On November 2, 2026, Google released the google-cloud-developer plugin, a new tool designed to integrate cloud-management capabilities directly into agentic workflows. Available through the Build with Google section of the Antigravity Marketplace, the plugin can be installed by running agy plugin install google-cloud-developer. This release addresses a specific set of challenges Google identified regarding AI agents in cloud environments: the risk of agents relying on stale training data, the potential for agents to invent deprecated SDK syntax or invalid flags, and the danger of unguided agents inadvertently enabling billable APIs or provisioning infrastructure without explicit permission.
To mitigate these risks, the plugin utilizes the Model Context Protocol (MCP)—a standard that enables AI agents to connect to external tools and data sources—to bridge the gap between agentic reasoning and cloud environments. The architecture is structured into three distinct layers: a documentation connection, a suite of on-demand skills, and a set of described safety guardrails for cloud operations.
The first layer, the Developer Knowledge MCP server, provides agents with direct access to official, public Google documentation. As detailed in the Developer Knowledge MCP documentation, this server offers three primary tools: search_documents, get_documents, and answer_query. These tools allow the agent to query official resources via streamable HTTP. A notable feature of the Antigravity Marketplace integration is that this documentation server is preauthorized by default. Consequently, developers do not need to manually manage API keys, configure OAuth token exchanges, or edit local configuration files to enable this specific connection.
The second layer consists of five specialized skills designed for on-demand routing. Google states that this approach is intended to maintain efficiency by fetching capabilities only when they are required, rather than preloading a large set of tools. These skills include finding-google-skills, which discovers and fetches specialized capabilities from a public catalog; retrieving-developer-knowledge, which provides grounded code samples and documentation; and gcloud, which is designed to validate syntax and perform safety checks on CLI commands. Additionally, the plugin includes google-cloud-recipe-auth, which steers users toward secure authentication patterns like Application Default Credentials (ADC) while discouraging risky practices such as downloading service-account JSON keys, and google-cloud-recipe-onboarding, which guides project setup, billing verification, and state audits before service activation.
The third layer involves safety design features intended to govern how agents interact with cloud infrastructure. Google claims the plugin is built to inspect local environment and CLI configurations to design safe execution roadmaps before any commands are actually run. Furthermore, Google states that the system is designed to enforce least-privilege IAM patterns and requires explicit user consent before enabling billable APIs or modifying cloud infrastructure. It is important to distinguish between the preauthorized documentation access and these operational safeguards. While the documentation server is preauthorized to streamline information retrieval, the plugin’s design intent for billable or infrastructure-changing actions is to require active user intervention.
Antigravity’s general MCP documentation says unconfigured MCP tools default to Ask mode, requiring user approval before execution. Its policy settings can allow an individual tool, all tools on a server, or all MCP tools. That describes the platform’s general permission model; it does not verify how this plugin implements the separate consent requirement Google describes.
For builders, the concrete change is an assembled route: one Marketplace install connects official documentation, task-specific instructions, and Google Cloud command guidance. Google says this path removes manual authentication setup for the documentation server and routes specialized skills on demand. That is a product-design change, not a measured claim about improved accuracy or safer deployments. The launch post describes intended controls; it is not an independent test and does not report adoption or operational outcomes. The distinction matters: preconfigured access to reference material does not itself establish permission to change a project, while a consent rule described by a vendor is not proof of how reliably it works in practice.