Using Google models in enterprise

The practical map for using Google’s models with a corporate/Workspace account — what’s officially supported, what it costs in quota, and where the friction actually is. Companion to Google models outside Antigravity (the harness-neutrality lane); this note is the enterprise-access lane.

KRZ status (2026-10-05)

Access: yes. Configured/installed: no. Reasons on record:

1. Google models are subpar (relative to the current frontier set in daily use).
2. Harness limitations — Google’s playbook funnels users toward Antigravity; using the models from other harnesses is second-class (see the timeline below).

The four officially supported rails

1. Gemini Code Assist Standard/Enterprise license — OAuth sign-in + GOOGLE_CLOUD_PROJECT = the licensed project; quotas 1,500 req/day/user (Standard) / 2,000 (Enterprise). Explicitly unaffected by the 2026-06-18 cutoff (that killed personal free/Pro/Ultra OAuth in gemini-cli only). Workspace/school accounts have always required a Cloud project for OAuth.
2. Vertex AI — three auth sub-paths: ADC (gcloud auth application-default login), service-account JSON, or Cloud API key; requires GOOGLE_CLOUD_PROJECT + GOOGLE_CLOUD_LOCATION. Buys IAM, audit logs, regional data residency — the CI/prod and enterprise-governance path, model-by-model via the aiplatform.googleapis.com API.
3. Gemini API key (AI Studio / Gemini Enterprise Agent Platform) — metered, billing-gated, simplest ToS-clean rail; free tier 250 req/day Flash-only (irrelevant for org use, the paid rail is what matters).
4. Antigravity + Gemini Enterprise (the 2026 flagship, antigravity.google/docs/enterprise) — corporate SSO; two modes: Cloud project + API (consumption billing) or Gemini Enterprise license Standard/Plus (included quotas, managed overages, centralized admin); ADC support via AGY_ADC_AUTH=true; VPC Service Controls + data residency (“prompts, responses, code, and telemetry never stored outside your private environments”). Supported clients: Antigravity 2.0/CLI, VS Code, Visual Studio, JetBrains (preview), Zed/Xcode (preview) — standalone Antigravity IDE is NOT supported for enterprise.

The timeline that explains the friction

• 2026-05-19 (I/O): Google consolidates all developer AI tooling under the Antigravity brand; ~30-day migration window announced.
• 2026-06-18: Gemini CLI / Code Assist IDE extensions stop serving requests for personal accounts (free, AI Pro, AI Ultra). The open-source gemini-cli survives on API key / Vertex / enterprise-license auth only. agy (Antigravity CLI, closed-source) becomes the personal-account client, with its own ANTIGRAVITY_API_KEY auth.
• The pattern: every consolidation step narrows the harness choice — personal accounts now have exactly one first-party client (Antigravity), and even the enterprise story funnels toward Antigravity-first clients. The ACP server (Google models outside Antigravity) is the one counter-move: an official protocol route that restores harness freedom — currently the most interesting way out of the funnel.

The decision logic (why “access but not configured”)

• If the models aren’t competitive with what’s already in daily rotation, quota and governance are moot — no reason to spend setup time.
• If/when a Google model earns a slot (a strong reasoning or long-context release), the configuration order is: Vertex AI via ADC first (least funnel, most governance, works headless), API key second (simplest), Code Assist license only if the org assigns one, Antigravity clients last (best model access, worst harness lock-in).
• Watch trigger: Antigravity ACP server model coverage — if the flagship models (not just effort-qualified variants) become ACP-reachable, the harness-freedom objection dissolves and the note flips to “configure it.”

The gray-market rail (for completeness): CLIProxyAPI

CLIProxyAPI (OSS) (54.2k★, Go; router-for-me) — wraps Antigravity, ChatGPT Codex, Claude Code, Grok Build, Muse Code, Devin CLIs behind an OpenAI/Gemini/Claude/Codex-compatible local API, with multiple CLI accounts per provider (OAuth logins harvested from the official CLIs, round-robin across subscriptions). It is the polished incarnation of exactly the pattern this note warns about: subscription OAuth repurposed as an API — ToS-gray, account-risk (the flagged-account era was the individual version of this), and the README’s sponsor wall (API relays, residential proxies “one-IP-per-account”) reads as the infrastructure of quota arbitrage at scale. Documented as the counter-example: what the official rails (above) exist to make unnecessary. Kin of the reverse-engineered-Cloud-Code-Assist lane in Google models outside Antigravity.
Related: Google models outside Antigravity · Agent Client Protocol · Inference engines · AI engineering stretch goals

The Amp-orb deployment recipe (2026-10-10)

ben-vargas gist (captured, 18.8 KB): the operational extreme of this gray-market rail — CLIProxyAPI running inside an Amp orb, published via Tailscale Funnel as a stable public HTTPS endpoint, wired into Amp as a Custom URL model-routing connection, serving Claude requests off a Claude Pro/Max OAuth subscription instead of Amp credits. The architecture is the interesting part: Amp’s servers call Custom URLs themselves, so the endpoint must be publicly reachable (orb ports aren’t) → Tailscale Funnel gives a stable *.ts.net name + real TLS + SSE (Cloudflare Quick Tunnels rejected: no SSE, new URL every run); the orb joins the tailnet via Amp workload identity (OIDC) — no Tailscale auth key stored anywhere, node ephemeral; orbs pause when idle, so an Amp plugin holds a keep-alive lease on the designated host orb. Cost honesty: the never-pausing host at a1.small ≈ $124/month to serve a subscription you already pay for (observed load 0.07, 700 MB). Ships as an Amp skill (amp skill add ben-vargas/amp-plugins/skills/amp-cli-proxy-api) that walks the setup. For this note: the ToS-gray flag stands (the gist itself: “using a subscription through a proxy is your decision under Anthropic’s terms”) — but the deployment pattern (OIDC-joined ephemeral orb + Funnel + keep-alive lease) is a clean reusable recipe for exposing any orb-hosted service publicly, independent of the gray-market payload.