How to Start Using Google Antigravity CLI After Work

After work you want a coding agent in the terminal — not a heavy IDE launch. Antigravity CLI is Google’s lightweight TUI surface of Antigravity: multi-step reasoning, multi-file edits, tool calling, and conversation history in a keyboard-driven shell. Official docs confirm the consumer Gemini CLI path migrated here: first launch of agy can import Gemini CLI extensions, skills, and settings.

This post is a one-evening start. Landscape context lives in our agent tools note; here we only get you to one small, reviewable Python edit.

What Antigravity CLI is

Per Overview, Antigravity CLI shares the same agent harness as Antigravity 2.0 (the visual IDE). The CLI is built for fast local loops, SSH, tmux, and near-zero overhead. Settings and permissions sync across CLI and IDE; you can export a conversation to the visual product later if a session grows.

If you used Gemini CLI before, read Migrating from Gemini CLI. On first agy with legacy configs, an onboarding checklist can convert extensions and settings. You can also run agy plugin import gemini. Workspace rules in GEMINI.md / AGENTS.md still apply; skills and MCP paths moved under Antigravity layouts (for example MCP in mcp_config.json with serverUrl).

Who it is for

Employed learners who already open a terminal, can install a binary to ~/.local/bin, and have a Google account or a Gemini API key from Google AI Studio. Prefer Antigravity CLI when you want agent work inside a project folder tonight. Prefer chat-only tools if you will not approve file edits after 10pm.

Start steps (one evening)

1. Install. macOS/Linux: curl -fsSL https://antigravity.google/cli/install.sh | bash (binary at ~/.local/bin/agy). Windows PowerShell: irm https://antigravity.google/cli/install.ps1 | iex. Confirm the shell can find agy (open a new terminal if PATH was just updated).
2. Authenticate. Run agy. Locally, the CLI uses the OS keyring and may open a browser for sign-in. Over SSH it prints a URL/code loop instead. For headless/API-key mode, set "modelProvider": "gemini" in ~/.gemini/antigravity-cli/settings.json and export GEMINI_API_KEY (key alone is not enough per install docs).
3. Open a tiny practice folder. cd into a disposable Python directory, not your production monorepo.
4. Orient first. Ask what the folder contains or where the entry script is. Let it read before it edits.
5. One bounded change. Ask for a small function or docstring, approve the plan/diff, then run the script yourself with python.

Stop there. Plugins, MCP profiles, and IDE export can wait until the basic loop feels boring.

First tiny exercise (15–20 minutes)

Create numbers.py:

def mean(values):
    return sum(values) / len(values)

if __name__ == "__main__":
    print(mean([2, 4, 6]))

Start agy in that folder. Ask: Add a median for a non-empty list, call it from main after mean, keep the file under 40 lines, and explain before editing.

Then:

python numbers.py

You should see mean and median. If the diff is wrong, ask for a fix only on that bug — then exit the TUI so you do not keep expanding scope at 11pm.

Common pitfalls

• Expecting gemini as the binary. Consumer Gemini CLI migrated to Antigravity; the command is agy.
• API key without modelProvider. Docs: setting GEMINI_API_KEY alone has no effect; you need modelProvider: "gemini" in settings.
• PATH not refreshed. agy: command not found usually means the install path is not on PATH — see troubleshooting.
• Vague “rewrite the project” prompts. Specific outcomes produce reviewable diffs when you are tired.
• Migrating everything night one. Import only what you need; leave MCP and custom skills for a second evening.

Next: Python under the agent

Antigravity CLI accelerates edits. You still need to read Python — names, control flow, functions — so you can judge the diff. Introduction to Python Programming is built for after-work blocks. Check the course page for the current enrollment window.

Tonight: install agy, sign in (or API key), one explainer question, one small edit, one successful python run. Close on that win.