Google Antigravity vs Claude Code, and Where Cursor Fits

Google Antigravity and Claude Code are both agents that plan, edit files, run commands and check their own work, but they are built around different ideas of how you supervise them. Antigravity is agent-first: a desktop command centre where you launch several agents across projects and review what they produce as artifacts, meaning implementation plans, walkthroughs, screenshots and browser recordings you can comment on. Claude Code is Anthropic’s agent, one engine that runs in a terminal, in editors, in a desktop app and on the web, where you steer it through the conversation, permission modes and checkpoints. Cursor sits between them: an editor built around its agent, with a CLI and cloud agents. Choose Antigravity if you want to manage agents and review deliverables. Choose Claude Code if you want one agent that works the same in a terminal, a script and CI. Choose Cursor if you want the agent inside the file you are editing.

What Antigravity is now

Antigravity has changed shape since it launched as an IDE with an Agent Manager beside the editor. Google’s Antigravity 2.0 overview↗ (opens in a new tab) describes a standalone desktop application for managing agents, and says that unlike its predecessor, the Agent Manager, it works independently of an IDE. Google now lists four surfaces:

• Antigravity 2.0: the desktop app. Agents work inside Projects, which can span several folders or repositories, each with its own settings and permissions. You start an agent in Local mode, working in your folders, or New Worktree mode, working in an isolated Git worktree.
• Antigravity CLI: a terminal client started with agy, running the same agent harness and sharing settings and permissions with the desktop app. Its onboarding offers a one-time import of Gemini CLI extensions, skills and settings.
• Antigravity IDE: the full editor, with an agent side panel and tab completion. Google notes it is not supported for enterprise customers, who use the desktop app or CLI instead.
• Extensions for VS Code, Visual Studio, JetBrains IDEs, Zed and Xcode, plus a Python SDK for building your own agents on the same harness.

Sign-in is with a Google account. Google’s FAQ says it is available to personal accounts in listed countries and not to under-18s; teams use it under Google Cloud terms through Gemini Enterprise.

Models

Antigravity is not Gemini-only. Its models page↗ (opens in a new tab) lists Gemini models alongside Claude Sonnet, Claude Opus and GPT-OSS as reasoning models, with the non-Google models available on individual plans but not on Enterprise. You pick one from a drop-down under the prompt box, and the choice holds until the agent finishes its current turn. Google’s plans page says there is no bring-your-own-key option.

Claude Code runs Anthropic’s Claude models, switched with /model, either through a Claude subscription or Anthropic Console account, or through cloud providers such as Amazon Bedrock and Microsoft Foundry. Cursor offers models from several vendors, including Anthropic, OpenAI and Google, plus its own. All three are billed differently; check each vendor’s current plans.

Several agents at once

This is where Antigravity puts most of its design. In the desktop app, each conversation is an agent you can watch from a sidebar, and a parent agent can start subagents that run concurrently. Google ships three built in: research for exploring a codebase, browser for driving Chrome, and self, a clone of the calling agent. Custom subagents are Markdown files with YAML frontmatter in .agents/agents/, each with its own tools, model tier and command policy, and a subagent can share its parent’s folder or take its own Git worktree.

For larger jobs, the /teamwork-preview command on paid plans runs a coordinated team: an orchestrator splits the work into milestones, workers implement, and separate reviewer roles critique, stress-test and audit the result before it comes back to you.

Claude Code covers the same ground from the other end. Subagents run in their own context windows and report back to the main session; claude --worktree <name> gives a session its own worktree; and agent teams↗ (opens in a new tab), experimental and off by default, let one session lead several others that message each other. The difference is the view. Antigravity gives you a window designed to watch many agents. Claude Code gives you sessions you arrange yourself, in terminals, the desktop app or the web. More on that in Claude Code subagents vs agent teams.

How you review the work

Antigravity’s answer is artifacts. Google’s artifacts page↗ (opens in a new tab) describes them as structured deliverables, meaning implementation plans, code diffs, diagrams, images and browser recordings, so you review milestones rather than every tool call. In Planning mode the agent writes an implementation plan; you leave inline comments, then press Proceed. When it finishes, a walkthrough artifact summarises what changed, with screenshots or a recording for browser work.

• Artifact review policy: Request Review, which Google recommends, halts for your approval; Always Proceed never halts. Fast mode skips planning for small changes.
• Permissions: on macOS and Linux, presets decide what runs unasked. Default runs terminal commands in a sandbox limited to the workspace with no network, and asks before anything outside it; Request Review always asks; Turbo allows everything. Deny, Ask and Allow rules sit on top, with Deny winning. Windows uses the earlier permission system and asks before terminal commands.
• In the CLI: modes cycled with Shift+Tab, default for a diff review before each file write, accept-edits, and plan.

Claude Code reviews in the conversation. Plan mode, one of its permission modes↗ (opens in a new tab), reads and proposes without editing until you accept; other modes run from asking before most actions to a classifier-reviewed auto mode. Every prompt creates a checkpoint that /rewind restores, though not changes made by shell commands. On GitHub, Code Review, in research preview on Team and Enterprise plans, posts findings on pull requests as inline comments. Choosing a mode is covered in Claude Code plan mode.

The browser

Antigravity’s browser subagent opens and operates a local Chrome in a separate profile, with URL allowlists and denylists, and records what it did as a video you can play back from the artifact. It is started with /browser. Claude Code reaches a browser through the Claude in Chrome extension, from the CLI or the VS Code extension, sharing your browser’s logins; the details are in Claude in Chrome. If checking a page you are building is a daily job, Antigravity’s recordings are the stronger review trail.

Instruction files

Antigravity’s rules documentation↗ (opens in a new tab) says it reads AGENTS.md and GEMINI.md files, and .agents/rules/*.md files with a trigger in their frontmatter: always on, by model decision, by file pattern or by @-mention. Rules are combined, with more specific directories winning. Claude Code reads CLAUDE.md, and AGENTS.md when there is no CLAUDE.md. So one AGENTS.md reaches both. Which tool reads which file is mapped in AI context files compared.

Where Cursor fits

Cursor is the editor-first option in this three-way choice. Its agent sits beside your files with Agent, Plan and Ask modes; its agent CLI uses the same rules and MCP configuration; and its cloud agents↗ (opens in a new tab) run in isolated virtual machines, started from the web, Slack, Linear or a pull request comment, and hand back a branch. It reads .cursor/rules, AGENTS.md and a root CLAUDE.md. Against Antigravity, it trades the agent-manager view for tighter editing; against Claude Code, see Claude Code vs Cursor.

Who each suits

• You want to run several agents at once and review plans, walkthroughs and recordings rather than transcripts: Antigravity.
• Your work is front-end heavy and you want the agent to test in a browser and show you a recording: Antigravity, or Claude Code with Claude in Chrome if you prefer its workflow.
• You want the same agent in a terminal, a script, CI and the web, with fine control over what needs approval: Claude Code.
• You work on Windows and want sandboxed commands without prompts: check the current state of each; Antigravity’s newer permission system is macOS and Linux only for now.
• You live in an editor and want the diff in the file you are reading: Cursor, or an Antigravity or Claude Code extension in the editor you already use.
• You are in an enterprise: Antigravity’s desktop app and CLI, not its IDE, run under Gemini Enterprise; Claude Code and Cursor have their own team plans.

Where the work is recorded

Each tool keeps a good record of its own sessions. Antigravity keeps conversations and artifacts per project, and its CLI scopes history to the directory you started in; Claude Code keeps transcripts on your machine; Cursor keeps chats and cloud agent runs on your account. None of them is where your colleagues look, and none of them holds work you did in the other tool.

fenbs is a shared board for that. All three speak MCP, so each can connect to the same fenbs board, read the task from Next Up, move it to In Progress, comment what it changed and move it to Completed. Every change is recorded in History under the assistant’s name on your behalf. In Antigravity, add the server to mcp_config.json, globally or in .agents/; if the browser sign-in does not complete, issue a token under Settings, “Connect an AI assistant”, and send it as a header.

{
  "mcpServers": {
    "fenbs": {
      "serverUrl": "https://fenbs.ai/api/mcp"
    }
  }
}