The Google Antigravity SDK is a Python library for building AI agents powered by Google Antigravity and Gemini.
The SDK addresses the complexity of implementing agentic loops by providing a secure, scalable, and stateful infrastructure layer that abstracts away the mechanics of agent execution. Developers define what their agent does through system instructions, custom tools, and multimodal inputs, while the SDK handles binary discovery, tool wiring, hook registration, and policy defaults. The library supports both streaming responses for real-time output and advanced features like intercepting internal model reasoning and tool call dispatches. By default agents run in read-only mode for safety, with write capabilities available through explicit configuration.
The SDK suits teams building AI agents on Google Cloud infrastructure who want to leverage Gemini's capabilities without managing the underlying agentic loop themselves. It supports two authentication modes for Gemini Enterprise Agent Platform: Express Mode using an API key for quick setup, and Standard Mode for enterprise deployments with regional endpoints and Application Default Credentials. The tool integrates with the Model Context Protocol and allows registration of custom Python functions as callable tools. Multimodal ingestion is supported, enabling agents to process images, videos, audio, and documents alongside text instructions.
The project maintains active development with regular updates to its core functionality and examples. The codebase includes comprehensive examples demonstrating common patterns like the hello world agent and vertex authentication modes. Documentation is embedded in the README with clear guidance on installation requirements, noting that the SDK depends on a compiled runtime binary distributed through PyPI wheels rather than being available through repository cloning alone. The project provides both simple and advanced usage patterns, from the basic Agent class for straightforward implementations to the Conversation class for scenarios requiring full lifecycle control.
