2026 Google Gemini CLI Policy Change: Open Source Trust Crisis, Antigravity Migration, and Developer Playbook

Gemini CLI is a TypeScript terminal AI coding assistant with Agent Skills, Hooks, Subagents, and Extensions—alongside Cursor and Claude Code in the 2025–2026 CLI agent wave. Google open-sourced it under Apache 2.0 in June 2025, drew 100k+ GitHub stars and 6,000+ merged PRs in a year, and won integrations from Dynatrace, Elastic, Figma, Shopify, and Stripe.

• Who loses service: Google AI free users, Google AI Pro ($19.99/mo), Google AI Ultra ($249.99/mo) subscribers, and Gemini Code Assist for Individuals.
• GitHub extension: new installs of Gemini Code Assist for GitHub blocked from June 18; existing personal requests wind down within weeks.
• Official replacement: affected users are routed to closed-source Antigravity CLI—not a maintained public endpoint for open Gemini CLI.
• Enterprise exemption: Gemini Code Assist Standard / Enterprise license holders, GCP-backed GitHub extension users, and paid Gemini / Gemini Enterprise Agent Platform API key holders are unaffected.
• Community mood: the GitHub discussion drew heavy negative feedback; a top comment sums it up: "As always, Google being Google." FOSS Force reporter Christine Hall wrote: Google "didn't change the open source license" but "shut down the infrastructure that made the tool useful."

The fight is not the license text—it is who controls runtime. Apache 2.0 lets you fork code, but auth, API endpoints, and quotas stay with Google. Without model API access, a forked CLI barely runs on its own.

Google's stated reason: focus on Antigravity as a single "multi-agent era" platform. The community counter: enterprise users may run both Gemini CLI and Antigravity CLI, while individuals and contributors are forced to switch—so the cutoff reflects commercial priority, not pure technical need. Contributor Andrea Alberti, after a 27-commit PR merged the same day, asked: "Are we essentially working for free on a codebase that only serves enterprises?"

Antigravity CLI launched at Google I/O 2026 as the terminal entry to a "premier agent-first development platform," with server-side harness and unified architecture. Google promises Agent Skills, Hooks, Subagents, and Extensions (now Antigravity plugins)—but there is no 1:1 feature parity, and the license moves from Apache 2.0 to proprietary closed source.

Early user reports: Pro subscribers hit limits after 6–7 prompts; Ctrl+C fails to exit cleanly in some environments; docs are thin and the launch felt rushed. The Register and The New Stack note most developers have no real choice: accept a quota-shrunk closed replacement, bring your own paid API key for the open binary, or switch tools entirely.

At Open Source Summit North America in May 2026, the Linux Foundation promoted isitopen.ai (Model Openness Tool)—transparency, reproducibility, and usage rights—to score how "open" AI tools really are. The Gemini CLI episode became a textbook case in pseudo-open-source debate.

1. Map dependencies: list every place the team uses Gemini CLI official endpoints—local terminal, IDE plugins, GitHub Actions, CI scripts—and flag Google-account OAuth vs API key binding.
2. Confirm account tier: use the Section 02 matrix to see if you are in the June 18 cutoff or enterprise/API-key exempt group; if the team mixes personal Pro with enterprise licenses, consolidate on enterprise paths to avoid single-point failure.
3. Evaluate API-key self-rescue: issue a paid API key from Google AI Studio or Vertex AI, wire it into Gemini CLI config, and keep Skills, Hooks, and MCP settings—official endpoints close, but the Apache 2.0 toolchain needs no rewrite.
4. Trial Antigravity CLI and stress quotas: before June 18, run real workloads against the new free tier (~20/day); if it fails daily use, switch early—do not discover failure at 9:01 AM on deadline day.
5. Compare rival CLIs: assess Claude Code (Anthropic), Codex CLI (OpenAI), Cursor CLI, and others; if you already use the Agent Skills open standard, some SKILL.md files migrate across tools and cut switch cost.
6. Institute "infrastructure dependency" review: before contributing PRs to open AI projects, use isitopen.ai or your own checklist on model API, auth, and quota lock-in; prefer tools with both open code and self-hosted API key paths.

If you are in the cutoff group and reject Antigravity quotas, the pragmatic path is bring your own API key and keep the open Gemini CLI binary, or move to Claude Code / Codex CLI and host agent processes on macOS.

• Open-source momentum: Gemini CLI passed 100k GitHub stars and 6,000+ merged community PRs in its first year—sources: TechTimes, official repo stats.
• Cutoff date: June 18, 2026—Gemini CLI and Gemini Code Assist IDE extensions stop serving Google AI Pro, Ultra, and free users—source: Google Developers Blog (published 2026-05-19).
• Free quota shift: Gemini CLI ~1,000 requests/day to Antigravity CLI ~20 requests/day, roughly a 98% reduction—sources: The Register, community testing.
• Three enterprise exemption paths: (1) Code Assist Standard / Enterprise license; (2) paid Gemini / Gemini Enterprise Agent Platform API key; (3) existing GCP-paid GitHub extension installs—source: Google transition announcement.
• Industry trend: "open code + private model/API" pseudo-open patterns are spreading; traditional OSS forks run standalone, but AI CLI forks without model access are barely useful—the Linux Foundation isitopen.ai framework targets exactly that gap.

FAQ: the license remains Apache 2.0; the repo is not closed. What changed is Google's first-party free/subscription endpoints. Enterprise users may run Gemini CLI and Antigravity CLI in parallel; individuals are pushed one-way into closed Antigravity. Google has a history of killing Reader, Google+, Stadia, and more—this episode further erodes trust in long-term developer commitments.

Whether you keep Gemini CLI with your own API key, move to Antigravity CLI, or switch to Claude Code / Codex CLI, production needs an always-on macOS host: laptop lids break long agent loops; personal Macs offline fail CI-triggered CLI jobs; mixing Xcode and multiple AI CLIs on one machine invites cert conflicts.

Suggested topology: before June 18, finish API key injection and Antigravity quota stress tests on a rented Mac; use launchd to keep CLI agent processes alive (see OpenClaw launchd guide); if the team also runs iOS CI, follow the hybrid CI guide to split PR builds from release archives—AI CLI handles generation and review; build machines handle compile speed.

Weak spots of oversold cloud hosts and "borrow a colleague's old Mac": (1) neighbor IO contention makes CLI agent validation scripts time out; (2) SSH jitter breaks multi-step agent loops; (3) no guaranteed Xcode / CLI versions, so post-migration behavior is not reproducible. For teams that must finish toolchain switches before the deadline and need stable 24/7 runs, JEXCLOUD multi-region bare-metal Macs with monthly elastic leases are usually the better fit: dedicated Apple Silicon, 120-second provisioning, Skills and API key config versioned with the repo. See pricing and order.