2026 Google Gemini CLI Policy Change: From 100k Stars to Antigravity — Why Developers Call It a Bait-and-Switch

In June 2025 Google released Gemini CLI as a TypeScript terminal coding assistant under Apache 2.0, inviting external contributions from day one. Within a year the repository passed 100,000 GitHub stars and accepted more than 6,000 merged pull requests—a scale rare even among developer CLIs. Contributor Andrea Alberti had a 27-commit PR merged on the same day the policy landed, then asked publicly whether volunteers had been building features for a codebase that would ultimately serve enterprise customers first.

The pivot arrived at Google I/O on May 19, 2026: Google announced the closed-source Antigravity CLI and stated that starting June 18, 2026, free users, Google AI Pro and Ultra subscribers, and most Gemini Code Assist for Individuals paths would no longer be served by Gemini CLI or its related IDE extensions. Discussion #27274 filled with downvotes within hours; a pinned comment captured the mood in four words: “As always, Google being Google.”

• Pain point 1 — labor versus reward: the community invested review cycles, bug fixes, and features in an open repo while service access was repriced behind enterprise licensing.
• Pain point 2 — hollow “open source” semantics: you can still fork the code, but without Google authentication and model APIs the toolchain is barely usable day to day.
• Pain point 3 — replacement product gaps: Antigravity’s free tier shrank sharply; ACP support, project memory, and docs feel incomplete; Pro users on Reddit report hitting limits after only 6–7 prompts.
• Citable data: free-tier daily requests fell from roughly 1,000/day on Gemini CLI to about 20/day on Antigravity—a roughly 98% drop by community and press estimates (verify against your account console).

Google’s stated rationale is consolidating effort into one platform built for the multi-agent era. Community pushback focuses on asymmetry: Gemini Code Assist Standard and Enterprise customers may keep using Gemini CLI—and can run it alongside Antigravity. If the migration were purely technical, why do enterprise tenants avoid the forced binary choice? The New Stack summarized it cleanly: individuals get pushed into closed Antigravity, while enterprise license holders and paid API key users stay on the original path.

Per Google Developers Blog, use the matrix below to self-check before the June 18 deadline (always confirm against official docs):

Antigravity CLI is closed source—a sharp contrast with Gemini CLI’s open license. Reported gaps include unclear ACP (Agent Client Protocol) support, missing project memory (Markdown files), odd Ctrl+C behavior, and rushed documentation. Integrators such as Dynatrace, Elastic, Figma, Shopify, and Stripe face migration reviews too; individual developers are not the only stakeholders under pressure.

Google emphasizes that the Gemini CLI repository remains Apache 2.0 and that the community can fork and submit PRs. FOSS Force reporter Christine Hall countered more bluntly: Google “did not change the open source license—it turned off the infrastructure that made the tool useful.” Traditional open source assumes a fork can run independently; an AI CLI’s “run” depends on proprietary models and quota systems controlled by the vendor.

At the May 2026 North America Open Source Summit, IBM’s Arnaud Le Hors used this episode to promote isitopen.ai (Model Openness Tool), scoring AI tools on transparency, reproducibility, and usage rights rather than LICENSE file text alone. Teams already wiring agents inside a GitHub agent execution workspace should treat orchestration as one layer and execution plane plus model egress as a separate vendor-risk review.

• Citable data: discussion #27274 saw 31+ downvotes concentrated on the policy announcement—a community sentiment signal, not a scientific sample.
• Historical context: Google Reader, Google+, Stadia, and a long product graveyard deepen skepticism about long-term commitments.
• Fork feasibility: you can fork the CLI technically, but without model API access daily developer value drops unless you bring your own Vertex or AI Studio paid key.

Vendors increasingly use open CLIs as acquisition and community-building surfaces, then sacrifice individual users and external contributors when monetization tightens. Meanwhile standards such as agentskills.io make Skills, Hooks, and MCP configs portable—if your procedures live only on Google’s closed endpoints, one policy change breaks the entire pipeline. Safer architecture: version SKILL.md and AGENTS.md in Git (see our Agent Skills guide), treat the model vendor as a swappable backend, and keep the 24/7 execution host inside a tenant boundary you control.

Some developers already canceled Google subscriptions and moved to Claude Code, Codex CLI, or a “Gemini CLI plus self-funded API key” combo. For Mac-first teams the terminal CLI is only the control plane; the expensive part is persistent gateways, Xcode builds, and long-lived connections—the same theme as our Hermes Agent install guide: the host matters more than any single model subscription.

Before June 18, split “Gemini in my terminal” into three audits: open assets (repos, Skills, scripts), service dependencies (OAuth, quotas, model names), and execution host (laptop, shared VPS, or dedicated Mac). Shared minute-pooled macOS hosts often add bandwidth jitter, oversubscription, and dropped long connections—so any CLI migration devolves into “works by day, dies overnight.”

1. 01
   Inventory auth paths: confirm whether you log in via free Google AI, Pro/Ultra, enterprise Code Assist, or a paid API key; mark 2026-06-18 on the calendar.
2. 02
   Export portable assets: back up .cursor/skills/, MCP configs, and custom scripts; record “service open vs code open” with isitopen.ai or an internal checklist.
3. 03
   Pick a CLI backend: enterprise tenants may keep Gemini CLI; individuals can trial Antigravity, Claude Code, Codex CLI, or attach AI Studio / Vertex paid API keys to the open binary.
4. 04
   Provision a dedicated Mac on the order page: move agent gateways, Telegram bots, and self-hosted runners off sleeping laptops; size unified memory for local inference (see the ds4 96GB threshold article).
5. 05
   Console provision and 72-hour soak: SSH from the console; keep processes alive with launchd; compare pre- and post-migration script exit codes and long-connection P95 latency.
6. 06
   Wire CI with human gates: on macOS runners executing Copilot coding agent or gh-aw, keep branch protection; treat CLI swaps as infra changes requiring review.

Antigravity’s closed source and tighter quotas do not automatically fix your need for a stable macOS execution plane. Shared VPS pools rarely deliver both auditable tenant boundaries and agent long connections. If you treat the CLI as a interface swap but keep gateways on a sleeping laptop or oversubscribed host, migration gains get eaten by infrastructure. Teams that need production-grade reliability get the same macOS CLI environment as owned hardware on NUKCLOUD multi-region bare-metal Mac / cloud Mac nodes, with hardware refresh and failover handled by the platform. Start on the pricing page for hourly validation, or compare host stability in our 30-day agent host report before committing to rent versus buying a Mac Mini.