2026 Google Gemini CLI Policy Shift:
Antigravity Migration, Open-Source Trust, and What Developers Should Do

Google’s public framing, in the Developers Blog and GitHub Discussion #27274, is pragmatic: terminal AI must evolve for the multi-agent era, so experiences consolidate under Antigravity CLI. For individual developers, the story reads differently—open source to gather momentum, then close the hosted front door.

• Promise versus rhythm: The Gemini CLI repository remains Apache 2.0 on GitHub, and Google states that enterprise customers will keep receiving model and security updates through that client. Yet the online inference entry for personal and Pro/Ultra accounts stops on 2026-06-18, which clashes with the mental model that “open source means I can keep using it as-is.”
• Contributor economics: Outlets including TechTimes cite thousands of merged pull requests and six-figure star counts on the public repo (numbers drift daily—verify on GitHub before you quote them in a contract). Contributors such as Andrea Alberti asked bluntly in the discussion thread whether the community effectively subsidized an enterprise-facing codebase.
• Replacement maturity: The Register and The New Stack relayed early user reports that Antigravity did not reach feature parity at launch. Community comparisons claimed free daily call limits dropped from roughly 1,000 requests per day under Gemini CLI-era expectations to about 20 per day on Antigravity—a ~98% reduction in that informal yardstick, not a figure Google published side by side. Treat any quota number as something you must measure on your own account.

Anchor dates (re-open official links before you ship runbooks):

• 2025-06: Gemini CLI open-sourced under Apache 2.0.
• 2026-05-19: Google I/O introduces Antigravity CLI and announces Gemini CLI service changes for consumer tiers.
• 2026-05-23 onward: GitHub discussion #27274 fills with negative feedback (vote counts change live).
• 2026-05-29: Linux Foundation highlights the case at the Open Source Summit North America and promotes the isitopen.ai Model Openness Tool.
• 2026-06-18: Service ends for free, Pro, Ultra, and personal Gemini Code Assist paths using Gemini CLI; new installs under the GitHub org face restrictions.

FOSS Force reporter Christine Hall captured the architectural lesson in one line: the license did not change, but the infrastructure that made the tool useful did.

The policy is not a repo takedown. It is a tiered shutdown of the default cloud auth path while leaving some enterprise and billing-backed routes intact. The matrix below synthesizes Google’s May 2026 announcements and subsequent press coverage; quotas and invoices always come from your console.

Antigravity is marketed as the single harness for background agents, a new terminal UX, and server-side orchestration. Early complaints focused on closed source, incomplete Agent Client Protocol (ACP) documentation, project memory via Markdown context files, and oddities such as broken Ctrl+C handling. Teams at Dynatrace, Elastic, Figma, Shopify, and Stripe that embedded Gemini CLI deeply now carry migration communication cost even when their contracts technically survive.

If you are architecting access today, the API-key row matters most: it is the clearest way to keep using the open client while paying for inference instead of betting on a free OAuth tier that can change after the next keynote. Read the primary sources yourself rather than trusting secondary star counts:

Google Developers Blog — Transitioning Gemini CLI to Antigravity CLI

GitHub Discussion #27274 — Transitioning Gemini CLI to Antigravity CLI

The fight is not about whether Apache 2.0 still applies to the repository. It is about the open client plus closed model API pattern that defines most 2026 AI CLIs. You can fork the UI shell; if authentication, routing, quotas, and weights stay in Google’s cloud, the fork’s practical value collapses the moment personal OAuth stops working.

• Double standard narrative: Google argues consolidation is inevitable in a multi-agent world. Critics counter that enterprise customers may keep both Gemini CLI and Antigravity while individuals are pushed one way—suggesting commercial prioritization, not a hard technical impossibility.
• Pseudo-open source: Classic open source let you fork and self-host. AI CLI “open source” often exposes plugins and terminal chrome while models, billing, and identity remain proprietary. Tools like isitopen.ai score transparency, reproducibility, and usage rights beyond the SPDX line on GitHub.
• Trust compounding: Google’s consumer product graveyard (Reader, Stadia, and others) primes developers to discount long-horizon free-tier promises. Some teams accelerated experiments with Claude Code, Codex CLI, or local Ollama stacks rather than waiting for Antigravity parity.

For Mac teams running OpenClaw, Hermes, or Cursor Agent Skills, the durable pattern is simple: treat cloud CLIs as replaceable inference backends and keep SOP assets—Skill folders, gateway configs, launchd plists—on hardware you control. That way a June policy email does not erase six months of workflow design.

# Pre-integration checklist (paste into architecture review)
1. Client license Apache/MIT — can you audit the code?
2. Inference tied to vendor OAuth only, or BYO API key?
3. Free-tier quota history — any one-way tightening events?
4. Closed successor — feature parity and docs complete?
5. Team skills locked to one CLI config format?

1. Map identity and billing: Confirm whether accounts are free, Pro/Ultra, personal Code Assist, or enterprise Standard/Enterprise. In Google Cloud Console, check for an existing paid Gemini or Vertex API key; that path remains viable for Gemini CLI per Google’s announcement.
2. Export local dependencies: Inventory ~/.gemini/, IDE extensions, and CI jobs invoking gemini. Flag anything that breaks on June 18 if it depends on consumer OAuth.
3. Benchmark Antigravity gaps: Install Antigravity on a sandbox Mac. Validate subagents, hooks, project memory files, and ACP against production flows. Plot daily quota burn; do not assume Pro subscriptions escape throttling without measurement.
4. Stand up alternate inference: Pilot Claude Code, Codex CLI, Ollama with OpenHuman, or your own Vertex/Gemini API key feeding the open Gemini CLI binary. Avoid wiring all automation to a single free OAuth curve.
5. Move appreciating state out of the CLI: Version Agent Skills (SKILL.md), Hermes ~/.hermes/, and OpenClaw gateway settings in Git or encrypted backups. CLIs are interchangeable; procedural capital is not.
6. Choose a 24/7 host for agents: Interactive terminals can stay on a laptop, but gateways, Telegram bots, and CI watchers need a macOS machine that does not sleep on lid close, with launchd health checks and log rotation. That is where bare-metal rental enters the picture.

Enterprise integrators should loop legal and procurement on Code Assist license scope and notify maintainers of GitHub org integrations so PR bots do not fail silently after the cutoff.

• Effective date: Consumer and subscription-tier Gemini CLI service ends June 18, 2026 per Google’s Developers Blog (check for amendments).
• Repository license: Gemini CLI stays Apache 2.0 on GitHub; enterprise customers are promised continued model and security updates through that client.
• Community scale (unaudited): Press reports cite 6,000+ merged PRs and 100,000+ stars during the open phase—confirm live on google-gemini/gemini-cli before citing externally.
• Openness scoring: Linux Foundation promoted isitopen.ai in May 2026 to evaluate AI products beyond license badges.

Further reading (open each link independently; do not treat blog summaries as contractual text):

The Register — Bye-bye, Gemini CLI

The New Stack — Google pushes users to Antigravity CLI

FOSS Force — Gemini CLI's Short Life and Antigravity Bait-and-Switch

Building a multi-agent practice entirely on a free cloud CLI whose quotas change at I/O is fast until it is not. You eventually face quota cliffs, closed-source migrations, or account policy shifts. The more resilient stack pairs your own API keys, an open client you can audit, and a stable macOS runtime you operate. Laptops that sleep kill gateways; generic Linux VPS hosts lack Xcode, launchd, and Metal paths many Mac agents expect.

For production teams that need 24/7 OpenClaw or Hermes gateways, shared Skill libraries, and bare-metal Apple Silicon delivered in minutes, CALMVPS bare-metal Mac rental is usually the better fit than hoping a notebook stays awake on a desk. You get dedicated M-series hardware, multi-region nodes, and monthly elasticity without capital expense on a machine that might be obsolete after the next Apple silicon cycle. Review models and pricing on the CALMVPS pricing page, place an order when your migration checklist is green, and use the help center for deployment questions.