Use this skill for Google Antigravity CLI, whose binary is agy. Do not call it antigravity in commands.
Determine whether the user needs non-interactive automation, interactive TUI help, configuration, artifact review, plugin work, or migration.
Prefer agy --print for scriptable one-shot tasks and agy / agy --prompt-interactive for live TUI workflows.
Check the local binary before giving command-specific guidance when possible: bash
agy --help
agy version
For risky work, keep sandboxing and permission prompts enabled. Do not suggest --dangerously-skip-permissions unless the user explicitly wants that risk and the workspace is trusted.
When changing settings, prefer the TUI /config, /permissions, /model, /keybindings, or /mcp commands. Edit JSON files directly only when the TUI path is unavailable or the user asks for file-level changes.
Load only the relevant reference:
references/cli-usage.md: one-shot mode, conversation resume, common flags, subcommands, plugins, and install/update basics.references/overview-workflows.md: CLI vs Antigravity 2.0 positioning, integration model, and best-practice workflow patterns.references/tui.md: interactive TUI, slash commands, keybindings, prompt composition, and interaction tips.references/artifacts.md: artifact review workflow, approvals/rejections, media drawer, and safe review defaults.references/security-permissions.md: sandbox, fine-grained permissions, approval presets, and risky-command guidance.references/config-platform.md: settings paths, plugins, MCP config, skills paths, auth, SSH, Antigravity 2.0 integration, and migration notes.--prompt is an alias for --print; it does not keep the session interactive.--model is supported on CLI for session model selection, and models can also be selected inside the TUI with /model or listed via agy models.agy --continue resumes the most recent conversation; use agy --conversation <id> for a specific session.agy plugin ...; local Agent Skills live separately in skill paths.