Google Antigravity

MCP tools connect the agent to outside systems under per tool allow, ask or deny rules, plugins bundle MCP servers and skills managed from the Customizations tab, and the agent operates the editor, terminal and a browser as tools.

Sourceantigravity.google/docs/permissions, docs/changelog and docs/sdk/overviewread 2026-09-29

The main agent delegates to custom subagents, whose invocations appear as cards with live status, custom agents can switch off default prompt sections and tools, a Teamwork preview coordinates agent teams, and the SDK builds multi agent systems with subagents.

Sourceantigravity.google/docs/changelog, docs/sdk/overview and llms.txt; antigravity.google/docs/subagentsread 2026-09-29

The agent reads the project's code with codebase search, rules (now with their own token budget) and installed skill bundles, and reaches outside data through MCP. Context is assembled per run, and no persistent index, graph or embeddings layer over the customer's code is named in the docs.

Sourceantigravity.google/docs/build-with-google, docs/changelog and llms.txtread 2026-09-29

Permission lists with Deny over Ask over Allow make the agent pause for explicit approval before matching commands, file access, URL navigation or MCP tool use, Request Review mode confirms every command, web browsing and MCP tools prompt by default, administrators can enforce enterprise policies, and /plan lets users read and edit the plan before any code is written.

Sourceantigravity.google/docs/permissions and docs/changelogread 2026-09-29

Enterprise use runs through Gemini Enterprise in the customer's Google Cloud project with single sign on, Workforce Identity Federation for providers such as Okta, Application Default Credentials for headless use, VPC Service Controls, request and response logging and enforced enterprise permission policies, under Google Cloud's enterprise security controls and terms. The limit: no attestation specific to Antigravity is named on the page, so ask Google which of its Google Cloud attestations cover Antigravity.

Sourceantigravity.google/docs/enterprise and docs/permissionsread 2026-09-29

The agent records its work as reviewable artifacts (implementation plans, task lists, walkthroughs, screenshots and browser recordings), subagent invocations show as cards with live status, and enterprise deployments log requests and responses for audit trails.

Sourceantigravity.google/docs/changelog, docs/enterprise and llms.txt (Artifacts, Walkthrough, Screenshots, Browser Recordings); antigravity.google/docs/artifactsread 2026-09-29

Projects group conversations across workspaces and worktrees, and conversations resume with /resume, which carries session history. No store the agent writes and reads across independent sessions with a stated scope, lifetime and delete path is documented; rules are standing instructions, and cron sidecars schedule work rather than remember it.

Sourceantigravity.google llms.txt (Projects, Rules, Sidecars) and docs/sidecarsread 2026-09-29

Enterprise deployment runs inside the organization's own Google Cloud project with global, US multi region or EU multi region endpoints for data residency and VPC Service Controls, and Google states enterprise prompts, responses, code and telemetry are never stored outside the customer's private environments.

Sourceantigravity.google/docs/enterpriseread 2026-09-29

Google publishes named bundles that each do their own job, among them Modern Web Guidance, a Firebase bundle, Android CLI, DeepMind Science skills, Chrome DevTools, Dart and Flutter, a Data Agent Kit starter pack and Google Maps Platform, and plugins are discovered, installed and managed from the Customizations tab.

Sourceantigravity.google/docs/build-with-google and docs/changelogread 2026-09-29

A sidecar's schedule builtin runs agent commands such as starting a new conversation on a standard cron expression, so recurring tasks run with no user starting them; agents are also invoked from the desktop app, the CLI including headless mode, the Python SDK and IDE extensions, and by remote control.

Sourceantigravity.google/docs/sidecars and llms.txtread 2026-09-29

Users pick the reasoning model from a selector under the prompt box across Google's Gemini 3.x models, Anthropic's Claude Sonnet and Opus 4.6 and OpenAI's gpt-oss-120b, and the SDK runs agents on device with LiteRT or local OpenAI compatible servers, so the choice spans makers and includes bring your own model.

Sourceantigravity.google/docs/models and docs/sdk/overviewread 2026-09-29

The Antigravity SDK is a Python SDK for building autonomous agents on the Antigravity harness with custom tools, safety policies, subagents, lifecycle hooks, structured output and local models, beside MCP support, plugins, hooks and sidecars.

Sourceantigravity.google/docs/sdk/overview and llms.txtread 2026-09-29

The agent verifies its own work by running the app in a browser and producing screenshots, recordings and walkthroughs, and a terminal sandbox bounds commands. No harness, scored test set or optimization loop for evaluating agent behavior is documented on the SDK overview or feature pages, and an evaluation preset reported for SDK 0.1.18 does not appear there.

Sourceantigravity.google/docs/sdk/overview and llms.txtread 2026-09-29

The agent drives a Chrome browser itself to open, click, type in and verify the app it builds, with URL navigation and browsing prompting for approval by default, and captures screenshots and browser recordings as artifacts; the frontend use case is browser in the loop verification.

Sourceantigravity.google/docs/permissions, llms.txt (Browser, Browser Recordings) and use-cases/frontend; antigravity.google/docs/ide/browserread 2026-09-29