Pilot now if you can isolate low-risk repositories, buy Gemini Enterprise Standard or Plus on Google Cloud, and assign an engineering owner with security oversight. Wait if required compliance coverage is missing, credentials cannot be isolated, or nobody can review agent actions. We recommend proving control before expanding autonomy.
What is Google Antigravity—and what is it not?
Google Antigravity is an agentic development platform: agents can plan, execute and verify software tasks across an editor, terminal and browser. It is not Gemini chat or simply a Gemini Code Assist rename. This guide is for Antigravity enterprise teams getting started on Google Cloud—access paths, pilot billing and controls—not a consumer Pro/Ultra walkthrough. Google’s Antigravity launch announcement explains the agent-first approach. Enterprise sign-in options are set out in the enterprise setup guide.
For a business pilot, use the Antigravity App, Antigravity CLI or Antigravity IDE extensions. Business sign-in is generally available for the VS Code and Visual Studio extensions, and still in Preview for JetBrains, Zed, Xcode and Eclipse. The standalone Antigravity IDE is not supported for enterprise deployments, according to Google’s Antigravity enterprise setup guide. Do not assume every downloadable surface supports the same account or terms.
Antigravity on Google Cloud: which access path fits?
Antigravity on Google Cloud is the path that matters when people search for “Antigravity cloud”: you run Antigravity under Cloud identity, billing and admin controls instead of a personal Google AI plan. The table below is how to choose that path.
The important buying decision is not the largest personal quota. It is which terms, data protections, billing model and administrative controls apply.
| Access path | What to know |
|---|---|
| Consumer Google AI Pro / Ultra | Individual terms. Fine for approved, non-sensitive experimentation; not a substitute for business governance. |
| Gemini Enterprise Standard or Gemini Enterprise Plus (seat-licensed) | Includes pooled AI developer tools credits for first-party Gemini models. No separate per-user Antigravity license. This guide’s focus. |
| Pay-as-you-go edition on Google Cloud | Usage-priced, with no bundled monthly AI developer tools credits. Verify your Cloud Billing account type and set spend limits before use. |
Some eligible customers can also use a Standard Emerging Market edition, which does not include default bundled AI developer tools credits; confirm eligibility with your Google account team.
Not every edition labelled “business” includes AI developer tools. On Google’s edition comparison, the similarly named Gemini Enterprise Business edition does not include them. AI developer tools are listed for the Gemini Enterprise Standard and Gemini Enterprise Plus and for Pay-as-you-go. Check that table before you buy.
Trial subscriptions do not include AI developer tools. Confirm a paid, non-trial subscription before pilot kickoff. Access is also rolling out to select customers, so confirm your location and rollout status too.
Google covers consumer vs Cloud terms on its Antigravity plans page and pooled credits on its credit and pay-as-you-go above-quota rules. Included credits are enforced as a rolling seven-day shared pool and do not roll over. They are not unlimited usage.
Scope of this guide: The step-by-step pilot below assumes paid, seat-licensed Gemini Enterprise Standard or Gemini Enterprise Plus subscriptions. Pay-as-you-go and API-key access are also valid paths, but their billing and control model is different and outside this guide’s scope. Most of the security settings below still apply.
Prices, quotas, models and rollout access change often. Recheck Google’s model catalog and quota rules before procurement and again before launch.
How should you buy it, and how is a pilot billed?
Buying through Gemini Enterprise on Google Cloud rather than a personal plan is the decision that gets you central admin controls, business terms and pooled credits; the rest is choosing an edition and setting limits. Included credits are not a spending cap. Plan for these billing behaviors, documented in Google’s credit and pay-as-you-go above-quota rules and AI developer tools settings:
- First-party Gemini usage above the pool: When the pooled credits run out, usage stops until the weekly reset unless an administrator turns on pay-as-you-go usage above quota. Once that setting is on, further Gemini usage is billed at Agent Platform rates.
- Third-party models: Claude models are pay-as-you-go only and require pay-as-you-go usage above quota to be turned on. Pooled credits do not cover them.
- Hybrid and background charges: Even when a developer selects a third-party model, Antigravity keeps using lightweight Gemini models in the background for coordination, and a Gemini image model for image generation. That usage can also incur charges.
- Monthly project spend limit: Set one. Google says it applies to all models, including third-party models.
Billing lands on prepaid credits for self-serve Cloud Billing accounts or on the monthly invoice for invoiced accounts. Name a budget owner and review consumption weekly.
Does Antigravity train on your code?
It depends on the subscription and terms, not on which login screen a developer uses. Under the Antigravity terms, individual consumer interactions may be used to improve Google products and machine-learning technologies, with possible human review. Turning off telemetry does not stop processing needed to provide and protect the service.
On the supported Gemini Enterprise Cloud subscription path, use is governed by the Google Cloud terms your administrator accepted. Under those terms, Google’s enterprise data-use statement says corporate data is not used for training. Prompts sent to administrator-enabled third-party models are also not used for training, per the settings documentation. During onboarding, verify that each developer signs in with the business account option against a licensed project. A work email address alone does not prove that.
Two caveats:
- Marketplace: Google’s terms state that enterprise or other signed or accepted terms do not apply to the Antigravity Marketplace. Treat Marketplace items as a separate approval decision.
- Data residency is limited: Business endpoints are available in
global,usandeulocations. Global endpoints are still required for license discovery, and preview models and image generation are global-only. Third-party model inference runs inglobal,usoreuonly, with no in-country regions (enterprise setup guide).
“No training” does not mean “no governance needed.” Decide what code, prompts, outputs and logs may contain before granting access.
Which security settings should you change first?
The defaults are not uniformly restrictive. Google’s AI developer tools settings document terminal auto-execution as Always proceed and sandbox mode as Disabled by default. Outside-workspace file access defaults to Deny.
| Control | Pilot setting |
|---|---|
| Terminal auto-execution | Require review |
| Sandbox mode | Enabled; test isolation |
| Outside-workspace file access | Keep Deny |
| Browser access | Off unless needed; then on with explicit Allowed URLs |
| Browser JavaScript | Disabled or Request review |
| MCP servers | Off, or named servers on an allowlist |
| Third-party models | Off until security and budget approval |
| Prompt/response and metadata logging | Set deliberately per evidence policy |
/teamwork-preview | Consider disabling until evaluated (on by default) |
| Monthly project spend limit | Set before first use |
Browser access: Turning browser access on without an Allowed URLs list blocks nothing, because no URLs are blocked by default. Enable it only for a task that needs it, with an explicit list.
ACP IDEs and MCP: IDEs that use the Agent Client Protocol (such as JetBrains IDEs and Android Studio) require MCP to be enabled. In those IDEs, Google says administrator controls are best-effort, because the IDE can inject its own MCP tools that are not subject to enforcement. Keep pilots on surfaces where your policies are fully enforced, or accept that gap explicitly.
Configure these policies before developers run their first task. Test them with harmless commands; a checked box is not proof that the workflow is contained. Logs are stored in your project, and Google says it does not access them for model training or human review. They can still contain sensitive material, so assign access, retention and incident-response ownership.
Which compliance requirements rule out a deployment?
Google states that Antigravity does not support SOC 1, SOC 2, SOC 3, ISO 27001, ISO 42001, FedRAMP Moderate or High, IL4/IL5, ITAR, or Access Transparency. See the AI developer tools overview and limitations and security and compliance controls.
Do not borrow certifications from the broader Google Cloud subscription. If your workload requires one of these assurances, pause that deployment and involve compliance.
What belongs in the governance checklist?
Use this as a pre-access sign-off:
- Identity and roles: Name the administrator and pilot owner. Configuration needs the Discovery Engine Admin role and day-to-day use needs the Discovery Engine User role; the exact IAM role names are in Google’s AI developer tools overview. The admin role alone does not grant user access.
- Repository scope: Approve named repositories. Exclude production configuration, customer datasets and unrelated code. Require human pull-request review.
- Secrets: Remove credentials from pilot fixtures. Use short-lived, least-privilege test credentials; deny production deployment authority.
- Network: Implement Google’s documented endpoint allowlist rather than broad outbound access.
- Third-party models, MCP and Marketplace: Review providers, terms, permissions and data flows individually.
- Cost: Name a budget owner, set the project spend limit and document who approves pay-as-you-go usage above quota.
How should a phased pilot work?
We suggest a four-week evaluation, led by one named engineering manager with a security counterpart.
Phase 1—Prepare: Approve two low-risk repositories, apply the settings above and test access revocation. Establish a baseline using comparable tasks completed without agents.
Phase 2—Run narrowly: Start with test generation, bug reproduction using synthetic inputs and small refactors, in line with Google’s documented use cases. Keep production access and autonomous merging out of scope.
Phase 3—Measure: Track accepted changes, correctness, reviewer minutes, rework, elapsed task time and spend. Generated code volume is not a success measure.
Phase 4—Decide: Expand only if agreed quality and cost thresholds are met, logs support investigation, and no unresolved access violations remain. Stop immediately for secret exposure, unauthorized execution or policy bypass.
Frequently asked questions
Can Antigravity enterprise teams use personal Pro or Ultra accounts for company code?
We recommend Antigravity on Google Cloud with Gemini Enterprise Standard or Plus for proprietary code. Personal subscriptions have different terms and data-use provisions; paying for a consumer tier does not establish business protections.
Does sandboxing make every task safe?
No. We still require scoped credentials, reviewed commands and human approval of changes. Treat sandboxing as one layer, not permission to expose production systems.
Should we promise a productivity uplift?
No. Set a hypothesis and measure your own results, including review and rework. Expand because your evidence supports it, not because a demonstration looked fast.
Can we use the standalone Antigravity IDE for our enterprise rollout?
No. Google says the standalone IDE is not supported for enterprise deployments. Use the App, CLI or IDE extensions, and note that business sign-in is GA for VS Code and Visual Studio but Preview for JetBrains, Zed, Xcode and Eclipse (enterprise setup guide).
Is Antigravity SOC 2 or ISO 27001 certified?
No. Google lists SOC 1/2/3, ISO 27001, ISO 42001, FedRAMP Moderate and High, IL4/IL5, ITAR and Access Transparency as not supported for Antigravity. Do not assume the broader Google Cloud subscription’s certifications cover it (limitations).
- Google Cloud edition comparison for AI developer tools
- Credit and pay-as-you-go above-quota rules
- AI developer tools administrative settings
- AI developer tools overview, roles and limitations
- Security and compliance controls
- Antigravity enterprise setup, surfaces and networking
- Antigravity terms of service
- Antigravity enterprise data-use statement
