Configure AI developer tools settings

Use the AI developer tools settings page in Gemini Enterprise to configure developer tools, enforce security and compliance controls, and manage model availability for your organization.

Before you begin

Make sure you have the following:

  • Invoiced Cloud Billing account: To use AI developer tools, your project must have Cloud Billing enabled. AI developer tools are available for invoiced accounts only. For more information, see Cloud Billing account types.

  • Supported editions: To access AI developer tools, users must have a subscription to the Gemini Enterprise Standard, Plus, Standard Emerging Market, or Pay-as-you-go edition.

  • Supported regions: For more information about regional availability and feature limitations, see Data residency for Gemini Enterprise Standard and Plus Editions and Gemini Notebook.

  • Required roles: You must have the Gemini Enterprise Admin (roles/discoveryengine.agentspaceAdmin) role to configure AI developer tools settings.

  • Required APIs: Your project must have the following APIs enabled:

    • Discovery Engine API (discoveryengine.googleapis.com): Provides core platform services, security and compliance controls, and settings management for Gemini Enterprise.

    • Business AI Code API (businessaicode.googleapis.com): Provides AI-powered code generation for Antigravity in Gemini Enterprise. You must enable this API on your project to let authorized principals reach the service from your project.

    When you turn on the AI developer tools toggle in the Google Cloud console, the system checks your project configuration and automatically enables these required APIs if you haven't already enabled them.

Turn on AI developer tools

The availability of AI developer tools depends on your Gemini Enterprise subscription type:

  • New Standard, Plus, or Pay-as-you-go subscriptions: AI developer tools is enabled by default in the location where you purchased the subscription if it is supported for that location or region.

  • Existing subscriptions: An administrator must manually enable AI developer tools in the Google Cloud console for each location to activate the tools for licensed users.

To manually enable AI developer tools for an existing Gemini Enterprise subscription, follow these steps:

  1. In the Google Cloud console, go to Gemini Enterprise.

    Go to Gemini Enterprise

  2. Click Settings, and select the AI developer tools tab.

  3. Turn on the AI developer tools toggle.

  4. Configure the settings in the Security, Compliance, and Model availability sections.

Configure security settings

As an administrator, you can prevent unauthorized data exposure across developer interfaces. Use the settings in this section to restrict access to external sources and manage automated execution permissions. For example, you can enforce mandatory security policies to prevent users from overriding safety rules, and require approval before AI agents run terminal commands or select links.

Manage security policies

You can configure security policies to control how AI agents access local files outside their working folders and execute terminal commands in developer workspaces.

Setting Description Default status
Outside of file access policy Specify the agent's permission level for accessing files outside of its designated working folders. The following options are available:
  • Deny: Prevents the agent from accessing files outside its working folders.
  • Always ask: Prompts the user for permission before the agent accesses files outside its working folders.
  • Allow: Permits the agent to access files outside its working folders without prompting.
 Deny
Terminal auto-execution mode Enable this setting to let agents run terminal commands automatically to streamline multi-step workflows. The following options are available:
  • Require review: Prompts the user for approval before running any terminal command.
  • Proceed in sandbox: Automatically runs commands in a secure, isolated environment.
  • Always proceed: Automatically runs commands in the current environment without prompting.
 Always proceed
Sandbox mode Enable this setting to restrict agent tools to a secure, isolated local sandbox. Disabled

Manage external web access

If your security policy requires controlling agent access to external websites to prevent data exfiltration, these settings let you limit browser access or restrict it to a list of allowed URLs.

Setting Description Default status
Browser access Enable this setting to control external web access. This prevents data exfiltration and defines organizational boundaries. Disabled
Allowed URLs Specify a list of allowed URLs that the agent can access in the browser. When configured, the system blocks all other URLs.

This setting applies only to the Antigravity browser agent. You can run the browser agent in Antigravity by typing the /browser command in the chat field.

 None. By default, no URLs are blocked.
Browser JavaScript auto-execution mode Enable this setting to let agents run JavaScript to support advanced web-based coding and testing tasks. The following options are available:
  • Disabled: Prevents the agent from running JavaScript.
  • Request review: Prompts the user for approval before the agent runs JavaScript.
  • Allowed: Permits the agent to run JavaScript automatically.
 Disabled

Manage Model Context Protocol (MCP)

You can configure access to Model Context Protocol (MCP) servers, which lets AI agents securely access additional data sources and tools.

Setting Description Default status
MCP servers Enable this setting to manage access to Model Context Protocol (MCP) servers to securely extend the AI's data context. Disabled
Allowed server configuration This setting lets administrators define specific server configurations for MCP server access. Administrators define the allowlist as a JSON object with configurations for local and remote MCP servers. The UI provides real-time validation for the following fields:
  • Local MCP servers (local_servers):
    • id (string): The unique identifier of the MCP server.
    • command (string, optional): The command to run the local MCP server. If specified, this overrides the user's local configuration.
    • args (array of strings, optional): The arguments passed to the command. If specified, these override the user's local configuration.
  • Remote MCP servers (remote_servers):
    • id (string): The unique identifier of the MCP server.
    • url (string, optional): The endpoint URL of the remote MCP server. Required for custom remote servers and optional for pre-configured remote servers.

Example configuration

{
  "mcpServers": {
    "local_servers": [
      {
        "id": "gopls-mcp-server",
        "command": "go",
        "args": [
          "run",
          "golang.org/x/tools/gopls@latest",
          "mcp"
        ]
      }
    ],
    "remote_servers": [
      {
        "id": "bigquery"
      },
      {
        "id": "custom-remote-server",
        "url": "https://mcp.example.com/mcp"
      }
    ]
  }
}
Not configured

Configure compliance settings

If your organization operates in a regulated industry or has specific internal policies, you might need to maintain audit logs of AI interactions for compliance. This section's settings let you configure logging and data retention policies to meet your legal and regulatory requirements.

Setting Description Default status
Prompts and responses logging Enable this setting to record interactions between developers and the AI to compute usage metrics and maintain an audit trail for security reviews. The system stores logs securely within your organization's project, and Google never accesses the logs for model training or human review. Disabled
Metadata logging Enable this setting to record product metadata to help you track usage and analytics across your organization, including third-party model token consumption. The system stores logs securely within your organization's project, and Google never accesses the logs for model training or human review. Disabled

Configure model availability

Use the settings in this section to authorize specific AI models and manage access levels. This ensures developers only use AI models that comply with your organization's safety and usage policies. For example, you can give developers access to preview models for experiments, or restrict tools to use only authorized models.

Key considerations before configuring models

Before you configure models, consider the following:

  • Policy alignment: Enabling a model in Gemini Enterprise doesn't override Gemini Enterprise Agent Platform organization policies. To prevent pay-as-you-go requests that exceed your standard quota from failing, ensure that the allowed models match across both platforms.

  • Data residency: Preview models and the Gemini 3 Pro Image model are available only in the global region. Therefore, they don't support data residency (DRZ) or machine learning (ML) regional processing commitments in the US or EU multi-regions.

  • Third-party model inference and data handling: Third-party models are available only after an administrator enables them and accepts the provider's terms. When a developer selects an enabled third-party model, Gemini Enterprise sends the prompt and the code context needed for that request to the model, with the same processing and protections as Gemini models. Prompts and responses aren't used to train models. Third-party model inference is available in the global, us, and eu locations. In-country regions aren't supported. To see the list of countries and regions that can access the model, see Anthropic's Supported countries and regions. For more information on data residency, see Data residency for Gemini Enterprise Standard and Plus Editions and Gemini Notebook.

  • Hybrid execution and billing disclosures: Gemini models draw from your organization's pooled edition seat credits, while administrator-enabled third-party models are billed as pay-as-you-go usage on your Cloud Billing account. When a developer selects an enabled third-party model (such as Anthropic Claude Opus 5.5 or Anthropic Claude Sonnet 5.5) for primary reasoning, planning, tool orchestration, and code generation tasks, Antigravity continues to use lightweight Gemini models in the background for client-side coordination—such as context management and summaries—and Gemini 3 Pro Image for image generation. Using these models can incur charges. For more information, see Quotas and overages and Track third-party model token usage.

Configure authorized models

You can configure authorized models to control which AI models are available to your AI developer tools.

Setting Description Default status
Antigravity authorized models Enable this setting to specify which AI models the Antigravity tool is authorized to use within the project. In addition to Gemini models made by Google, administrators can opt in to enable third-party models, including Anthropic Claude Opus 5.5 and Anthropic Claude Sonnet 5.5. For information about each model's capabilities, limitations, and safety evaluations, see the model cards for Anthropic Claude Opus 5.5 and Anthropic Claude Sonnet 5.5. Gemini models are enabled by default. Third-party models are turned off by default and require administrator consent.

Supported models

AI developer tools (Antigravity 2.0, the Antigravity CLI, and Antigravity for IDEs) support the following Gemini models and partner models:

Model display name Provider Availability in Gemini Enterprise
Gemini 3.8 Flash Google Enabled by default for licensed users; configurable in Antigravity authorized models
Gemini 3.7 Flash Google Enabled by default for licensed users; configurable in Antigravity authorized models
Gemini 3.6 Flash Google Enabled by default for licensed users; configurable in Antigravity authorized models
Gemini 3.1 Pro (Preview) Google Enabled by default for licensed users; configurable in Antigravity authorized models
Anthropic Claude Opus 5.5 Anthropic Turned off by default; requires administrator opt-in (see Enable or turn off third-party models)
Anthropic Claude Sonnet 5.5 Anthropic Turned off by default; requires administrator opt-in (see Enable or turn off third-party models)

For information about quotas, pooled edition credits, and pay-as-you-go billing for these models, see Quotas and overages and Configure overages.

Thinking level for third-party models

Developers can use Anthropic Claude Opus 5.5 and Anthropic Claude Sonnet 5.5 at a thinking level of Low, Medium, High, or Max. Each level appears as a separate entry in the model picker, for example Anthropic Claude Opus 5.5 (High). Medium is the default. When you enable a third-party model, all of its thinking levels are available to developers. You can't restrict individual thinking levels.

Enable or turn off third-party models

Third-party models are billed as pay-as-you-go usage and require overages. Before you enable a third-party model, turn on overages. If overages are turned off later, developers can't use third-party models until overages are turned back on. To control costs, you can set a monthly spend limit for your project. The limit applies to all models, including third-party models.

To enable or manage third-party models (such as Anthropic Claude Opus 5.5 and Anthropic Claude Sonnet 5.5) in the Google Cloud console:

  1. In the Google Cloud console, go to Gemini Enterprise.

    Go to Gemini Enterprise

  2. Click Settings, and then select the AI developer tools tab.

  3. In the Antigravity authorized models setting, turn on the toggle for the third-party model that you want to enable.

  4. In the embedded Terms of Service dialog that appears, review and accept the third-party terms.

If you turn off a third-party model, developers lose access to it shortly afterwards.

If the model provider restricts access to a third-party model for your organization, developers can't use that model. To resolve the issue, contact Google Cloud Support.

For developer instructions on signing in to Antigravity and selecting models, see the Antigravity enterprise documentation.

Security considerations for third-party models

AI models, including third-party models, can follow instructions hidden in content they process, such as files, web pages, or tool output. This is known as prompt injection. Antigravity provides the following protections when developers use third-party models:

  • Administrator opt-in: Third-party models are off by default. Developers can use a third-party model only after an administrator enables it.
  • Provider safety protections: Third-party models include the provider's built-in safety protections. For more information, see Claude safety.
  • Your security policies: The security policies and external web access settings that you configure apply to agents regardless of the model that a developer selects.

Best practices

  • Enable only the third-party models that your developers need.
  • Set Terminal auto-execution mode to Require review or Proceed in sandbox, so that developers approve commands before they run.
  • Keep Outside of file access policy set to Deny.
  • Turn on Browser access and limit agents to a list of Allowed URLs.
  • Turn on Sandbox mode for projects that work with untrusted code, dependencies, or web content.
  • Encourage developers to review the commands and code changes that agents propose before approving them.
  • Set a monthly spend limit for your project.

What's next