Google Antigravity Enterprise Security, MCP, and Browser Policies
On this page
Quick answer
Gemini Enterprise exposes administrative settings for AI developer tools, including workspace sandboxing and browser and MCP access. Google also describes centralized audit logging for prompts, responses, and metadata. Use those controls as one layer in a larger authorization design.
An enabled MCP server is not permission for every tool call, and an audit log does not prevent an unsafe action.
Separate the boundaries
- Workspace: allowed repositories, files, generated artifacts, symlinks, and data classifications.
- Terminal: command allow and deny rules, environment variables, credentials, package installation, and confirmation points.
- Browser: destinations, login state, uploads, downloads, form submission, purchases, and external communication.
- MCP: server identity, transport, tools, read and write scope, secrets, argument validation, destructive actions, and revocation.
- Models: enabled models, global endpoint warnings, data residency requirements, and fallback behavior.
- Audit: event coverage, retention, access, export, alerting, incident correlation, and privacy review.
Start read-only. Add one write capability at a time with a named owner, explicit resource scope, confirmation for consequential actions, idempotency, and rollback. Test denial paths and verify that the log captures enough context without exposing secrets.
Use the setup guide for roles and eligibility and the readiness checklist for the pilot gate.
Official sources
Source check: August 23, 2026. Verify settings, custom roles, audit coverage, model routing, data residency, browser, and MCP behavior in the target project.