = "", **kwargs) -> None:
+        """Initialize Antigravity provider.
+
+        Args:
+            api_key: Not used (authentication via OAuth2 refresh tokens)
+            **kwargs: Additional configuration options
+        """
+        # Initialize registry
+        if AntigravityProvider._registry is None:
+            AntigravityProvider._registry = AntigravityModelRegistry()
+            AntigravityProvider.MODEL_CAPABILITIES = dict(self._registry.model_map)
+            logger.info("Antigravity provider loaded %d models", len(self._registry.model_map))
+
+        super().__init__(api_key, **kwargs)
+
+        # Initialize token manager
+        self._token_manager = AntigravityTokenManager()
+
+        # HTTP client for API requests
+        self._http_client: httpx.Client | None = None
+
+        # Base URL configuration
+        self._base_url = get_env("ANTIGRAVITY_BASE_URL") or PRODUCTION_ENDPOINT
+
+        # Show warning on first use
+        if not AntigravityProvider._warning_shown:
+            logger.warning(TOS_WARNING)
+            AntigravityProvider._warning_shown = True
+
+        self._invalidate_capability_cache()
+
+    @property
+    def http_client(self) -> httpx.Client:
+        """Lazy initialization of HTTP client."""
+        if self._http_client is None:
+            timeout = httpx.Timeout(
+                connect=30.0,
+                read=600.0,  # 10 minutes for long responses
+                write=60.0,
+                pool=600.0,
+            )
+            self._http_client = httpx.Client(timeout=timeout, follow_redirects=True)
+        return self._http_client
+
+    # ------------------------------------------------------------------
+    # Provider identity
+    # ------------------------------------------------------------------
+
+    def get_provider_type(self) -> ProviderType:
+        """Return the provider type."""
+        return ProviderType.ANTIGRAVITY
+
+    # ------------------------------------------------------------------
+    # Capability surface
+    # ------------------------------------------------------------------
+
+    def _lookup_capabilities(
+        self,
+        canonical_name: str,
+        requested_name: str | None = None,
+    ) -> ModelCapabilities | None:
+        """Fetch capabilities from the registry."""
+        if self._registry:
+            return self._registry.get_capabilities(canonical_name)
+        return None
+
+    def get_all_model_capabilities(self) -> dict[str, ModelCapabilities]:
+        """Return registry-backed capabilities."""
+        if not self._registry:
+            return {}
+        return dict(self._registry.model_map)
+
+    def list_models(
+        self,
+        *,
+        respect_restrictions: bool = True,
+        include_aliases: bool = True,
+        lowercase: bool = False,
+        unique: bool = False,
+    ) -> list[str]:
+        """Return available Antigravity model names."""
+        if not self._registry:
+            return []
+
+        from utils.model_restrictions import get_restriction_service
+
+        restriction_service = get_restriction_service() if respect_restrictions else None
+        allowed_configs: dict[str, ModelCapabilities] = {}
+
+        for model_name in self._registry.list_models():
+            config = self._registry.resolve(model_name)
+            if not config:
+                continue
+
+            if restriction_service:
+                if not restriction_service.is_allowed(self.get_provider_type(), model_name):
+                    continue
+
+            allowed_configs[model_name] = config
+
+        if not allowed_configs:
+            return []
+
+        return ModelCapabilities.collect_model_names(
+            allowed_configs,
+            include_aliases=include_aliases,
+            lowercase=lowercase,
+            unique=unique,
+        )
+
+    def _resolve_model_name(self, model_name: str) -> str:
+        """Resolve aliases to canonical model names."""
+        if self._registry:
+            config = self._registry.resolve(model_name)
+            if config and config.model_name != model_name:
+                logger.debug("Resolved Antigravity alias '%s' to '%s'", model_name, config.model_name)
+                return config.model_name
+        return model_name
+
+    # ------------------------------------------------------------------
+    # Content generation
+    # ------------------------------------------------------------------
+
+    def generate_content(
+        self,
+        prompt: str,
+        model_name: str,
+        system_prompt: str | None = None,
+        temperature: float = 0.7,
+        max_output_tokens: int | None = None,
+        thinking_mode: str = "medium",
+        images: list[str] | None = None,
+        **kwargs,
+    ) -> ModelResponse:
+        """Generate content using the Antigravity API.
+
+        Args:
+            prompt: User prompt to send to the model
+            model_name: Model identifier (e.g., 'claude-opus-4-5-thinking')
+            system_prompt: Optional system instructions
+            temperature: Sampling temperature (0.0-2.0)
+            max_output_tokens: Maximum tokens in response
+            thinking_mode: Thinking budget level for thinking models
+            images: Optional list of image paths or data URLs
+            **kwargs: Additional parameters
+
+        Returns:
+            ModelResponse with generated content and metadata
+        """
+        # Validate and resolve model name
+        self.validate_parameters(model_name, temperature)
+        capabilities = self.get_capabilities(model_name)
+        resolved_model = self._resolve_model_name(model_name)
+
+        # Get authentication
+        access_token, project_id, fingerprint_headers = self._token_manager.get_access_token(model=resolved_model)
+
+        # Build request
+        request_body = self._build_request(
+            prompt=prompt,
+            model_name=resolved_model,
+            project_id=project_id,
+            system_prompt=system_prompt,
+            temperature=temperature,
+            max_output_tokens=max_output_tokens,
+            thinking_mode=thinking_mode,
+            images=images,
+            capabilities=capabilities,
+        )
+
+        # Build headers
+        headers = {
+            "Authorization": f"Bearer {access_token}",
+            "Content-Type": "application/json",
+            **fingerprint_headers,
+        }
+
+        # Retry logic
+        max_retries = 4
+        retry_delays = [1.0, 3.0, 5.0, 8.0]
+        attempt_counter = {"value": 0}
+
+        def _attempt() -> ModelResponse:
+            attempt_counter["value"] += 1
+
+            url = f"{self._base_url}{GENERATE_CONTENT_PATH}"
+            logger.debug("Antigravity request to %s for model %s", url, resolved_model)
+
+            response = self.http_client.post(
+                url,
+                json=request_body,
+                headers=headers,
+            )
+
+            # Handle errors
+            if response.status_code != 200:
+                self._handle_error_response(response, resolved_model)
+
+            return self._parse_response(response.json(), resolved_model, capabilities, thinking_mode)
+
+        try:
+            return self._run_with_retries(
+                operation=_attempt,
+                max_attempts=max_retries,
+                delays=retry_delays,
+                log_prefix=f"Antigravity API ({resolved_model})",
+            )
+        except Exception as exc:
+            attempts = max(attempt_counter["value"], 1)
+            error_msg = f"Antigravity API error for {resolved_model} after {attempts} attempt(s): {exc}"
+            raise RuntimeError(error_msg) from exc
+
+    def _build_request(
+        self,
+        prompt: str,
+        model_name: str,
+        project_id: str,
+        system_prompt: str | None = None,
+        temperature: float = 0.7,
+        max_output_tokens: int | None = None,
+        thinking_mode: str = "medium",
+        images: list[str] | None = None,
+        capabilities: ModelCapabilities | None = None,
+    ) -> dict[str, Any]:
+        """Build the Antigravity API request body."""
+        # Build contents array (Gemini-style format)
+        contents = []
+
+        # Build user message parts
+        user_parts = []
+
+        # Add text prompt
+        user_parts.append({"text": prompt})
+
+        # Add images if provided
+        if images and capabilities and capabilities.supports_images:
+            for image_path in images:
+                try:
+                    image_part = self._process_image(image_path)
+                    if image_part:
+                        user_parts.append(image_part)
+                except Exception as e:
+                    logger.warning("Failed to process image %s: %s", image_path, e)
+        elif images:
+            logger.warning("Model %s does not support images, ignoring %d image(s)", model_name, len(images))
+
+        contents.append({"role": "user", "parts": user_parts})
+
+        # Build generation config
+        generation_config: dict[str, Any] = {
+            "temperature": temperature,
+        }
+
+        if max_output_tokens:
+            generation_config["maxOutputTokens"] = max_output_tokens
+
+        # Add thinking config for thinking-capable models
+        if capabilities and capabilities.supports_extended_thinking and thinking_mode in self.THINKING_BUDGETS:
+            max_thinking = capabilities.max_thinking_tokens or 8192
+            thinking_budget = int(max_thinking * self.THINKING_BUDGETS[thinking_mode])
+            generation_config["thinkingConfig"] = {
+                "thinkingBudget": thinking_budget,
+                "includeThoughts": True,
+            }
+            # Ensure maxOutputTokens > thinkingBudget
+            if not max_output_tokens or max_output_tokens <= thinking_budget:
+                generation_config["maxOutputTokens"] = thinking_budget + 10000
+
+        # Build system instruction (must be object with parts, not string)
+        system_instruction = None
+        if system_prompt:
+            system_instruction = {"parts": [{"text": system_prompt}]}
+
+        # Build full request
+        request: dict[str, Any] = {
+            "project": project_id,
+            "model": model_name,
+            "request": {
+                "contents": contents,
+                "generationConfig": generation_config,
+            },
+            "userAgent": "antigravity",
+            "requestId": f"pal-{uuid.uuid4().hex[:12]}",
+        }
+
+        if system_instruction:
+            request["request"]["systemInstruction"] = system_instruction
+
+        return request
+
+    def _parse_response(
+        self,
+        data: dict[str, Any],
+        model_name: str,
+        capabilities: ModelCapabilities | None,
+        thinking_mode: str,
+    ) -> ModelResponse:
+        """Parse the Antigravity API response."""
+        response_data = data.get("response", data)
+
+        # Extract content from candidates
+        content = ""
+        thinking_content = ""
+        candidates = response_data.get("candidates", [])
+
+        if candidates:
+            candidate = candidates[0]
+            candidate_content = candidate.get("content", {})
+            parts = candidate_content.get("parts", [])
+
+            for part in parts:
+                if part.get("thought"):
+                    # This is thinking content
+                    thinking_content += part.get("text", "")
+                elif "thoughtSignature" in part and "text" in part:
+                    # Gemini thinking with signature
+                    thinking_content += part.get("text", "")
+                elif "text" in part:
+                    content += part.get("text", "")
+
+        # Extract usage metadata
+        usage_metadata = response_data.get("usageMetadata", {})
+        usage = {
+            "input_tokens": usage_metadata.get("promptTokenCount", 0),
+            "output_tokens": usage_metadata.get("candidatesTokenCount", 0),
+            "total_tokens": usage_metadata.get("totalTokenCount", 0),
+        }
+
+        # Add thinking tokens if available
+        if "thoughtsTokenCount" in usage_metadata:
+            usage["thinking_tokens"] = usage_metadata["thoughtsTokenCount"]
+
+        # Extract finish reason
+        finish_reason = "UNKNOWN"
+        if candidates:
+            finish_reason = candidates[0].get("finishReason", "STOP")
+
+        return ModelResponse(
+            content=content,
+            usage=usage,
+            model_name=model_name,
+            friendly_name="Antigravity",
+            provider=ProviderType.ANTIGRAVITY,
+            metadata={
+                "thinking_mode": thinking_mode if capabilities and capabilities.supports_extended_thinking else None,
+                "thinking_content": thinking_content if thinking_content else None,
+                "finish_reason": finish_reason,
+                "model_version": response_data.get("modelVersion"),
+                "response_id": response_data.get("responseId"),
+                "trace_id": data.get("traceId"),
+            },
+        )
+
+    def _handle_error_response(self, response: httpx.Response, model_name: str) -> None:
+        """Handle error responses from the API."""
+        try:
+            error_data = response.json()
+            error = error_data.get("error", {})
+            error_code = error.get("code", response.status_code)
+            error_message = error.get("message", response.text)
+            error_status = error.get("status", "UNKNOWN")
+
+            # Handle rate limiting
+            if error_code == 429:
+                # Extract retry delay from response
+                retry_delay = 60.0
+                details = error.get("details", [])
+                for detail in details:
+                    if detail.get("@type", "").endswith("RetryInfo"):
+                        delay_str = detail.get("retryDelay", "60s")
+                        # Parse delay string (e.g., "3.957525076s")
+                        match = re.match(r"([\d.]+)s?", delay_str)
+                        if match:
+                            retry_delay = float(match.group(1))
+                        break
+
+                # Mark current account as rate limited
+                self._token_manager.mark_rate_limited(model_name, retry_delay)
+
+                raise RuntimeError(f"Rate limited (429): {error_message}. Retry after {retry_delay}s")
+
+            # Re-raise with formatted message
+            raise RuntimeError(f"Antigravity API error {error_code} ({error_status}): {error_message}")
+
+        except json.JSONDecodeError:
+            raise RuntimeError(f"Antigravity API error {response.status_code}: {response.text}")
+
+    def _process_image(self, image_path: str) -> dict[str, Any] | None:
+        """Process an image for the Antigravity API."""
+        try:
+            image_bytes, mime_type = validate_image(image_path)
+
+            if image_path.startswith("data:"):
+                # Extract base64 data from data URL
+                _, data = image_path.split(",", 1)
+                return {"inlineData": {"mimeType": mime_type, "data": data}}
+            else:
+                # Encode file bytes
+                image_data = base64.b64encode(image_bytes).decode()
+                return {"inlineData": {"mimeType": mime_type, "data": image_data}}
+
+        except ValueError as e:
+            logger.warning("Image validation failed: %s", e)
+            return None
+        except Exception as e:
+            logger.error("Error processing image %s: %s", image_path, e)
+            return None
+
+    def _is_error_retryable(self, error: Exception) -> bool:
+        """Determine if an error should be retried."""
+        error_str = str(error).lower()
+
+        # Rate limits should be retried (we'll rotate accounts)
+        if "429" in error_str or "rate limit" in error_str:
+            return True
+
+        # Check for retryable indicators
+        retryable_indicators = [
+            "timeout",
+            "connection",
+            "network",
+            "temporary",
+            "unavailable",
+            "retry",
+            "500",
+            "502",
+            "503",
+            "504",
+        ]
+
+        return any(indicator in error_str for indicator in retryable_indicators)
+
+    # ------------------------------------------------------------------
+    # Preferred model selection
+    # ------------------------------------------------------------------
+
+    def get_preferred_model(self, category: ToolModelCategory, allowed_models: list[str]) -> str | None:
+        """Get preferred model for a category from allowed models."""
+        from tools.models import ToolModelCategory
+
+        if not allowed_models:
+            return None
+
+        capability_map = self.get_all_model_capabilities()
+
+        def find_best(candidates: list[str]) -> str | None:
+            return sorted(candidates, reverse=True)[0] if candidates else None
+
+        if category == ToolModelCategory.EXTENDED_REASONING:
+            # Prefer thinking-capable models
+            thinking_models = [
+                m for m in allowed_models if m in capability_map and capability_map[m].supports_extended_thinking
+            ]
+            if thinking_models:
+                # Prefer Claude Opus, then Claude Sonnet, then Gemini
+                opus = [m for m in thinking_models if "opus" in m]
+                if opus:
+                    return find_best(opus)
+                sonnet = [m for m in thinking_models if "sonnet" in m]
+                if sonnet:
+                    return find_best(sonnet)
+                return find_best(thinking_models)
+
+        elif category == ToolModelCategory.FAST_RESPONSE:
+            # Prefer Flash models
+            flash_models = [m for m in allowed_models if "flash" in m]
+            if flash_models:
+                return find_best(flash_models)
+
+        # Default: prefer models by capability rank
+        return find_best(allowed_models)
+
+    # ------------------------------------------------------------------
+    # Resource cleanup
+    # ------------------------------------------------------------------
+
+    def close(self) -> None:
+        """Clean up resources."""
+        if self._http_client:
+            self._http_client.close()
+            self._http_client = None
+        if self._token_manager:
+            self._token_manager.close()
--- a/providers/antigravity_auth.py
+++ b/providers/antigravity_auth.py
@@ -0,0 +1,299 @@
+"""OAuth2 token manager for Antigravity provider.
+
+This module handles authentication with Google's Antigravity unified gateway API,
+including refresh token management and multi-account rotation.
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import time
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any
+
+import httpx
+
+from utils.env import get_env
+
+logger = logging.getLogger(__name__)
+
+# OAuth2 endpoints
+TOKEN_URL = "https://oauth2.googleapis.com/token"
+
+# Google OAuth2 client credentials (from Antigravity/Gemini CLI)
+# These are publicly documented OAuth client credentials
+OAUTH_CLIENT_ID = "764086051850-6qr4p6gpi6hn506pt8ejuq83di341hur.apps.googleusercontent.com"
+OAUTH_CLIENT_SECRET = "d-FL95Q19q7MQmFpd7hHD0Ty"
+
+# Required OAuth2 scopes for Antigravity API access
+OAUTH_SCOPES = [
+    "https://www.googleapis.com/auth/cloud-platform",
+    "https://www.googleapis.com/auth/userinfo.email",
+    "https://www.googleapis.com/auth/userinfo.profile",
+    "https://www.googleapis.com/auth/cclog",
+    "https://www.googleapis.com/auth/experimentsandconfigs",
+]
+
+# Default accounts file location
+DEFAULT_ACCOUNTS_FILE = Path.home() / ".config" / "opencode" / "antigravity-accounts.json"
+
+
+@dataclass
+class AntigravityAccount:
+    """Represents a single Antigravity OAuth account."""
+
+    email: str
+    refresh_token: str
+    project_id: str
+    enabled: bool = True
+    access_token: str | None = None
+    token_expiry: float = 0.0
+    rate_limit_reset_times: dict[str, float] = field(default_factory=dict)
+    fingerprint: dict[str, Any] = field(default_factory=dict)
+
+    def is_token_valid(self) -> bool:
+        """Check if the current access token is still valid."""
+        if not self.access_token:
+            return False
+        # Refresh 60 seconds before expiry
+        return time.time() < (self.token_expiry - 60)
+
+    def is_rate_limited(self, model: str | None = None) -> bool:
+        """Check if this account is rate limited for a specific model or globally."""
+        now = time.time() * 1000  # Convert to milliseconds
+        if model:
+            reset_time = self.rate_limit_reset_times.get(model, 0)
+            if reset_time > now:
+                return True
+        # Check for any active rate limits
+        return any(reset > now for reset in self.rate_limit_reset_times.values())
+
+    def set_rate_limited(self, model: str, reset_delay_seconds: float) -> None:
+        """Mark this account as rate limited for a model."""
+        reset_time = (time.time() + reset_delay_seconds) * 1000
+        self.rate_limit_reset_times[model] = reset_time
+        logger.info("Account %s rate limited for %s until %.0f", self.email, model, reset_time)
+
+
+class AntigravityTokenManager:
+    """Manages OAuth2 tokens for Antigravity API access.
+
+    Supports:
+    - Loading accounts from ~/.config/opencode/antigravity-accounts.json
+    - Environment variable override via ANTIGRAVITY_REFRESH_TOKEN
+    - Token refresh with automatic expiry handling
+    - Multi-account rotation on rate limits
+    """
+
+    def __init__(self, accounts_file: Path | str | None = None) -> None:
+        self._accounts: list[AntigravityAccount] = []
+        self._current_account_index: int = 0
+        self._http_client: httpx.Client | None = None
+
+        # Try environment variable first
+        env_token = get_env("ANTIGRAVITY_REFRESH_TOKEN")
+        env_project = get_env("ANTIGRAVITY_PROJECT_ID")
+
+        if env_token:
+            logger.info("Using Antigravity refresh token from environment variable")
+            self._accounts.append(
+                AntigravityAccount(
+                    email="env-configured",
+                    refresh_token=env_token,
+                    project_id=env_project or "rising-fact-p41fc",  # Default project
+                )
+            )
+        else:
+            # Load from accounts file
+            accounts_path = Path(accounts_file) if accounts_file else DEFAULT_ACCOUNTS_FILE
+            self._load_accounts(accounts_path)
+
+        if self._accounts:
+            logger.info("Antigravity token manager initialized with %d account(s)", len(self._accounts))
+        else:
+            logger.warning("No Antigravity accounts configured")
+
+    @property
+    def http_client(self) -> httpx.Client:
+        """Lazy-init HTTP client."""
+        if self._http_client is None:
+            self._http_client = httpx.Client(timeout=30.0)
+        return self._http_client
+
+    def _load_accounts(self, accounts_path: Path) -> None:
+        """Load accounts from the antigravity-accounts.json file."""
+        if not accounts_path.exists():
+            logger.debug("Antigravity accounts file not found at %s", accounts_path)
+            return
+
+        try:
+            with open(accounts_path, encoding="utf-8") as f:
+                data = json.load(f)
+
+            version = data.get("version", 1)
+            if version < 3:
+                logger.warning(
+                    "Antigravity accounts file has old format (version %d), may have limited support", version
+                )
+
+            for account_data in data.get("accounts", []):
+                if not account_data.get("enabled", True):
+                    continue
+
+                email = account_data.get("email", "unknown")
+                refresh_token = account_data.get("refreshToken")
+                project_id = account_data.get("projectId") or account_data.get("managedProjectId")
+
+                if not refresh_token:
+                    logger.warning("Skipping account %s: no refresh token", email)
+                    continue
+
+                account = AntigravityAccount(
+                    email=email,
+                    refresh_token=refresh_token,
+                    project_id=project_id or "rising-fact-p41fc",
+                    enabled=account_data.get("enabled", True),
+                    rate_limit_reset_times=account_data.get("rateLimitResetTimes", {}),
+                    fingerprint=account_data.get("fingerprint", {}),
+                )
+                self._accounts.append(account)
+                logger.debug("Loaded Antigravity account: %s (project: %s)", email, project_id)
+
+        except json.JSONDecodeError as e:
+            logger.error("Failed to parse Antigravity accounts file: %s", e)
+        except Exception as e:
+            logger.error("Error loading Antigravity accounts: %s", e)
+
+    def has_accounts(self) -> bool:
+        """Check if any accounts are configured."""
+        return len(self._accounts) > 0
+
+    def get_account_count(self) -> int:
+        """Return the number of configured accounts."""
+        return len(self._accounts)
+
+    def get_access_token(self, model: str | None = None) -> tuple[str, str, dict[str, str]]:
+        """Get a valid access token, refreshing if necessary.
+
+        Args:
+            model: Optional model name for rate limit checking
+
+        Returns:
+            Tuple of (access_token, project_id, fingerprint_headers)
+
+        Raises:
+            RuntimeError: If no valid token can be obtained
+        """
+        if not self._accounts:
+            raise RuntimeError("No Antigravity accounts configured")
+
+        # Try each account starting from current index
+        attempts = 0
+        original_index = self._current_account_index
+
+        while attempts < len(self._accounts):
+            account = self._accounts[self._current_account_index]
+
+            # Skip rate-limited accounts for this model
+            if model and account.is_rate_limited(model):
+                logger.debug("Skipping rate-limited account %s for model %s", account.email, model)
+                self._rotate_account()
+                attempts += 1
+                continue
+
+            # Refresh token if needed
+            if not account.is_token_valid():
+                try:
+                    self._refresh_token(account)
+                except Exception as e:
+                    logger.warning("Failed to refresh token for %s: %s", account.email, e)
+                    self._rotate_account()
+                    attempts += 1
+                    continue
+
+            # Build fingerprint headers
+            fingerprint_headers = self._build_fingerprint_headers(account)
+
+            return account.access_token, account.project_id, fingerprint_headers
+
+        # All accounts exhausted
+        self._current_account_index = original_index
+        raise RuntimeError("All Antigravity accounts are rate limited or have invalid tokens")
+
+    def mark_rate_limited(self, model: str, retry_delay: float = 60.0) -> None:
+        """Mark the current account as rate limited and rotate to next."""
+        if self._accounts:
+            account = self._accounts[self._current_account_index]
+            account.set_rate_limited(model, retry_delay)
+            self._rotate_account()
+
+    def _rotate_account(self) -> None:
+        """Rotate to the next available account."""
+        if len(self._accounts) > 1:
+            self._current_account_index = (self._current_account_index + 1) % len(self._accounts)
+            logger.debug("Rotated to account index %d", self._current_account_index)
+
+    def _refresh_token(self, account: AntigravityAccount) -> None:
+        """Refresh the OAuth2 access token for an account."""
+        logger.debug("Refreshing access token for %s", account.email)
+
+        response = self.http_client.post(
+            TOKEN_URL,
+            data={
+                "client_id": OAUTH_CLIENT_ID,
+                "client_secret": OAUTH_CLIENT_SECRET,
+                "refresh_token": account.refresh_token,
+                "grant_type": "refresh_token",
+            },
+        )
+
+        if response.status_code != 200:
+            error_msg = f"Token refresh failed: {response.status_code} - {response.text}"
+            logger.error(error_msg)
+            raise RuntimeError(error_msg)
+
+        data = response.json()
+        account.access_token = data["access_token"]
+        # Token typically expires in 3600 seconds
+        expires_in = data.get("expires_in", 3600)
+        account.token_expiry = time.time() + expires_in
+
+        logger.debug("Token refreshed for %s, expires in %d seconds", account.email, expires_in)
+
+    def _build_fingerprint_headers(self, account: AntigravityAccount) -> dict[str, str]:
+        """Build fingerprint headers for API requests."""
+        fingerprint = account.fingerprint
+        headers = {}
+
+        if fingerprint.get("userAgent"):
+            headers["User-Agent"] = fingerprint["userAgent"]
+        else:
+            headers["User-Agent"] = "antigravity/1.15.8 linux/amd64"
+
+        if fingerprint.get("apiClient"):
+            headers["X-Goog-Api-Client"] = fingerprint["apiClient"]
+        else:
+            headers["X-Goog-Api-Client"] = "google-cloud-sdk intellij/2024.1"
+
+        # Build client metadata
+        client_metadata = fingerprint.get("clientMetadata", {})
+        metadata = {
+            "ideType": client_metadata.get("ideType", "CLOUD_SHELL_EDITOR"),
+            "platform": client_metadata.get("platform", "LINUX"),
+            "pluginType": client_metadata.get("pluginType", "GEMINI"),
+        }
+        headers["Client-Metadata"] = json.dumps(metadata)
+
+        # Quota user for rate limiting
+        if fingerprint.get("quotaUser"):
+            headers["X-Goog-Quota-User"] = fingerprint["quotaUser"]
+
+        return headers
+
+    def close(self) -> None:
+        """Close the HTTP client."""
+        if self._http_client:
+            self._http_client.close()
+            self._http_client = None
--- a/providers/registries/antigravity.py
+++ b/providers/registries/antigravity.py
@@ -0,0 +1,56 @@
+"""Antigravity model registry for managing model configurations and aliases."""
+
+from __future__ import annotations
+
+from ..shared import ModelCapabilities, ProviderType
+from .base import CAPABILITY_FIELD_NAMES, CapabilityModelRegistry
+
+
+class AntigravityModelRegistry(CapabilityModelRegistry):
+    """Capability registry backed by ``conf/antigravity_models.json``.
+
+    The Antigravity provider accesses Claude, Gemini, and other models through
+    Google's unified gateway API. Models are split across two quota pools:
+
+    * **antigravity**: Claude models and Gemini 3 variants
+    * **gemini-cli**: Gemini 2.5 and preview models using Gemini CLI quota
+    """
+
+    # Extra fields beyond standard ModelCapabilities
+    EXTRA_KEYS = {"quota_pool"}
+
+    def __init__(self, config_path: str | None = None) -> None:
+        super().__init__(
+            env_var_name="ANTIGRAVITY_MODELS_CONFIG_PATH",
+            default_filename="antigravity_models.json",
+            provider=ProviderType.ANTIGRAVITY,
+            friendly_prefix="Antigravity ({model})",
+            config_path=config_path,
+        )
+
+    def _extra_keys(self) -> set[str]:
+        """Include quota_pool as a valid entry key."""
+        return self.EXTRA_KEYS
+
+    def _finalise_entry(self, entry: dict) -> tuple[ModelCapabilities, dict]:
+        """Build ModelCapabilities and extract extras like quota_pool."""
+        # Extract extras before building capabilities
+        extras = {}
+        quota_pool = entry.pop("quota_pool", None)
+        if quota_pool:
+            extras["quota_pool"] = quota_pool
+
+        # Build capabilities from remaining fields
+        filtered = {k: v for k, v in entry.items() if k in CAPABILITY_FIELD_NAMES}
+        filtered.setdefault("provider", ProviderType.ANTIGRAVITY)
+        filtered.setdefault("friendly_name", f"Antigravity ({entry.get('model_name', 'unknown')})")
+        capability = ModelCapabilities(**filtered)
+
+        return capability, extras
+
+    def get_quota_pool(self, model_name: str) -> str:
+        """Return the quota pool for a model ('antigravity' or 'gemini-cli')."""
+        extras = self.get_entry(model_name)
+        if extras and "quota_pool" in extras:
+            return extras["quota_pool"]
+        return "antigravity"  # Default to antigravity quota pool
--- a/providers/registry.py
+++ b/providers/registry.py
@@ -41,6 +41,7 @@ class ModelProviderRegistry:
        ProviderType.AZURE,  # Azure-hosted OpenAI deployments
        ProviderType.XAI,  # Direct X.AI GROK access
        ProviderType.ZEN,  # OpenCode Zen curated models
+        ProviderType.ANTIGRAVITY,  # Antigravity unified gateway (Claude/Gemini via Google OAuth)
        ProviderType.DIAL,  # DIAL unified API access
        ProviderType.CUSTOM,  # Local/self-hosted models
        ProviderType.OPENROUTER,  # Catch-all for cloud models
@@ -140,6 +141,19 @@ class ModelProviderRegistry:
                azure_endpoint=azure_endpoint,
                api_version=azure_version,
            )
+        elif provider_type == ProviderType.ANTIGRAVITY:
+            # Antigravity uses OAuth tokens from file or env var
+            # The provider handles token loading internally
+            from pathlib import Path
+
+            # Check if refresh token is available from env or accounts file
+            accounts_file = Path.home() / ".config" / "opencode" / "antigravity-accounts.json"
+            if not api_key and not accounts_file.exists():
+                logging.debug("Antigravity: No refresh token in env and no accounts file found")
+                return None
+
+            # Initialize provider (it will load tokens internally)
+            provider = provider_class(api_key=api_key or "")
        else:
            if not api_key:
                return None
@@ -341,6 +355,7 @@ class ModelProviderRegistry:
            ProviderType.OPENROUTER: "OPENROUTER_API_KEY",
            ProviderType.CUSTOM: "CUSTOM_API_KEY",  # Can be empty for providers that don't need auth
            ProviderType.DIAL: "DIAL_API_KEY",
+            ProviderType.ANTIGRAVITY: "ANTIGRAVITY_REFRESH_TOKEN",  # Uses OAuth refresh token, not API key
        }

        env_var = key_mapping.get(provider_type)
--- a/providers/shared/provider_type.py
+++ b/providers/shared/provider_type.py
@@ -16,3 +16,4 @@ class ProviderType(Enum):
    CUSTOM = "custom"
    DIAL = "dial"
    ZEN = "zen"
+    ANTIGRAVITY = "antigravity"
--- a/server.py
+++ b/server.py
@@ -392,6 +392,7 @@ def configure_providers():
        value = get_env(key)
        logger.debug(f"  {key}: {'[PRESENT]' if value else '[MISSING]'}")
    from providers import ModelProviderRegistry
+    from providers.antigravity import AntigravityProvider
    from providers.azure_openai import AzureOpenAIProvider
    from providers.custom import CustomProvider
    from providers.dial import DIALModelProvider
@@ -490,6 +491,21 @@ def configure_providers():
        else:
            logger.debug("OpenCode Zen API key is placeholder value")

+    # Check for Antigravity OAuth accounts
+    antigravity_token = get_env("ANTIGRAVITY_REFRESH_TOKEN")
+    antigravity_accounts_file = Path.home() / ".config" / "opencode" / "antigravity-accounts.json"
+    has_antigravity = False
+    if antigravity_token:
+        valid_providers.append("Antigravity (OAuth)")
+        has_antigravity = True
+        logger.info("Antigravity refresh token found - Claude/Gemini models via unified gateway")
+    elif antigravity_accounts_file.exists():
+        valid_providers.append("Antigravity (OAuth)")
+        has_antigravity = True
+        logger.info("Antigravity accounts file found - Claude/Gemini models via unified gateway")
+    else:
+        logger.debug("Antigravity: No refresh token or accounts file found")
+
    # Check for custom API endpoint (Ollama, vLLM, etc.)
    custom_url = get_env("CUSTOM_API_URL")
    if custom_url:
@@ -551,7 +567,13 @@ def configure_providers():
        registered_providers.append(ProviderType.ZEN.value)
        logger.debug(f"Registered provider: {ProviderType.ZEN.value}")

-    # 4. OpenRouter last (catch-all for everything else)
+    # 4. Antigravity (Google unified gateway for Claude/Gemini)
+    if has_antigravity:
+        ModelProviderRegistry.register_provider(ProviderType.ANTIGRAVITY, AntigravityProvider)
+        registered_providers.append(ProviderType.ANTIGRAVITY.value)
+        logger.debug(f"Registered provider: {ProviderType.ANTIGRAVITY.value}")
+
+    # 5. OpenRouter last (catch-all for everything else)
    if has_openrouter:
        ModelProviderRegistry.register_provider(ProviderType.OPENROUTER, OpenRouterProvider)
        registered_providers.append(ProviderType.OPENROUTER.value)
--- a/tests/test_antigravity_provider.py
+++ b/tests/test_antigravity_provider.py
@@ -0,0 +1,311 @@
+"""Unit tests for the Antigravity provider."""
+
+import json
+import os
+import tempfile
+from pathlib import Path
+from unittest import mock
+
+import pytest
+
+from providers.antigravity import AntigravityProvider
+from providers.antigravity_auth import AntigravityAccount, AntigravityTokenManager
+from providers.registries.antigravity import AntigravityModelRegistry
+from providers.shared import ProviderType
+
+
+class TestAntigravityModelRegistry:
+    """Tests for the Antigravity model registry."""
+
+    def test_registry_loads_models(self):
+        """Verify registry loads models from config file."""
+        registry = AntigravityModelRegistry()
+        models = registry.list_models()
+
+        assert len(models) > 0
+        assert "claude-opus-4-5-thinking" in models
+        assert "claude-sonnet-4-5" in models
+        assert "gemini-3-pro-high" in models
+
+    def test_registry_aliases(self):
+        """Verify alias resolution works correctly."""
+        registry = AntigravityModelRegistry()
+
+        # Test alias resolution
+        capabilities = registry.resolve("antigravity-claude-opus-4-5-thinking")
+        assert capabilities is not None
+        assert capabilities.model_name == "claude-opus-4-5-thinking"
+
+        # Test short alias
+        capabilities = registry.resolve("ag-opus-thinking")
+        assert capabilities is not None
+        assert capabilities.model_name == "claude-opus-4-5-thinking"
+
+    def test_model_capabilities(self):
+        """Verify model capabilities are correct."""
+        registry = AntigravityModelRegistry()
+
+        opus = registry.get_capabilities("claude-opus-4-5-thinking")
+        assert opus is not None
+        assert opus.supports_extended_thinking is True
+        assert opus.max_thinking_tokens == 32768
+        assert opus.supports_images is True
+
+        sonnet = registry.get_capabilities("claude-sonnet-4-5")
+        assert sonnet is not None
+        assert sonnet.supports_extended_thinking is False
+
+    def test_quota_pool(self):
+        """Verify quota pool assignment."""
+        registry = AntigravityModelRegistry()
+
+        # Antigravity quota models
+        assert registry.get_quota_pool("claude-opus-4-5-thinking") == "antigravity"
+        assert registry.get_quota_pool("gemini-3-pro-high") == "antigravity"
+
+        # Gemini CLI quota models
+        assert registry.get_quota_pool("gemini-2.5-pro") == "gemini-cli"
+        assert registry.get_quota_pool("gemini-3-flash-preview") == "gemini-cli"
+
+
+class TestAntigravityTokenManager:
+    """Tests for the OAuth2 token manager."""
+
+    def test_empty_initialization(self):
+        """Verify manager handles no accounts gracefully."""
+        # Create temp dir without accounts file
+        with tempfile.TemporaryDirectory() as tmpdir:
+            fake_path = Path(tmpdir) / "nonexistent.json"
+            manager = AntigravityTokenManager(accounts_file=fake_path)
+            assert not manager.has_accounts()
+            assert manager.get_account_count() == 0
+
+    def test_env_var_override(self, monkeypatch):
+        """Verify environment variable takes precedence."""
+        monkeypatch.setenv("ANTIGRAVITY_REFRESH_TOKEN", "test_token_123")
+        monkeypatch.setenv("ANTIGRAVITY_PROJECT_ID", "test-project")
+
+        manager = AntigravityTokenManager()
+        assert manager.has_accounts()
+        assert manager.get_account_count() == 1
+
+    def test_load_accounts_from_file(self):
+        """Verify loading accounts from JSON file."""
+        accounts_data = {
+            "version": 3,
+            "accounts": [
+                {
+                    "email": "test@example.com",
+                    "refreshToken": "token_1",
+                    "projectId": "project-1",
+                    "enabled": True,
+                },
+                {
+                    "email": "test2@example.com",
+                    "refreshToken": "token_2",
+                    "projectId": "project-2",
+                    "enabled": True,
+                },
+                {
+                    "email": "disabled@example.com",
+                    "refreshToken": "token_3",
+                    "projectId": "project-3",
+                    "enabled": False,  # Should be skipped
+                },
+            ],
+        }
+
+        with tempfile.NamedTemporaryFile(mode="w", suffix=".json", delete=False) as f:
+            json.dump(accounts_data, f)
+            temp_path = f.name
+
+        try:
+            # Clear env to ensure file is used
+            with mock.patch.dict(os.environ, {}, clear=True):
+                manager = AntigravityTokenManager(accounts_file=temp_path)
+                assert manager.has_accounts()
+                assert manager.get_account_count() == 2  # Disabled account skipped
+        finally:
+            os.unlink(temp_path)
+
+
+class TestAntigravityAccount:
+    """Tests for the AntigravityAccount dataclass."""
+
+    def test_token_validity(self):
+        """Test token validity checking."""
+        import time
+
+        account = AntigravityAccount(
+            email="test@example.com",
+            refresh_token="token",
+            project_id="project",
+        )
+
+        # No access token
+        assert not account.is_token_valid()
+
+        # Set valid token
+        account.access_token = "access_token"
+        account.token_expiry = time.time() + 3600
+        assert account.is_token_valid()
+
+        # Expired token
+        account.token_expiry = time.time() - 100
+        assert not account.is_token_valid()
+
+    def test_rate_limit_tracking(self):
+        """Test rate limit tracking."""
+        import time
+
+        account = AntigravityAccount(
+            email="test@example.com",
+            refresh_token="token",
+            project_id="project",
+        )
+
+        # No rate limits initially
+        assert not account.is_rate_limited()
+        assert not account.is_rate_limited("claude-opus-4-5-thinking")
+
+        # Set rate limit
+        account.set_rate_limited("claude-opus-4-5-thinking", 60.0)
+        assert account.is_rate_limited("claude-opus-4-5-thinking")
+        assert account.is_rate_limited()  # Any rate limit active
+
+
+class TestAntigravityProvider:
+    """Tests for the Antigravity provider."""
+
+    def test_provider_type(self):
+        """Verify provider type is correct."""
+        # Mock token manager to avoid needing real credentials
+        with mock.patch.object(AntigravityTokenManager, "__init__", return_value=None):
+            with mock.patch.object(AntigravityTokenManager, "has_accounts", return_value=True):
+                provider = AntigravityProvider()
+                assert provider.get_provider_type() == ProviderType.ANTIGRAVITY
+
+    def test_model_capabilities(self):
+        """Verify model capabilities are accessible."""
+        with mock.patch.object(AntigravityTokenManager, "__init__", return_value=None):
+            with mock.patch.object(AntigravityTokenManager, "has_accounts", return_value=True):
+                provider = AntigravityProvider()
+                caps = provider.get_all_model_capabilities()
+                assert len(caps) > 0
+                assert "claude-opus-4-5-thinking" in caps
+
+    def test_model_resolution(self):
+        """Verify alias resolution works through provider."""
+        with mock.patch.object(AntigravityTokenManager, "__init__", return_value=None):
+            with mock.patch.object(AntigravityTokenManager, "has_accounts", return_value=True):
+                provider = AntigravityProvider()
+
+                # Test alias resolution
+                resolved = provider._resolve_model_name("ag-opus-thinking")
+                assert resolved == "claude-opus-4-5-thinking"
+
+                # Test canonical name
+                resolved = provider._resolve_model_name("claude-sonnet-4-5")
+                assert resolved == "claude-sonnet-4-5"
+
+    def test_request_building(self):
+        """Verify request body is built correctly."""
+        with mock.patch.object(AntigravityTokenManager, "__init__", return_value=None):
+            with mock.patch.object(AntigravityTokenManager, "has_accounts", return_value=True):
+                provider = AntigravityProvider()
+
+                capabilities = provider.get_capabilities("claude-opus-4-5-thinking")
+
+                request = provider._build_request(
+                    prompt="Hello, world!",
+                    model_name="claude-opus-4-5-thinking",
+                    project_id="test-project",
+                    system_prompt="You are a helpful assistant.",
+                    temperature=0.7,
+                    max_output_tokens=1000,
+                    thinking_mode="medium",
+                    capabilities=capabilities,
+                )
+
+                assert request["project"] == "test-project"
+                assert request["model"] == "claude-opus-4-5-thinking"
+                assert "contents" in request["request"]
+                assert "systemInstruction" in request["request"]
+                assert "generationConfig" in request["request"]
+
+                # Check thinking config for thinking models
+                config = request["request"]["generationConfig"]
+                assert "thinkingConfig" in config
+                assert config["thinkingConfig"]["thinkingBudget"] > 0
+
+    def test_list_models(self):
+        """Verify list_models returns expected models."""
+        with mock.patch.object(AntigravityTokenManager, "__init__", return_value=None):
+            with mock.patch.object(AntigravityTokenManager, "has_accounts", return_value=True):
+                provider = AntigravityProvider()
+                models = provider.list_models(respect_restrictions=False)
+
+                assert len(models) > 0
+                # Check that both canonical names and aliases are present
+                assert any("claude" in m.lower() for m in models)
+                assert any("gemini" in m.lower() for m in models)
+
+
+@pytest.mark.integration
+class TestAntigravityIntegration:
+    """Integration tests (require actual credentials and valid tokens).
+
+    These tests require:
+    - Valid Antigravity credentials (env var or accounts file)
+    - Working OAuth tokens that can be refreshed
+
+    Run with: pytest -m integration tests/test_antigravity_provider.py
+    Skip with: pytest -m "not integration" tests/test_antigravity_provider.py
+    """
+
+    @pytest.mark.skipif(
+        not os.getenv("ANTIGRAVITY_REFRESH_TOKEN")
+        and not Path.home().joinpath(".config/opencode/antigravity-accounts.json").exists(),
+        reason="Antigravity credentials not available",
+    )
+    def test_token_refresh(self):
+        """Test actual token refresh with real credentials."""
+        manager = AntigravityTokenManager()
+        assert manager.has_accounts()
+
+        # This will trigger a token refresh - may fail if tokens are expired
+        try:
+            token, project_id, headers = manager.get_access_token()
+            assert token is not None
+            assert len(token) > 0
+            assert project_id is not None
+        except RuntimeError as e:
+            if "rate limited or have invalid tokens" in str(e):
+                pytest.skip("Antigravity tokens are expired or invalid")
+            raise
+
+    @pytest.mark.skipif(
+        not os.getenv("ANTIGRAVITY_REFRESH_TOKEN")
+        and not Path.home().joinpath(".config/opencode/antigravity-accounts.json").exists(),
+        reason="Antigravity credentials not available",
+    )
+    def test_simple_generation(self):
+        """Test actual content generation with Antigravity API."""
+        provider = AntigravityProvider()
+
+        try:
+            response = provider.generate_content(
+                prompt="Say 'hello' and nothing else.",
+                model_name="claude-sonnet-4-5",
+                temperature=0.0,
+                max_output_tokens=100,
+            )
+        except RuntimeError as e:
+            if "rate limited or have invalid tokens" in str(e):
+                pytest.skip("Antigravity tokens are expired or invalid")
+            raise
+
+        assert response is not None
+        assert response.content is not None
+        assert "hello" in response.content.lower()
+        assert response.provider == ProviderType.ANTIGRAVITY