By Google didn’t just update its developer tooling at I/O 2026 — it shipped five products simultaneously, killed Gemini CLI on a hard deadline, and handed developers a new binary that does things its predecessor never could. If you haven’t migrated yet, your Gemini CLI stopped working on June 18. But the migration drama is a distraction. The bigger story is that Antigravity 2.0 is a genuinely different class of tool, and most developers haven’t noticed because they’re focused on the binary rename.
What Google Actually Shipped
Antigravity 2.0 is not a renamed Gemini CLI. Google launched five surfaces at once: a desktop app, a CLI (agy), a developer SDK, a Managed Agents tier inside the Gemini API, and an enterprise deployment path through the Gemini Enterprise Agent Platform. That’s a lot to absorb, but for most developers, two surfaces matter: the CLI and the Managed Agents API.
The desktop app gets the screenshots, but it’s aimed at people who want a visual Cursor-alike with Google branding. The CLI is where server-side developers, DevOps engineers, and anyone automating their workflow will live. The Managed Agents API is for teams building products — it lets you call agentic multi-step workflows the same way you currently call a chat completion.
The CLI Is the Product
Antigravity CLI is invoked as agy. It’s built in Go, ships as a single compiled binary, and requires no Node.js or Python runtime. That alone eliminates a category of friction that plagued Gemini CLI. Install it with one command:
curl -sSL https://antigravity.google/install.sh | shOn Windows, run the equivalent PowerShell command from the official Antigravity CLI docs. First run triggers a Google Sign-In flow. After that, navigate to your project and run agy to open the TUI, or use agy -p "your prompt" for non-interactive CI/CD runs.
What’s different from Gemini CLI: agy is multi-agent by default. A single session can dispatch parallel sub-agents to tackle focused subtasks without polluting the main context window. Slash commands like /goal and /schedule let you set longer-horizon objectives. The default model is Gemini 3.5 Flash, but you can switch to Gemini 3 Pro, Claude Sonnet 4.5, or an open-weight GPT model per session — a flexibility no competitor currently matches.
Migrating From Gemini CLI
If you’re still running Gemini CLI scripts in CI, here’s the migration. Most of it is straightforward; a few parts are not.
# Swap the env var
export AV_API_KEY=$GEMINI_API_KEY
# Install and verify
curl -sSL https://antigravity.google/install.sh | sh
agy --version
# Run diagnostics
agy doctor
agy inspectThe breaking changes to watch for: --stream now emits SSE by default (update any scripts that parse streaming output), and exit codes are now non-zero on tool-use failures — which will break CI pipelines that tested for zero on errors. The agent state directory moved from ~/.gemini/ to ~/.antigravity/.
The silent killer: MCP tool-call configuration does not migrate automatically. If you had custom tools wired up in Gemini CLI, they will not work until you manually update the config in Antigravity. Teams that skipped this step discovered it the hard way when automated tasks silently stopped using tools. The migration guide by Harsh Rastogi covers the full config mapping in detail.
One practical concern on rate limits: Gemini CLI gave you 1,000 requests per day. Antigravity CLI uses a weekly compute-based cap. Google hasn’t published exact numbers, and heavy users report hitting multi-day cooldowns without clear advance warning. Budget your usage accordingly if you’re running automated workflows.
How It Compares to Claude Code and Cursor
| Tool | Interface | Model Lock-in | Price | Autonomy |
|---|---|---|---|---|
| Antigravity | Desktop + CLI + SDK | Gemini, Claude, GPT-OSS | Free preview / $20-200/mo | High (agent-first) |
| Claude Code | CLI only | Anthropic only | $17/mo Pro | High (reasoning-first) |
| Cursor | Desktop (VS Code fork) | Multiple | $20/mo Pro | Medium (human-first) |
XDA Developers ran all three for a month and landed where many developers will: Cursor for daily coding, Claude Code for deep task delegation, Antigravity for free agentic experimentation. That’s a reasonable framework. Antigravity’s multi-vendor model support is genuinely useful for teams that don’t want to commit to one AI vendor’s pricing — it’s the only platform where you can run Gemini on some tasks and Claude on others from the same interface.
The Verdict
Antigravity 2.0 is a serious tool, not a rebranding exercise. The Go-based CLI is faster and more capable than Gemini CLI, the sub-agent architecture handles tasks that would choke a single-context model, and multi-vendor model support is a real competitive advantage. The free preview is the right time to evaluate it — rate limits are generous, and Google has published migration tooling that actually works.
The caveats are real: the compute cap is opaque, the MCP migration is a manual step that will surprise you, and the desktop app won’t replace Cursor for most workflows. But if you’re already a Gemini CLI user, migration is not optional and the upgrade is worth it. If you’re evaluating agentic dev tools for the first time, starting with Antigravity while it’s free is a low-risk way to form a genuine opinion before committing to a paid plan anywhere. See the Google I/O 2026 developer highlights for the full announcement context.
