Clarification on Codex/Claude Skill invoking the official agy CLI through a local PowerShell wrapper

Google Antigravity
1

Hello Antigravity team,
I’m seeking clarification on a specific integration before using or distributing it.
My setup is:
Codex CLI or Claude Code → a locally installed Skill → scripts/invoke.ps1 → the official, unmodified agy CLI → Antigravity
The Skill is invoked for selected user-requested tasks, such as read-only code review or research. The PowerShell script launches agy locally, passes a prompt, and returns its output to the calling agent. It does not use an MCP server, extract or forward OAuth credentials, call private Antigravity APIs, share account access, or bypass quotas. Authentication remains with the official agy CLI. I would like to publish the Skill’s source code for others to install locally.
I found forum answers that appear to differ:

  • A local application spawning agy, with Codex participating in the workflow, was described as permitted.
  • A third-party agent calling agy through a local MCP wrapper was described as not permitted.
  • A local orchestrator spawning agy was described as conditionally permitted.
    Could someone authorized to clarify Antigravity policy answer these questions for the exact Skill → PowerShell wrapper → official agy setup?
  1. Is it permitted when Codex CLI or Claude Code chooses to invoke the Skill and receives agy’s result?
  2. Does the answer change if a person explicitly requests each invocation, rather than the agent selecting the Skill during a task?
  3. May this setup use a [personal Google account / Google AI Pro subscription — select the applicable one], or does it require a separate API-based integration?
  4. Is publishing the Skill’s source code permitted, assuming users install and authenticate the official agy CLI themselves?
  5. If it is not permitted, what supported route would provide this workflow?
    I would appreciate a clear yes/no answer for this architecture and an explanation of how it differs from the examples above, given Section 6 of the Antigravity Additional Terms and the FAQ.
2

Hi @nothing2,

Thanks for laying out your setup so clearly!

Short answer: yes, your setup is fine. What’s not allowed is extracting sign-in tokens, calling private Antigravity endpoints, or using agy as a raw model backend for another agent. Running the official agy CLI locally to do a specific task and reading its output is allowed. That’s the same rule behind all the forum answers you found.

To your questions:

  1. Codex or Claude calling the Skill: Yes, as long as it sends agy a real task (like a code review) and doesn’t pass another agent’s prompts through it.
  2. You triggering it vs. the agent: Both are fine. Triggering it yourself is safer, so you don’t hit rate limits with rapid bursts.
  3. Personal or AI Pro account: Yes, within your plan’s normal limits.
  4. Publishing it: Yes. Just don’t bundle agy or any credentials, stick to documented flags, and mark it as an unofficial community Skill.
  5. Other options: headless mode (antigravity.google/docs/cli/headless), the Antigravity SDK or Gemini API for API-key setups, or the official ACP adapter.