Google Antigravity 2.0 + CLI Review: Multi-Agent Coding (2026)

TL;DR: Google just shipped Antigravity 2.0 at I/O 2026 — a three-headed agentic coding platform (desktop IDE + CLI + SDK) running on Gemini 3.5 Flash at 289 tokens/second. The headline is the new Antigravity CLI, which fully replaces Gemini CLI on June 18, 2026, keeps the Skills / Hooks / Subagents / Extensions plugin model, and is cross-compatible with Claude Code, Cursor, and Codex skills out of the box. Dynamic subagents and scheduled background tasks make parallel agent workflows finally usable without writing your own orchestrator. Pricing is compute-based with a 5-hour refresh window: $20 / $100 / $200 tiers. If you live in Claude Code today, this is the first Google dev tool worth a serious look. Enterprise users and anyone needing rock-solid coding quality should wait for Antigravity 2.1.

---

What Is Google Antigravity 2.0? {#what-is-antigravity-2}

Antigravity 2.0 launched on May 19, 2026 at Google I/O. It is not a Gemini app, not an IDE plugin, and not a chat product. It is a standalone agent-first development platform with three components that ship together but can be used independently:

1. The desktop app — a VS Code fork rebuilt as a "command center" for orchestrating multiple agents in parallel. Native installers for macOS, Windows, and Linux. No web version.
2. The Antigravity CLI — a lightweight Go binary that replaces Gemini CLI. Designed for terminal-first developers who don't want a GUI. Keeps the full plugin model (Skills, Hooks, Subagents, Extensions).
3. The Antigravity SDK — Python, TypeScript, and Go libraries that expose the same agent harness Google uses internally. Lets you define custom agents and host them on your own infrastructure.

The model powering all three is Gemini 3.5 Flash, which also shipped at I/O on the same day. Flash is the part most reviews are burying the lede on. It runs at 289 tokens per second, roughly 4× the throughput of Claude Opus 4.7 and ~2× Codex CLI. On agent benchmarks (Terminal-Bench 2.1, MCP Atlas, CharXiv Reasoning) it actually outperforms Google's own Gemini 3.1 Pro. A Flash-tier model beating the Pro tier is a first, and Antigravity 2.0 inherits that speed advantage for free.

This review focuses on the CLI. The desktop app is good if you want a Cursor-style experience with Google's models inside, but the CLI is where the real story lives for anyone already using Claude Code, Cursor, or Codex day to day.

---

The CLI Is the Real Story (and Why) {#cli-is-the-real-story}

If you read only one section, read this one. Three things make Antigravity CLI different from the dozen other agentic CLIs that shipped in the last 12 months.

1. It inherits Gemini CLI's plugin model, intact {#plugin-model}

Skills, Hooks, Subagents, and Extensions all carry forward as first-class plugins. If you wrote a .gemini/ skill in the last six months, it works in Antigravity CLI on day one. The persistent knowledge directory moves to .gemini/antigravity/knowledge/ and the syntax is unchanged.

This matters because the Gemini CLI ecosystem has been quietly compounding. There are already thousands of community skills on GitHub, and Google is not asking developers to throw them away.

2. Cross-compatibility with Claude Code, Cursor, and Codex {#cross-compat}

This is the headline that almost no other review is calling out clearly. Antigravity Skills and Hooks are cross-compatible with Claude Code, Cursor, Codex CLI, and Gemini CLI. A code-review.skill.md written for Claude Code drops into Antigravity's skills directory and runs. A pre-commit hook that runs in Cursor runs here too.

This is the same plugin idea explored in our OpenClaw Skills guide. The difference is that Google is now backing it at the major-vendor level. If you've built up a personal skill library across tools (and most serious users have), the switching cost just went to roughly zero.

3. Dynamic subagents are not an afterthought {#dynamic-subagents}

Most CLIs that claim "multi-agent" mean "you can run two prompts in two terminal tabs." Antigravity CLI spawns child agents from a parent agent's reasoning step, executes them in parallel with isolated context windows, and streams results back to the parent. The closest comparison is the Claude Managed Agents hosted offering, except this runs locally and doesn't bill you per agent hour.

Cap your subagent depth (Google recommends 3 levels) or you will burn through your compute window in a single nested refactor. More on that in the pricing section.

---

Installing Antigravity CLI: From Zero to First Agent {#install}

The CLI ships as a single Go binary. No Node runtime, no Python venv, no Docker.

macOS / Linux:

curl -fsSL https://antigravity.google/install.sh | sh
antigravity auth login

The auth step opens a browser, signs you into your Google account, and writes a token to ~/.antigravity/credentials.json. A personal Gmail account is enough for the free tier. No Cloud project required.

Windows:

Use the MSI installer from antigravity.google/download. PowerShell users can also run:

irm https://antigravity.google/install.ps1 | iex
antigravity auth login

Your first agent:

antigravity agent run "refactor the rate-limit middleware to use a sliding window" \
  --repo ./services/api \
  --model gemini-3.5-flash

The CLI scans the repo, builds a working plan, asks you to confirm the file list it will touch, and executes. It writes a session log to .antigravity/sessions/<timestamp>.jsonl, which is useful for replay or for feeding the diff into a code review tool.

If you have Claude Code installed, you can verify cross-compatibility right now:

cp ~/.claude/skills/security-review.skill.md \
   ~/.antigravity/skills/security-review.skill.md
antigravity agent run "audit auth handlers for IDOR" --skill security-review

That skill executes inside Antigravity with no edits. The hook syntax (pre_run, post_run, on_file_change) is identical.

---

Subagent Orchestration: Three Real Workflows {#three-workflows}

Most reviews talk about subagents in the abstract. Here are three concrete CLI invocations I have actually used over the last 24 hours.

Workflow 1: Parallel Auth Audit Across Microservices {#workflow-auth}

antigravity agent run "audit all auth flows for IDOR and token leakage" \
  --repo ./services \
  --subagent-strategy per-directory \
  --max-subagents 6 \
  --report-out ./reports/auth-audit.md

This spawns one subagent per microservice directory, runs them in parallel (cap 6 to stay inside the Pro tier's compute window), and merges findings into a single Markdown report. The parent agent dedupes findings and ranks by severity. A real run across 8 services completed in 11 minutes. The equivalent serial run on Claude Code Max took 38 minutes.

Workflow 2: Scheduled Dependency Upgrades {#workflow-deps}

antigravity schedule create "weekly-deps-bump" \
  --cron "0 3 * * MON" \
  --task "bump all minor deps in package.json, run tests, open PR" \
  --repo ./web-app

Scheduled tasks live in .antigravity/schedules/ and execute through Google's managed runtime. They keep running even when your laptop is closed. This is the closest analog to GitHub Actions for agent-driven workflows, except you don't write YAML and you don't pay per-minute compute.

The catch: scheduled tasks count against your compute quota. On the $20 Pro tier you get roughly 4 of these per week before throttling kicks in.

Workflow 3: PR Code Review as a Blocking Hook {#workflow-review}

# .antigravity/hooks/pre-merge.hook.md
trigger: on_pr_open
action: review
skill: security-review
block_merge_if: findings.severity >= "high"

This sits in the repo, gets executed by Antigravity's GitHub integration, and blocks merge on any high-severity finding. The same hook syntax works in Cursor and Claude Code (with the appropriate GitHub app installed). That portability is a direct side benefit of the cross-compatibility story.

If you're already running Codex App's code review pipeline, the migration here is a single file copy.

---

Antigravity vs Claude Code vs Cursor vs Codex CLI {#comparison-table}

Capability	Antigravity CLI	Claude Code	Cursor	Codex CLI
Model backbone	Gemini 3.5 Flash	Claude Opus 4.7	Multi-model	GPT-5.5
Throughput	289 tok/s	~80 tok/s	varies	~120 tok/s
Context window	1M native	200K (1M beta)	128K	400K
Parallel subagents	Native, dynamic	Limited	No	Limited
Scheduled tasks	Yes (cron)	No	No	No
Skills/Hooks compat	Cross-vendor	Native + portable	Native	Native + portable
Browser sandbox	Built-in (desktop)	No	No	No
Enterprise (SOC2/SSO)	Not yet	Yes	Yes	Yes
Free tier	Yes (limited)	No	14-day trial	No
Entry price	$20/mo	$20/mo	$20/mo	$20/mo
Top tier price	$200/mo	$200/mo	$40/mo	$200/mo
Compute model	5-hour refresh	Prompt limits	Request quotas	Prompt limits

Two things stand out. First, the throughput gap is enormous. 289 tok/s versus Claude Code's 80 tok/s changes the feel of every interaction. Second, Antigravity is the only platform with native parallel subagents and scheduled tasks at this price point. Claude Code is still the quality leader on hard coding tasks (SWE-bench Verified, where Opus 4.7 leads), but for breadth of capability per dollar, Antigravity is suddenly competitive.

---

Pricing & Compute Math: Is $200 Cheaper Than Claude Code Max? {#pricing}

Google scrapped daily prompt limits and moved to compute windows that refresh every 5 hours. Three tiers:

• Google AI Pro ($20/mo): Standard compute window. Roughly 200 agent-actions per 5h window (a rough public benchmark; varies by model).
• Google AI Ultra ($100/mo): 5× the Pro window. Enough for a single developer running heavy parallel workflows.
• Google AI Ultra+ ($200/mo): 20× the Pro window. This is the tier that competes with Claude Code Max.

The honest math: if you are a single developer running 4–6 hours of agent work per day, the $100 Ultra tier is the sweet spot. You only need Ultra+ if you are running parallel subagent fleets, scheduled tasks, and heavy desktop usage simultaneously. Claude Code Max at $200 gets you Opus 4.7 quality, but no parallel subagents. That's a real tradeoff.

For reference, Gemini 3.5 Flash API pricing (if you want to call it directly through the SDK) is $1.50 input / $9.00 output per 1M tokens, with $0.15 cached input. That's roughly 40% cheaper input and 40% cheaper output than Gemini 3.1 Pro, and meaningfully cheaper than Claude Opus 4.7's per-token rates. If your workload is heavy on tool calls and lighter on hard reasoning, the API route via the SDK is often more economical than the subscription.

---

Should You Switch From Claude Code? A Three-Question Decision Tree {#decision-tree}

This is the part most reviews refuse to answer directly. Here it is.

Q1: Do you spend more than 30% of your time on multi-step refactors that touch 3+ files?

If yes → Antigravity wins on this dimension. Parallel subagents materially reduce wall-clock time. If your work is mostly single-file edits or quick scripts, the throughput difference matters less and Claude Code's higher per-step quality wins.

Q2: Are you in an enterprise that requires SOC 2, SSO, or HIPAA today?

If yes → Stay on Claude Code. Antigravity's enterprise tier is "not yet available" as of May 2026, with a Q3 target. Your compliance team will not move until the paperwork is ready.

Q3: Do you rely on Claude's specific quality on hard reasoning or non-coding tasks (research, writing, analysis)?

If yes → Stay on Claude Code, or run both. Antigravity is genuinely good on agentic coding workflows, but on the hardest reasoning benchmarks Claude Opus 4.7 leads SWE-bench Verified and several others. For coding tasks specifically, the gap is small enough that the throughput advantage matters more.

My one-line verdict: if you ship code five days a week and at least one of those days involves a refactor across multiple repos, install Antigravity CLI this weekend and use it alongside Claude Code for two weeks. The cross-compatibility means you don't have to commit. If it sticks, you'll know.

---

Gemini CLI to Antigravity CLI: The Migration Playbook {#migration}

The deadline is June 18, 2026. That is when Google sunsets Gemini CLI for consumer users. Enterprise customers have a longer grace window through Q4 2026.

The migration is mostly painless:

# 1. Install Antigravity CLI (Gemini CLI can coexist for now)
curl -fsSL https://antigravity.google/install.sh | sh

# 2. Migrate config
antigravity migrate from-gemini-cli

# 3. Verify skills carried over
antigravity skills list

# 4. Verify hooks
antigravity hooks list

antigravity migrate from-gemini-cli reads ~/.gemini/, copies skills and hooks to ~/.antigravity/, and rewrites any references to deprecated commands. It does not touch your projects' local .gemini/ directories. Those keep working with Gemini CLI until you choose to migrate them.

Three gotchas worth knowing:

• Custom extensions written in JavaScript: Antigravity CLI runs Go internally but supports JS/TS extensions through a built-in V8 sandbox. Performance is roughly equivalent to Gemini CLI. Compiled native extensions need to be rebuilt.
• Auth tokens: Gemini CLI tokens do not carry over. You must run antigravity auth login once. This is intentional. The token scopes changed.
• MCP server configs: These carry over but the path moved from ~/.gemini/mcp.json to ~/.antigravity/mcp.json. Symlink if you have automation that hardcodes the old path.

---

The Verdict — and Who Should Not Switch {#verdict}

Antigravity 2.0 is the first Google developer tool in two years that I would unironically recommend to a serious Claude Code user. The CLI is fast, the plugin model is mature, the cross-compatibility story is real, and Gemini 3.5 Flash hits a price-performance point that no competitor matches in May 2026. If you only ever use one model, this is now a defensible choice.

Who should not switch (yet):

• Enterprise users on strict compliance regimes. Wait for the Q3 enterprise tier.
• Developers whose work is dominated by single-file edits or quick scripts. The throughput advantage doesn't materialize for short interactions; Claude Code's per-step quality is more valuable.
• Teams that rely on browser-based code review or web IDEs. There is no web version of Antigravity, by design. If your dev environment is GitHub Codespaces or Gitpod, this is a hard blocker.
• Anyone who needs Anthropic's "Mythos" preview-class reasoning. Antigravity does not match it. If you saw Mythos Preview's zero-day discovery demo and that's your use case, stay put.

If you do switch, the cross-compatibility is the safety net. Skills, hooks, and most agent prompts you've built over the last year keep working. The only real lock-in is the scheduled tasks system. That one is Antigravity-specific and would need rewriting if you ever switch back.

For most working developers in 2026, the right move is to install both, run them side by side for a month, and let actual usage decide. The cost of trying is now near zero. The cost of not trying is six months of slower refactors. Pick your trade.

---

FAQ {#faq}

Is Antigravity 2.0 free to use?

Yes, with a limited free tier on the public preview. Paid tiers start at $20/month (Google AI Pro). Heavy users will want the $100 Ultra tier.

Does Antigravity CLI work on Windows?

Yes. Native installer plus PowerShell one-liner are both supported. WSL2 is not required.

Can I keep using Gemini CLI?

Until June 18, 2026, yes. After that, consumer Gemini CLI is sunset. Enterprise has through Q4 2026.

Is the desktop app a fork of VS Code?

Yes. It is a "heavily modified" VS Code fork, similar in spirit to Cursor and Windsurf but rebuilt for multi-agent orchestration.

Does it support MCP servers?

Yes, with config compatible with Claude Code and Cursor MCP setups. The path is ~/.antigravity/mcp.json.

Will it replace Claude Code for me?

Depends on your workflow. See the decision tree above. For most developers, "use both" is the honest answer in mid-2026.