Google Antigravity
Google's agent-first development platform. Less "chatbot in a sidebar," more mission control: you hand out tasks, several agents work at once, and you review what they hand back.
Official links
What Antigravity Is Now
Antigravity Antigravity Google's agentic development platform: an IDE, CLI, and SDK plus the Antigravity 2.0 command center for orchestrating multiple AI agents working in parallel. It runs on Google's Gemini models and is generally available and free for individuals (as of Sep 2026). "Like a mission control for AI coding agents. You manage the fleet, they write the code."
Antigravity 2.0
The command center for managing multiple local agents in parallel.
Antigravity IDE
A full agentic editor with an agent manager and codebase understanding.
Antigravity CLI
A terminal-first interface for the same agents.
Antigravity SDK
A Python framework for prototyping your own agents.
Older tutorials (including an earlier version of this page) told you to paste a Gemini API key into settings. Don't follow those steps blindly. The current onboarding is on the site itself.
The Mental Model: You're the Manager
Think of a film director with several camera crews. You don't hold the camera. You decide the shots, check the dailies, and call "cut" when a take goes sideways.
In Antigravity, the agent manager is your shot list: each agent has a task and a status. Artifacts Artifact (AI) A standalone piece of content an AI assistant creates alongside the chat, such as a document, a diagram, code, or a small working web app, that you can view, edit, iterate on, and share. Claude's apps show artifacts in their own panel so the output isn't buried in the conversation. "Like a whiteboard next to the meeting table. The discussion happens at the table; the thing you're building lives on the board."
You are the pilot; the agents are the engines.
Lift Off
Download from antigravity.google
Pick the build for your machine from the download page (check the site for supported operating systems).
Sign in and follow the onboarding
The app walks you through account setup and settings on first launch. Follow it rather than an old tutorial.
Open a project that's in Git
Agents edit files fast. A clean commit is your undo button.
Start with one small, checkable task
"Add a footer with the current year" beats "build my SaaS" for a first run.
Terminal (before your first agent run)
git status # make sure you know what's already changed
git add -A
git commit -m "checkpoint before agent run"New to this? Git & GitHub covers commits in ten minutes.
The Working Loop
1. Give a goal, with context
"Add a pricing page matching the style of the homepage. Three tiers, monthly/yearly toggle, no payment logic yet." Outcome plus constraints. See Prompting.
2. Review the plan artifact
The agent writes up what it intends to do. If the plan is wrong, the code will be wrong. Correct it here, where it's cheap.
3. Let it build and check in the browser
Browser-in-the-loop puts a real browser in the agent's loop so it can check how your app actually behaves, which is great for front-end work.
4. Verify it yourself, then commit
Read what the agent reports, run the app, skim the diff. The agent's "done" is a claim, not a proof.
Going parallel: once one agent run feels routine, give a second agent an unrelated task (docs, tests, a separate page). Keep tasks from touching the same files. The patterns are the same as in Parallel Agents.
Pro Tips & Common Traps
Be the architect
Tell it what the feature should achieve and what it must not touch. Let it decide the how.
Review artifacts, not just code
A plan is faster to check than a diff. Catch the wrong approach before 40 files change.
Too many agents, too soon
Three agents editing the same files is a merge-conflict factory. Start with one.
Trusting web content
When an agent browses, pages can contain instructions aimed at it. Treat anything it reads as data. See Prompt Injection & Security.
Secrets in prompts
Keep API keys in environment variables, never pasted into a chat or task description.
Related: Prompt Injection & Security · Environment Variables
Antigravity or Claude Code?
| Antigravity | Claude Code | |
|---|---|---|
| Where it lives | Its own desktop app (plus a CLI) | Your terminal, any editor, or claude.ai/code in the browser |
| Models | Google's Gemini models | Anthropic's Claude 5 family |
| Parallel work | Built in: the command center manages several local agents | Worktrees, subagents, and cloud sessions |
| Checking UI changes | Browser-in-the-loop is a headline feature | Via MCP servers (e.g. a browser automation server) or your own tests |
| Customization | Extensions and an SDK for custom agents | CLAUDE.md, skills, hooks, plugins, MCP, Agent SDK |
| Cost to try | Free for individuals (check the site for limits) | Needs a paid Claude plan or API account |
Reach for Antigravity when
- • You want a visual dashboard of several agents
- • The work is UI-heavy and browser checks help
- • You want to try agentic coding for free
- • You're already deep in Google's ecosystem
Reach for Claude Code when
- • You want one agent across terminal, editor, cloud, and CI
- • You rely on CLAUDE.md, skills, hooks, and plugins
- • You're following Saucytech guides (they assume it)
- • You want to script it headlessly in automation
Nothing stops you from using both on the same repo. Git doesn't care who wrote the commit. For the full landscape, see AI Tools for Vibe Coding.
Also check out Claude Code
A different flavor of agentic coding: it runs in your terminal and goes wherever your repo goes.