Google models outside Antigravity

How to run Google’s frontier models (Gemini/Antigravity family) from harnesses that aren’t Google’s own Antigravity IDE or agy CLI — a lane that mutated hard in 2026 and now has a safe official route.

The danger: reverse-engineered OAuth (the flagged-account era)

Most third-party integrations went through the private Cloud Code Assist login — reverse-engineered OAuth against endpoints Google never offered for this purpose. Consequence: main Google accounts getting flagged (@hqmank thread, captured). The now-retired @google/gemini-cli individual Code Assist path was exactly this. Rule of thumb: if the auth flow was discovered by packet-sniffing a first-party client, the account risk sits with you.

The official routes (2026 state)

• Antigravity’s own ACP server — Google exposes an official Agent Client Protocol endpoint; anything that speaks ACP can drive Antigravity models with Google’s blessing. Protocol v1, official TS SDK.
• pi-antigravity-acp-provider (OSS) (via @hqmank; TS) — the showcase instance: pi install npm:pi-antigravity-acp-provider → /login → Google Antigravity (ACP) → browser sign-in → /model → antigravity-acp/*. Implementation details worth stealing: collapses Antigravity’s effort-qualified model IDs into one entry with Pi’s Shift+Tab picking low/medium/high reasoning dynamically; streaming, cancellation, usage/quota metadata; persisted ACP session restoration across restarts; Pi tools routed through an authenticated loopback MCP bridge; permission modes (default yolo, fail-closed broker when prompting); auth-health, quota, and runtime-update subcommands; advertises no ACP filesystem/terminal capabilities (principle-of-least-capability, matching Pi agent harness’s no-list).
• Gemini API key — the boring path (AI Studio key → any OpenAI-compatible harness); what Pi agent harness’s pi-ai provider list means by “Google Gemini”. Limits differ from the Antigravity subscription quota.

The officially supported enterprise rails (2026)

For Workspace/corporate accounts, four rails are documented (cross-checked against Google’s own docs + the antigravity.google enterprise page):

1. Gemini Code Assist Standard/Enterprise license — OAuth sign-in + GOOGLE_CLOUD_PROJECT = licensed project; 1,500–2,000 req/day/user; explicitly unaffected by the 2026-06-18 personal-OAuth cutoff (Workspace accounts always needed a Cloud project anyway).
2. Vertex AI — ADC (gcloud auth application-default login), service-account JSON, or Cloud API key + GOOGLE_CLOUD_PROJECT/GOOGLE_CLOUD_LOCATION; IAM, audit logs, regional residency; the CI/prod path.
3. Gemini API key (AI Studio / Gemini Enterprise Agent Platform) — metered, billing-gated, documented; the simplest ToS-clean rail.
4. Antigravity + Gemini Enterprise — corporate SSO; two modes (Cloud project+API consumption, or Gemini Enterprise Standard/Plus license with quotas/managed overages/admin controls); ADC via AGY_ADC_AUTH=true; VPC Service Controls + data residency (“never stored outside your private environments”); supported clients = Antigravity 2.0/CLI, VS Code, VS, JetBrains/Zed/Xcode previews — standalone Antigravity IDE not supported for enterprise.

Why this matters here

• The ACP server move makes Google the first frontier lab to treat harness-neutrality as a product surface — the same bet as Agent Client Protocol’s thesis (agents as standard clients), coming from the model vendor. Watch: does Anthropic follow with a Claude ACP server?
• The provider is also the best-documented ACP-client integration in the vault: session restoration, capability negotiation, permission brokering — the practical checklist any ACP integration needs.
• Fits the Pi agent harness ecosystem tier (“Pi first” as compatibility class) and MCP’s loopback-bridge pattern (tools over MCP without exposing a socket).

Related: Using Google models in enterprise (the access/quota lane)
Related: Pi agent harness · Agent Client Protocol · MCP · Inference engines (the provider-surface layer)