Official ACP 1.3.0: supported consultation-only controls on an individual Google AI Pro account

I would like to use Google’s unmodified official Windows Antigravity ACP 1.3.0 server as a local child process of a single-user AI coding assistant. Authentication and provider traffic would remain inside Google’s supported unmodified client. The controlling assistant would not extract tokens or call Google’s backend endpoints directly. It would not operate a shared relay or multi-user service, or use high-frequency polling.

The Google-marked moderator’s clarification for the official local CLI also points to the official ACP adapter. Please confirm whether this narrow ACP use is supported for an individual Google AI Pro account, alongside the broader FAQ/Terms wording.

The intended operation is one fresh session, one explicit text prompt, and a complete Gemini text answer. The coding assistant would retain responsibility for PC tools. The proposed ACP client advertises no file or terminal services and uses this session envelope:

{
  "cwd": "<fresh empty consultation workspace>",
  "mcpServers": [],
  "_meta": {
    "agy": {
      "enabledTools": []
    }
  }
}

The client identifies itself accurately rather than impersonating an IDE, rejects permission/client-service requests, and would use a reviewed fresh configuration root and pinned companion executable. No model prompt has been submitted through this ACP candidate.

A bounded static inspection of the shipped wrapper shows the empty allowlist reaching SDK built-in capability settings. It also shows global MCP/hooks/skills being loaded separately. The companion’s built-in default selects the core agent, whose fundamental-component suppression is a separate setting; its final tool-override stage appears to decorate existing tools rather than remove them. These observations are limited to the inspected code, not a claim about the complete effective runtime or a reported unwanted action.

Could you provide the supported procedure for these requirements?

  1. Disable every PC/tool action before a prompt, including command/file/web/image tools, MCP, hooks, subagents and internal task/message/scheduling actions. Plain text-answer completion is still needed. Does enabledTools: [] cover the fundamental components, and if so, where is that enforced? If it does not, what official consumer-ACP configuration provides this boundary?
  2. Restrict implicit context to the explicitly supplied prompt and fresh session. What supported controls exclude global/ancestor/project rules, skills, hooks, plugins, prior-session data and other implicit local inputs? Is a fresh GEMINI_HOME plus empty cwd sufficient across both the wrapper and companion, after configuration/experiment merges?
  3. Inspect the effective tool/context configuration before inference, through a guaranteed metadata-only diagnostic or handshake. The declared settings diagnostic and model/mode options do not appear to report the final constructed registry.
  4. Select only the account-advertised Gemini text model, disable any credit/spend fallback, and verify effective telemetry settings. Please identify the supported controls and their scope; an API-key, paid API, Vertex or enterprise fallback is outside this proposed individual-account route.

If this consultation-only boundary is currently unsupported, please state that explicitly and identify any official planned or documented consumer route. I will keep the candidate inactive until its controls are verified.

Hi @K_K7,

Welcome to the Forum and thank you for reaching out.

Here is a summary for each of your questions :

  • 1. Complete Tool Suppression (_meta.agy.enabledTools: []): Unsupported. The official ACP adapter is built for full agentic editor integrations (such as our Zed extension). _meta.agy.enabledTools is an internal field that does not suppress core agent components, and there is no supported consumer ACP setting to strip all tools while retaining plain text completion.
  • 2. Excluding All Implicit Context: Unsupported. The runtime automatically loads customizations and rules from global (~/.gemini/), ancestor, and workspace directories, and neither GEMINI_HOME nor an empty cwd is a guaranteed isolation boundary across both the ACP wrapper and the companion binary.
  • 3. Pre-Inference Tool/Context Inspection Handshake: Unsupported. Neither the ACP server nor the CLI Reference provides a pre-inference dry-run handshake that outputs the final merged tool and context registry before a prompt runs.
  • 4. Model Selection, Credit Fallback & Telemetry Controls: In the official Antigravity CLI (agy), you can select a model via --model <model_name>, and overage credit usage is off by default unless explicitly enabled in account settings (AI Credits documentation). However, there is no documented consumer route in agy or agy_acp_server for a guaranteed zero-tool, plain text-completion mode on an individual Google AI Pro subscription.

Please let us know if you have any further questions.