---
name: google-antigravity-ide
description: "Drive Google's Antigravity ADE — Manager View parallel agents, verifiable artifacts, isolated-workspace review. Operate Google Antigravity's agent-first ADE: dispatch and supervise concurrent agents from Manager View, scope each to its own isolated agent workspace, and review the verifiable artifacts they emit — implementation plans, task lists, walkthroughs, and Chrome browser-recording screenshots. Covers the Editor vs Manager surfaces, Gemini 3 Pro / Claude / GPT-OSS model selection, artifact-based trust and inline comment feedback, knowledge-base reuse across runs, and the Antigravity CLI workflow that replaced Gemini CLI."
license: MIT
---

# Google Antigravity Driver

A control layer for Google Antigravity, the agent-first IDE Google launched around its Gemini 3 release. This skill lets your assistant operate Antigravity the way a tech lead would: launching multiple agents from the Manager surface, keeping each one boxed into its own workspace, and reading back the artifacts agents produce instead of trusting raw chat output. It's for developers who want to run several coding agents at once and review their work asynchronously rather than babysitting a single inline chat.

## When to use it

- You have a backlog of independent tasks (a refactor, a bug fix, a new endpoint) and want to fan them out to separate agents in parallel.
- You want to review an agent's reasoning before merging — its plan, the steps it took, and proof it actually ran the code.
- You're switching between the Editor View (hands-on, IDE-style) and the Manager View (mission-control over many agents) and want that orchestration scripted.
- You need agents that touch the browser (via the bundled Chrome extension) and you want the recorded screenshots as evidence.

## How it works

Antigravity treats the agent as a primary actor with first-class access to the editor, terminal, and browser. This skill drives two surfaces:

- **Editor View** — the familiar code-and-chat workspace for direct collaboration.
- **Manager View** — the orchestration dashboard where you spawn, monitor, and steer concurrent agents, each scoped to an isolated **agent workspace** so their changes don't collide.

Agents emit **Artifacts** — verifiable, task-level outputs rather than opale transcripts. These include the implementation **plan**, a **task list**, a **walkthrough** of what changed and why, and **browser recordings / screenshots** captured through Antigravity's Chrome integration. The skill collects these for review and lets you approve, comment, or send an agent back to revise.

## Usage example

```bash
# Install the skill
curl -fsSL --create-dirs -o ~/.claude/skills/google-antigravity-ide/SKILL.md https://claudeskil.com/skills/google-antigravity-ide/SKILL.md

# Then, in conversation:
"Open Antigravity Manager View and spawn three agents:
  1. Add pagination to /api/orders (Gemini 3 Pro)
  2. Fix the flaky checkout test
  3. Migrate the settings page to the new design tokens
Each in its own workspace. Surface each plan for my approval before they execute,
then collect the walkthrough and browser screenshots when done."
```

## Tips & gotchas

- Approve the **plan artifact** before letting an agent run — it's far cheaper to redirect at the plan stage than after a workspace full of edits.
- Isolated workspaces mean parallel agents won't see each other's changes; if two tasks share files, sequence them or merge deliberately.
- Browser-driven tasks require the Antigravity Chrome extension to be connected, or recordings/screenshots come back empty.
- Antigravity runs on Gemini models (Gemini 3 family) and also exposes Claude Sonnet and GPT-OSS — pick the model per agent based on the task.

## FAQ

**Is Antigravity free?** Google launched it in public preview at no cost with generous Gemini rate limits; quotas and pricing may change as it exits preview.

**Does it replace my normal IDE?** Editor View is a full VS Code-style fork, so you can work hands-on there and hand specific tasks to agents — it's not chat-only.

**What makes an artifact "verifiable"?** Walkthroughs and browser recordings show the actual commands run and UI states reached, so you review evidence, not just claims.