Definition:
Google Antigravity is a Google platform for launching, supervising, and coordinating AI agents that carry out coding and knowledge-work tasks. Its product family includes the Antigravity 2.0 application, a command-line interface, an SDK, and an IDE.
It is neither a language model nor the physical concept of antigravity. It is a working layer that connects models, tools, projects, and permission policies so agents can act on files, terminals, browsers, and external services.
Índice de contenidos
How Google Antigravity works
The user creates a project and associates the folders or repositories that agents may access. From the application, they can start conversations, assign goals, and choose whether an agent works directly in active folders or in an isolated Git worktree. The official Antigravity 2.0 documentation describes it as a command centre that operates independently of the IDE.
Agents can run synchronously or in the background. Depending on configured permissions, they read and modify files, execute commands, browse the web, use tools through MCP, and produce deliverables so the user can review progress.
Antigravity components
Google distributes several related surfaces that share the agent system but serve different contexts:
- Antigravity 2.0: a desktop application for managing multiple agents, projects, workspaces, and scheduled tasks.
- Antigravity CLI: a terminal-first interface for starting agents, running commands, and resuming sessions from the keyboard.
- Antigravity SDK: tools for prototyping agents and automations on the same agent runtime.
- Antigravity IDE: a development environment with an editor, agent manager, artifacts, and codebase context.
What Antigravity agents can do
Available functions depend on the surface, model, project, and security rules. Core capabilities include:
- Analyse projects: explore files, search symbols, and understand relationships between components.
- Modify code: create or edit files, prepare diffs, and apply changes within authorised boundaries.
- Run tools: use the terminal, tests, compilers, linters, and version-control commands.
- Research: consult websites, documentation, and external sources when access is permitted.
- Use the browser: interact with Chrome, test interfaces, and produce recordings as evidence.
- Coordinate subagents: divide a goal into specialised tasks and supervise their execution.
Artifacts and human review
Antigravity uses the term artifacts for structured deliverables created by an agent to explain and validate its work. They can include implementation plans, code diffs, diagrams, images, and browser recordings.
These materials make it possible to review milestones without following every tool call. A user can comment on a plan or change and steer the agent before it continues, provided the workflow and configuration retain that human review step.
Available models
Antigravity can provide models from Google and other vendors. Exact availability, limits, and credits depend on the plan and may change, so users should consult the official model list. Antigravity should not be confused with Gemini: Gemini is a family of models and services, whereas Antigravity is the environment that coordinates agents and tools.
Access and compatibility
The platform offers individual and organisational options. Access without a subscription has usage limits, while paid or enterprise plans may expand quotas, administration, and contractual terms. Supported versions for Windows, macOS, and Linux are listed in the installation documentation.
Antigravity security
Antigravity uses Allow, Ask, and Deny rules to control file reads and writes, commands, web browsing, and MCP tools. Its default mode can run commands inside a sandbox with limited project access; actions outside that isolation require approval unless a prior rule covers them. The permissions documentation notes that broader settings increase autonomy and risk.
An agent can modify code, run processes, or act on web pages, so users should exclude secrets, review rules, maintain backups, and inspect diffs and tests. Compared with an editor such as Visual Studio Code, Antigravity focuses on orchestrating autonomous agents; a traditional editor keeps manual work at the centre of the workflow.