google-antigravity-sdk

The saved excerpt is a snapshot from when this skill was listed. The external source remains the complete and most current version.

---
name: google-antigravity-sdk
description: "Design, implement, and debug autonomous AI agents and multi-agent systems using the Google Antigravity (AGY) SDK. ACTIVATE this skill when the user wants to create, configure, or orchestrate Google Antigravity agents."
---

# Google Antigravity SDK

## Installation & Setup

Before proceeding with any Google Antigravity tasks, ensure the environment is
ready:

-   **Verify Applicability**: If operating in an existing codebase, verify that
    using this Python SDK is possible and appropriate for the project.
-   **Check Dependencies**: Check if `google-antigravity` is listed in the
    project's dependencies (e.g., `requirements.txt`, `pyproject.toml`).
-   **Install Package**: Ensure the `google-antigravity` Python package is
    installed.
-   **Authentication Setup**: Check for a valid `GEMINI_API_KEY` environment
    variable or a `.env` file (required to access Gemini models).
    -   If credentials are missing, you MUST actively help the user get set up
        with an API key by providing the following link:
        -   Default to Google AI Studio:
            `https://aistudio.google.com/app/api-keys`
    -   Explain that the API key can be passed explicitly in code as shorthand
        (e.g., `LocalAgentConfig(api_key="...")`) or automatically read from the
        environment.
    -   For Gemini Enterprise Agent Platform (formerly Vertex AI)
        authentication, the SDK supports both Standard Mode and Express Mode:
        -   **Standard Mode (ADC)**: Instruct the user to run
            `gcloud auth application-default login` and configure the agent with
            `vertex=True` along with `project` and `location` in
            `LocalAgentConfig`.
        -   **Express Mode (API Key)**: Configure the agent with `vertex=True`
            along with `api_key="your-express-api-key"` in `LocalAgentConfig`
            (no ADC or regional project/location needed).
    -   **Note**: For local models (`LiteRTAgentConfig` or
        `LocalOpenAIAgentConfig`), no API key or cloud credentials are needed.
        See `references/local_models.md` for setup details.

## Routing Table

Use the following information to dig deeper into specific topics based on the
user request. Read the referenced files or explore the directories to find
relevant information.

### References

-   If the user needs to understand the high-level overview and core concepts of
    the Google Antigravity SDK (Agent, Conversation, Connection), read
    `references/architecture.md`.
-   If the user needs to perform advanced agent configuration (e.g., selecting
    appropriate models, configuring execution behavior via `agent_behavior`—defaulting
    to autonomous vs interactive—or configuring connection reliability), or
    understand the critical rules for model identifiers to avoid assumptions,
    read `references/agent_configuration.md`.
-   If the user needs to extend an agent's capabilities by integrating Model
    Context Protocol (MCP) servers, or configure tool permissions for the agent,
    read `references/mcp_integration.md`.
-   If the user needs to define safety policies, resolve execution order,
    restrict agent actions using predicates, or run terminal commands inside an
    OS-level sandbox, read `references/safety_policies.md`.
-   If the user needs to debug failed agents, stream logs, or implement error
    recovery using hooks to make agents robust, read
    `references/error_handling.md`.
-   If the user needs to monitor costs, track token usage (including thinking
    tokens), or build custom audit logs for advanced monitoring, read
    `references/observability.md`.
-   If the user needs to see a list of built-in tools and understand their default state, read `references/built_in_tools.md`.
-   If the user needs to run agents locally using on-device models (e.g., Gemma
    via LiteRT, or via OpenAI-compatible APIs), understand hardware
    requirements, or set up a local model envir