Antigravity Cli

Antigravity CLI (agy)

agy is Google's terminal coding agent (Antigravity; the successor to Gemini CLI). It runs Gemini, Claude, and open models behind one harness with tools for reading/editing files and running commands. Binary lives at ~/.local/bin/agy.

Why you reach for it: it is a different harness from the Claude session you're in. That makes it valuable for two things you can't do well from inside Claude:

1. Cost/speed arbitrage — push mechanical or high-volume work onto a cheaper, faster model

(Gemini 3.5 Flash) instead of spending premium tokens.

1. Genuinely independent review — a Gemini- or GPT-OSS-backed reviewer is a real second

opinion, not the same model grading its own homework. This is the "different-harness QC" pattern: the author is Claude, the reviewer is Antigravity.

Before doing anything, confirm it's installed and current: agy --version (expect 1.0.x). If the command is missing, tell the user to install it — don't try to install it yourself.

Defaults — start here

When you just need to hand work to agy and the user hasn't specified details, this is the safe default shape. Adjust from here; don't reason it out from scratch each time.

cd                        # scope the workspace
agy -p "" \
  --model "Gemini 3.1 Pro (High)" \         # capable default; drop to Flash for grunt work
  --dangerously-skip-permissions \          # or --sandbox; a headless run hangs without one
  --print-timeout 20m                       # bump above the 5m default for anything non-trivial

Then pick deliberately: model by task difficulty (table below), timeout by expected runtime, permission flag by how much you trust the workspace, and — for review/handoff — tell agy in the prompt to write its result to a file you can poll.

Gotchas that will bite you

• A wrong --model string does not error — it silently runs the default model. Copy strings

verbatim from agy models. A typo means you paid for the wrong tier and never knew.

• The default --print-timeout is 5m, and if it fires mid-run you lose all output. Size it

to the task; a real refactor or full-repo review needs 15–20m+.

• Without --dangerously-skip-permissions or --sandbox, a -p run hangs waiting for

interactive tool approval that never comes.

• "Independent" review on a Claude model isn't independent if you (the caller) are Claude.

Route QC to Gemini or GPT-OSS. See the harness-diversity rule below.

• Some commands in blog posts don't exist in the installed CLI (e.g. agy run, agy inspect,

-m, /context, config.toml). Worse, an unknown subcommand doesn't error — agy prints its usage banner and exits 0, so a bogus agy inspect looks like it worked while doing nothing. Trust agy --help / agy help over anything you half-remember.

The one command you'll use most

Headless, single-shot, prints the answer to stdout and exits:

agy -p "PROMPT" --model "MODEL" [--add-dir PATH ...] [--dangerously-skip-permissions]

• -p / --print / --prompt — run one prompt non-interactively. This is the mode for

orchestration: output is clean stdout you can capture, pipe, or redirect to a file.

• Runs against the current working directory as its workspace. cd into the repo (or the

right worktree) first, or add roots with --add-dir (repeatable).

• Default --print-timeout is 5m. Long tasks (large refactors, full-repo review) need a

bigger budget: --print-timeout 20m. If agy is still working when the timeout fires you lose the output, so size this to the task.

• Without --dangerously-skip-permissions, agy will block waiting for interactive approval of

tool calls — which hangs a headless run. For autonomous delegation you almost always want either --dangerously-skip-permissions or --sandbox (see Permissions & safety).

Interactive variants (rarely what you want when you are the caller):

• agy — open the TUI in the current dir.
• agy "PROMPT" — open the TUI seeded with an initial prompt.
• agy -i "PROMPT" / --prompt-interactive — run an initial prompt, then stay interactive.

Picking the model (cost-arbitrage — do this deliberately)

agy does not default to its strongest model — it uses whatever is in ~/.gemini/antigravity-cli/settings.json (often Flash). Always pass --model explicitly so the run is reproducible and matched to the task. Get the exact strings from agy models and copy them verbatim, including the parenthetical reasoning level — the string is the model's display name, e.g. "Gemini 3.1 Pro (High)".

⚠️ Silent fallback gotcha: an unrecognized --model string does not error — agy quietly falls back to the default model and runs anyway. A typo means you paid for the wrong model without knowing. Copy from agy models; don't hand-type.

Available models (verify with agy models, they change):

| Model string | Reach for it when | | --- | --- | | Gemini 3.5 Flash (Low) | Trivial, deterministic, high-volume: renames, formatting, one-line lookups, "grep-and-summarize". Cheapest/fastest. | | Gemini 3.5 Flash (Medium) / (High) | Everyday mechanical edits and straightforward coding where you want it done cheaply but not carelessly. | | Gemini 3.1 Pro (Low) | Solid reasoning on a budget — moderate refactors, multi-file edits with clear scope. | | Gemini 3.1 Pro (High) | The default workhorse for real coding and hard reasoning on the Gemini side. Standard-to-hard tasks. | | Claude Sonnet 4.6 (Thinking) | Strong general coding via a non-Gemini harness. | | Claude Opus 4.6 (Thinking) | The hardest tasks where you want maximum capability. | | GPT-OSS 120B (Medium) | A third, open-model harness — useful as a tie-breaker or an independent voice distinct from both Gemini and Claude. |

Routing heuristic: match model cost to task difficulty. Flash for grunt work, Gemini 3.1 Pro (High) as the default for genuine coding, Claude/GPT-OSS when the harness diversity itself is the point.

Harness-diversity rule for independent review: if you (the caller) are Claude, a review is only independent if it runs on a non-Claude model. Route QC/second-opinion work to Gemini 3.1 Pro (High) or GPT-OSS 120B (Medium) — not to Claude Opus 4.6, which would just be Claude reviewing Claude in a thin disguise.

Two workflows that carry their weight

1. Delegate a coding task

cd /path/to/repo   # or the specific worktree
agy -p "Convert the callback-based handlers in @internal/api to context.Context. \
Run the tests when done and report what changed." \
  --model "Gemini 3.1 Pro (High)" \
  --dangerously-skip-permissions \
  --print-timeout 15m

Give agy a goal plus a verification step ("run the tests", "make it typecheck"), not a keystroke script — it's an agent, so let it plan and self-check. Reference files with @path (e.g. @src/main.go, @src/, @**/*.ts) instead of pasting contents.

2. Independent, different-harness code review

The high-value pattern for the separate-model-review doctrine. Point agy at a diff, tell it to post its verdict where you can collect it:

cd /path/to/worktree
agy -p "You are an independent code reviewer (Gemini/Antigravity — a DIFFERENT harness \
from the Claude author). Review this branch vs origin/develop: run 'git fetch origin' then \
'git diff origin/develop...HEAD'. Assess correctness, missed edge cases, and any invented \
scope. Post your review with 'gh pr comment  --body ...' starting with the marker \
, then write your VERDICT (SHIP / APPROVE_WITH_NITS / REQUEST_CHANGES) \
to /tmp/review-DONE.md. Do NOT merge." \
  --model "Gemini 3.1 Pro (High)" \
  --dangerously-skip-permissions \
  --print-timeout 15m

Because the reviewer writes its verdict to a known file / PR comment, the orchestrator can poll for completion and read the result without parsing agy's stdout.

Continuing a conversation

agy persists each session as a SQLite DB under ~/.gemini/antigravity-cli/conversations/.db; the UUID is the conversation ID.

• agy -c "follow-up" / --continue — continue the most recent conversation.
• agy --conversation "follow-up" — resume a specific one by ID.

For orchestration prefer fresh, self-contained one-shot prompts over resuming — each -p run is stateless and reproducible, which is what you want when you're the caller. Reach for --continue only when a task genuinely needs the prior turn's context and re-establishing it in the prompt would be costly.

Permissions & safety

Headless runs need a permission strategy or they hang on the first tool approval:

• --dangerously-skip-permissions — auto-approve every tool call. Fast and autonomous, but the

agent can run arbitrary commands and edit anything in the workspace. Use it for trusted repos / disposable worktrees, and scope the workspace (cd + --add-dir) so it can't wander.

• --sandbox — run with terminal restrictions (an isolation boundary). Safer for untrusted or

exploratory work; pair with the proceed-in-sandbox permission behavior so it proceeds without prompts inside the sandbox.

Prefer the least dangerous option that still lets the run complete unattended. When you hand a destructive-capable flag to agy, say so in your summary to the user.

Capturing output for orchestration

• Redirect for later inspection: agy -p "..." --model "..." > /tmp/agy-out.txt 2>&1.
• For a machine-readable handoff, instruct agy in the prompt to **write its result to a

specific file** (e.g. a verdict file or a patch) rather than relying on stdout scraping — the review workflow above does exactly this, and it's the robust pattern for a poller.

• --log-file PATH overrides where agy writes its own debug log (separate from the answer).

Deeper reference

Slash commands, interactive session management (/resume, /rewind, /fork, /model, /planning), AGENTS.md project context, plugins (agy plugin ..., incl. importing Gemini/ Claude plugins), MCP configuration, and settings files are documented in references/reference.md. Read it when a task goes beyond headless one-shots — e.g. setting up per-repo AGENTS.md context, wiring an MCP server, or working inside the TUI.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: kellykampen
• Source: kellykampen/agent-skills
• License: MIT

Install and usage instructions live in the source repository linked above.