Blog

Antigravity vs Claude Code: headless CLI test

Antigravity vs Claude Code on one failing Python test: Claude Code needs explicit permissions to fix it; Antigravity hit errors on a free Gemini API key.

On this page · 8 sections
  1. Antigravity vs Claude Code at a glance
  2. What I tested
    1. The failing Python fixture
    2. Antigravity with a Gemini API key
    3. Claude Code with its own login
  3. Permissions when nobody approves
    1. Claude Code: an answer can succeed while the fix is denied
    2. Antigravity: failed attempts, with defaults from Google's docs
  4. Failures a script must catch
  5. Models and plans
    1. Claude Code access
    2. Antigravity access and model availability
  6. Codex CLI on the same test
  7. Where Orbi fits
  8. Related

Google Antigravity is Google's agent-first coding product, with a desktop app, IDE extensions and a terminal CLI called agy; Claude Code is Anthropic's terminal coding agent. For this Antigravity vs Claude Code test, I ran both CLIs headless on the same failing Python test, for anyone picking one for scripts or unattended runs.

I'm Larry 🍀 Orbi, the creator of Orbi. Orbi turns GitHub Issues into reviewed, merged PRs using Pi, a different open-source agent; it does not run Antigravity or Claude Code.

I did not run the desktop app. My Antigravity attempts hit Gemini API errors or timed out, so I have no successful agy fix run to compare with Claude Code's fixes. The Antigravity permission defaults below are attributed to Google's documentation.

Antigravity vs Claude Code at a glance

Your requirement Start with What to check
A terminal agent in a Claude subscription Claude Code Free excludes it; paid plans include it
Google's coding product with a terminal CLI Antigravity Account authentication versus Gemini API key configuration
An unattended fix verified on this fixture Claude Code with explicit permissions Allow Write as well as Edit (or use acceptEdits), plus the test command
Unattended Antigravity runs agy, with a timeout API availability and JSON status (from my failed runs and Google's docs)
A result your script can accept Either Inspect the diff and run tests independently

What I tested

I tested on 19 November 2026 on Linux with agy 1.5.12, Claude Code 2.1.322 and Python 3.14.7. Each attempt started from a fresh copy of the same committed, failing fixture. The runner recorded the command, process exit code, wall time, output, Git status and diff, then ran the tests independently.

Antigravity used a Gemini API key and an isolated HOME. Its settings file contained only the provider selection below. Claude Code used its own login; I passed --safe-mode --setting-sources project so my user settings and hooks did not apply; the fixture has no project settings. The --model sonnet runs reported claude-sonnet-5-5 in JSON modelUsage.

The two agents used different models, so these runs say how each CLI behaves on one small task and nothing about which model writes better code.

The failing Python fixture

Save these files in a throwaway Git checkout and commit the failing version. Start each attempt from that version.

slug.py:

def slugify(title: str) -> str:
    """Turn a post title into a URL slug."""
    return title.lower().replace(" ", "-")

test_slug.py:

import unittest
from slug import slugify

class SlugTest(unittest.TestCase):
    def test_spaces(self):
        self.assertEqual(slugify("Hello World"), "hello-world")

    def test_punctuation_dropped(self):
        self.assertEqual(slugify("Antigravity vs. Claude Code!"), "antigravity-vs-claude-code")

if __name__ == "__main__":
    unittest.main()

The spacing test passes. The punctuation test fails because the function returns antigravity-vs.-claude-code! instead of antigravity-vs-claude-code.

Antigravity with a Gemini API key

I installed the CLI with Google's documented installer, curl -fsSL https://antigravity.google/cli/install.sh | bash. It put version 1.5.12 at ~/.local/bin/agy, a 257 MB binary, and appended a PATH line to ~/.zshrc, ~/.bashrc, ~/.bash_profile and ~/.profile. Check those files afterwards if you manage your shell configuration by hand.

Google's authentication guide describes the API key route: set modelProvider to gemini and export GEMINI_API_KEY; setting only the variable has no effect. In the isolated HOME, I used ~/.gemini/antigravity-cli/settings.json:

{"modelProvider":"gemini"}

Set the key in the environment before running agy:

export GEMINI_API_KEY="<your Gemini API key>"

I pointed HOME at a separate directory holding only that settings file, so my own Antigravity configuration could not affect the run. Create that directory and settings file, then run the command:

export AGY_HOME="$HOME/agy-home"
mkdir -p "$AGY_HOME/.gemini/antigravity-cli"
echo '{"modelProvider":"gemini"}' > "$AGY_HOME/.gemini/antigravity-cli/settings.json"
HOME="$AGY_HOME" GEMINI_API_KEY="<your Gemini API key>" timeout 480 ~/.local/bin/agy \
  -p 'Fix slugify in slug.py so every test in test_slug.py passes, then run python3 -m unittest and report the result.' \
  --model gemini-3.6-flash-low \
  --output-format json \
  --print-timeout 7m

-p runs one prompt without the interactive interface and exits. -low is the low thinking level listed by agy models. I used 3.6 Flash because, right before this run, a direct API request to it returned HTTP 200 while 3.8 Flash returned 503. Pick a model from agy models that still has quota. Keep the outer timeout longer than --print-timeout: in my runs a print timeout exited 3, while an outer timeout kill exits 124.

This attempt, ag-e2-default-b2, returned a Gemini API 503 error and reached its print timeout. It did not fix the file.

Claude Code with its own login

With Claude Code already authenticated, the default-permission attempt was:

timeout 600 claude --safe-mode --setting-sources project \
  --model sonnet --output-format json \
  -p 'Fix slugify in slug.py so every test in test_slug.py passes, then run python3 -m unittest and report the result.'

After each agent exits, run python3 -m unittest yourself and inspect git status --short and git diff. A successful fix changes slug.py, leaves test_slug.py untouched and passes both tests. Some runs also left an untracked __pycache__/ directory.

Permissions when nobody approves

Claude Code: an answer can succeed while the fix is denied

Anthropic's permission-mode docs say that since v2.1.302, claude -p without an explicit permission mode starts in Manual mode. Tool calls needing approval are denied unless another mechanism allows them.

My default run, cc-e2-default, attempted to write slug.py with the Write tool and run the tests with Bash. Both were denied. Claude Code printed the proposed fix in its answer, left the file unchanged and exited 0 with is_error: false. The independent test run still failed.

All runs used the same fix prompt. Every run in this table exited 0, including the three that left the test failing.

Permission configuration Run Result Exit
No permission mode or allow rules cc-e2-default Write and Bash denied; no fix 0
--allowedTools Edit "Bash(python3 -m unittest:*)" cc-e3-allow, cc-e3-allow-r2 Tests ran; Write denied; still failing 0
--allowedTools "Edit(./**)" "Bash(python3 -m unittest:*)" cc-e3c-editpath Fixed; 2/2 passed 0
--permission-mode acceptEdits --allowedTools "Bash(python3 -m unittest:*)" cc-e3b-acceptedits Fixed; 2/2 passed 0
--permission-mode dontAsk --allowedTools "Read,Edit,Write,Bash(python3 -m unittest:*)" cc-e3d-dontask Fixed; 2/2 passed 0
--dangerously-skip-permissions cc-e4-skip Fixed; 2/2 passed 0

In both failed allow-rule runs, the model chose Write and the bare Edit rule did not authorize it. The allowed test command ran against the unchanged source. The path-scoped Edit(./**) configuration succeeded without a permission denial, but its result JSON does not identify which tool changed the file. These runs do not prove whether the path form or a different tool choice made the difference. In practice, allow Write explicitly, as the dontAsk command below does, or use acceptEdits.

For CI, I would use the tested dontAsk command and list the required tools explicitly:

timeout 600 claude --safe-mode --setting-sources project \
  --model sonnet --output-format json \
  --permission-mode dontAsk \
  --allowedTools 'Read,Edit,Write,Bash(python3 -m unittest:*)' \
  -p 'Fix slugify in slug.py so every test in test_slug.py passes, then run python3 -m unittest and report the result.'

The documented dontAsk mode denies calls that would otherwise prompt, so the session does not wait for input. The acceptEdits configuration also worked when I allowed the test command separately. I would reserve --dangerously-skip-permissions for isolated containers. Run Claude Code unattended covers the surrounding process setup.

With the diagnostic prompt, "List the files in this directory and tell me which unit test fails and why. Do not change any files.", Claude Code changed nothing (cc-e1-readonly). One Bash call attempting to run tests was denied; it diagnosed the punctuation bug by reading the files and exited 0.

Antigravity: failed attempts, with defaults from Google's docs

The recorded fix attempt ag-e2-default-b2 used gemini-3.6-flash-low, returned status: "ERROR" with an API 503 message and ended with finished_reason: "print_timeout". Its process exited 3 after 420.9 seconds. The Git diff was empty and the punctuation test still failed.

The diagnostic attempts also hit 503 errors or print timeouts. I have no successful agy fix run and cannot report measured permission behavior for the default mode, an allow rule, plan mode or skipped permissions. In particular, an unchanged file after an API failure does not show that a permission rule blocked a write.

Google's CLI reference says agentMode defaults to "auto" where supported from v1.5.7 onward, otherwise to the standard default mode. The same reference says auto mode requires the terminal sandbox. It auto-approves edits and runs commands in that sandbox by default. A command that needs to run outside the sandbox is checked automatically: safe, reversible ones run without prompting, and irreversible actions still prompt. Google's sandbox page says Linux uses kernel namespaces for that isolation.

The headless documentation says actions that need approval and cannot obtain it are soft-denied: the run continues, exits 0 and prints a permission notice to stderr. My runs ended on API errors and timeouts before reaching that point.

Google documents advance grants through permissions.allow, including command(git) and write_file(src/). It also documents --dangerously-skip-permissions as approving every tool request. I could not test any of them.

Failures a script must catch

My Antigravity attempts used a free-tier Gemini API key. Many returned Gemini API 503 errors reporting high demand. I reproduced a 503 with a direct curl request to the Gemini API using the same key, without agy.

Even a short response and exit 0 did not establish success. The model-availability probes asked only for Reply with exactly: ok:

Antigravity probe model Response JSON status Process exit
gemini-3.6-flash-low ok ERROR, with API 503 0
gemini-3.8-flash-low ok ERROR, with API 503 0
gemini-4-hydrogen-low ok SUCCESS 0

In the two ERROR rows, one JSON envelope held both the reply ok and a 503 message of the form API error (attempt N), and the status stayed ERROR.

The SUCCESS probe only answered a one-line prompt; it edited nothing and ran no tests. Probes for gemini-3.1-pro-low and gemini-4-carbon-low returned Free Tier quota-exceeded errors on their first request, so those models appear to have no usable free quota on this key. Later fix attempts on gemini-3.6-flash-low failed in under a second with a 429: the free tier allowed 20 generate_content requests per day for that model, and my earlier attempts and probe requests had used them up. One 503 probe logged attempt 5, so the CLI retried that request at least four times; a failed run can therefore use several requests from the daily quota. The 429 runs ended in under a second. These limits belong to the Gemini API key's free tier; I did not test Antigravity's own plan quotas.

Authentication and model-selection failures were different:

Setup Observed result Wall time Exit
agy, not logged in Printed a Google OAuth URL, waited, then authentication failed or timed out 60.2s 1
agy, Gemini provider selected but no key Clear message that GEMINI_API_KEY was not set 0.1s 1
Claude Code, not logged in Not logged in · Please run /login 0.7s 1
Claude Code, unknown model Model-selection error; is_error: true 2.0s 1

For Antigravity, check .status as well as the process exit code, and retain stderr for permission notices. For Claude Code, check .is_error and permission_denials; the default run had no JSON error but did have denied tools and an unfixed test. Run against my saved JSON, the first check exits 1 for the failed agy run ag-e2-default-b2; the second exits 1 for Claude Code's default run and 0 for the dontAsk run:

jq -e '.status == "SUCCESS"' agy-result.json
jq -e '.is_error == false and (.permission_denials | length) == 0' claude-result.json

In both cases, inspect the diff and run tests independently before accepting the change.

Set a time limit on the job. I used Antigravity's --print-timeout as well as an outer timeout; several attempts exited 3 with partial output at the print limit. Google's headless guide documents that timeout flag. These failure times do not measure coding speed.

Models and plans

I checked the official plan pages on 19 November 2026. My Gemini API key runs are separate from Antigravity's Google AI plan quotas; I did not test those subscription allowances.

Claude Code access

Claude plan Price and Claude Code access
Free Does not include Claude Code
Pro $20/month billed monthly; $17/month with annual billing; includes Claude Code
Max $100/month; includes Claude Code
Ultra $200/month; includes Claude Code

I did not measure what the fixture runs cost.

Antigravity access and model availability

Google's plans page says all plans receive baseline quota, including Gemini 3.8 Flash and access to the CLI. On Antigravity's free plan, quota refreshes weekly. Google AI Pro and Ultra raise limits and refresh quota every five hours, with weekly limits also applying. Pro and Ultra users can use purchased AI credits for overages.

The same page says BYOK and custom endpoints are not supported for additional Antigravity rate limits. The Gemini API provider setting used in this test sends requests directly to the Gemini API; it does not demonstrate an increase to a Google AI subscription's quota. See Google's plan details for the available routes; I have not verified US dollar prices for those plans.

Google's model table lists Claude Sonnet 5.5 (thinking) and Claude Opus 5.5 (thinking) for Google AI Pro with a non-trial subscription, Ultra and Enterprise. When I checked it on 19 November 2026, it still listed Claude Sonnet 4.6, Claude Opus 4.6 and GPT-OSS-120b, with a note that they would be removed on 18 November 2026. I did not verify whether the removal had happened.

In my Gemini API key configuration, agy models listed only Gemini models. So don't assume a model in the plan table is available through the API key route.

Codex CLI on the same test

If your shortlist is Antigravity vs Codex instead, here is Codex on the same fixture. I tested Codex CLI 0.167.0 with gpt-6.2-sol through a ChatGPT login on the same Python fixture. In this test environment, the run without a sandbox override reported a read-only sandbox and approval never. It left the source unchanged, as Claude Code's default run did, but for a different reason: Codex could not write, while Claude Code tried and was denied. Google documents Antigravity's auto mode as approving workspace edits.

These commands rely on that environment's default model; I did not pass a model flag.

timeout 400 codex exec --skip-git-repo-check \
  'Fix slugify in slug.py so every test in test_slug.py passes, then run python3 -m unittest and report the result.'

Allowing workspace writes fixed the source:

timeout 400 codex exec -s workspace-write \
  'Fix slugify in slug.py so every test in test_slug.py passes, then run python3 -m unittest and report the result.'
Codex configuration Run Result Wall time / exit
No sandbox override in this environment cx-e2-default No change; one test still failed 18.9s / 0
-s workspace-write cx-e4-workspace-write Fixed; both tests passed 16.1s / 0

OpenCode vs Codex has a separate CLI comparison; Codex yolo mode covers broader permission settings.

Where Orbi fits

The commands above run an agent against a checkout. You still need to arrange review and merging if you put them in a delivery workflow.

Orbi starts from a GitHub Issue labelled ai-ready, implements with Pi, has a separate session review the work, and merges when review passes and CI is green.

In Claude Code, the Orbi plugin provides /orbi:ship to turn the current task into an Issue. In a repository Orbi delivers, it creates the Issue with ai-ready after confirmation. /orbi:status checks delivery status.

Antigravity is not in that README's supported list. I have not tested an Antigravity integration and cannot claim the Orbi skills work there. Orbi compared with Claude Code explains the Issue-to-PR handoff; Orbi is a poor fit when you want an agent to pair with you live.

What is new in Orbi, every week

One email a week: new features, hands-on posts and numbers from our own work with Orbi. Unsubscribe anytime.