If you adopted Google Gemini CLI after its June 2025 Apache 2.0 release, you are not alone: the repo crossed 100,000 GitHub stars and absorbed 6,000+ merged pull requests from a community that treated it as a credible terminal-native agent. At Google I/O on May 19, 2026, that story changed. Google announced Antigravity CLI, said non-enterprise Gemini CLI API access ends June 18, 2026, and cut the Antigravity free tier from 1,000 to 20 requests per day on closed-source infrastructure. Critics call it bait-and-switch; FOSS Force put it plainly: they did not change the license, they shut down the infrastructure. This article is for developers who need a decision, not a hot take: a timeline, a comparison matrix, three hard numbers, a six-step operational playbook, and why stable CLI Agent and iOS CI work often lands on a dedicated remote Mac with SSH session isolation instead of a laptop that sleeps mid-session.
The controversy is not about whether open source is good. It is about what happens when a vendor trains a community on free hosted inference, then moves that inference behind a different product with a 98% quota cut. If you built scripts, CI jobs, or team habits around Gemini CLI's default endpoint, these six failures show up immediately.
Your README still says "install and run," but the default backend is gone. New contributors clone the repo, follow upstream docs, and hit auth or quota errors on June 18 because the public Gemini CLI API path no longer serves non-enterprise users.
Free-tier math no longer works for daily agent loops. Twenty requests per day on Antigravity cannot sustain an agent that reads a repo, plans, edits, runs tests, and retries. A single multi-step task can burn the entire allowance.
Trust in "Apache 2.0" as a promise collapses. The license still allows fork and modify, but without a compatible hosted model route the upstream binary becomes a shell that expects keys you were not budgeting for.
Enterprise escape hatches do not help indie devs or small teams. Gemini Code Assist Standard and Enterprise, plus paid API keys, continue. Everyone else must replatform or pay before the next sprint starts.
CI and long SSH sessions become fragile on a laptop. When you switch to Claude Code or Codex CLI with your own keys, session drops, sleep, and disk contention on a MacBook become the new bottleneck, especially next to Xcode builds.
Due diligence now requires "service openness," not license text alone. At Open Source Summit in May 2026 the Linux Foundation promoted isitopen.ai to score whether AI products are open in practice, not only on paper, and Gemini CLI is the case study everyone cites.
These symptoms share one root cause: open code does not imply open service. The next sections separate what Google actually changed from what the repository still guarantees, then map realistic replacements.
June 2025. Google open-sourced Gemini CLI under Apache 2.0. The launch narrative emphasized terminal-native agent workflows, tight integration with Google's models, and a generous free tier for individual developers. Community velocity followed: stars climbed past six figures and thousands of external contributors landed fixes, docs, and integrations. Many teams treated the default Google-hosted route as production-adjacent for personal automation and internal experiments.
May 19, 2026 — Google I/O. Google announced Antigravity CLI as the forward path for agentic coding in its stack. In the same beat, it communicated that Gemini CLI API access for non-enterprise users would end on June 18, 2026. Antigravity is not a community continuation of the open repo; it is a separate, closed-source product line with its own quota model. The free allowance on Antigravity was described as dropping from roughly 1,000 requests per day to 20 per day, a 98% reduction that reads less like tuning and more like a funnel into paid tiers.
June 18, 2026 — operational cliff. After this date, developers without enterprise Gemini Code Assist or a paid API key cannot rely on the same hosted path that powered the 2025 growth curve. The GitHub repository remains public; issues and PRs may continue; but the default "it just works" cloud attach for individuals does not. That is why FOSS Force summarized the episode as shutting infrastructure while leaving the license untouched, and why social channels label it bait-and-switch: the invitation was "star the repo and build on our API," the withdrawal is "use Antigravity or enterprise."
"They didn't change the license. They shut down the infrastructure." — how the FOSS community framed the Gemini CLI transition in June 2026
Enterprise customers are explicitly out of scope for the panic: Gemini Code Assist Standard and Enterprise retain supported access, and organizations with paid Gemini API keys can wire those into compatible clients. The pain concentrates on indie developers, startups without assist contracts, and open-source maintainers who marketed Gemini CLI as the zero-friction on-ramp.
Note: Dates and quotas here reflect Google's May 2026 I/O communications and subsequent reporting. Treat Antigravity limits as product policy that can change; bake monitoring and key rotation into any production workflow.
Choosing a path after June 18 is a procurement and architecture decision, not a git pull. Use the matrix below to see what each option optimizes for. None of them is "wrong," but they assume different budgets, compliance needs, and tolerance for vendor-controlled inference.
| Option | Code license | Hosted inference | Typical free quota | Best fit |
|---|---|---|---|---|
| Gemini CLI (upstream repo) | Apache 2.0 | Non-enterprise API ends Jun 18, 2026 | Was ~1000 req/day via Google route; ending | Forks, custom keys, study of Google's CLI design |
| Antigravity CLI | Closed source | Google-hosted, new product surface | ~20 req/day free tier (reported 98% cut) | Teams already on Google enterprise stack |
| Gemini Code Assist + paid API | N/A (commercial) | Contracted | Enterprise / paid key billing | Regulated orgs with existing Google agreements |
| Claude Code / Codex CLI | Vendor CLI (varies) | Anthropic / OpenAI endpoints | Plan-specific; BYO key supported | Terminal agents with predictable third-party billing |
| Self-hosted + own keys | Your fork / scripts | Any provider you pay for | You set spend caps | Long-running agents, CI, multi-repo automation |
Read the table once more with the open code vs open service lens. Gemini CLI still teaches how Google wanted terminal agents to behave: tool calls, repo awareness, tight edit loops. Antigravity trades community inspectability for a managed funnel. Claude Code and Codex CLI are the pragmatic exits if you want mature terminal agents today, provided you accept their vendors' terms and bring your own spend controls.
If your workflow already spans IDE agents and terminal agents, this policy change is also a reminder that the agent runtime and the model endpoint are separable. The runtime can stay on a stable Mac node while endpoints swap. That is where SSH-isolated remote Mac practice stops being optional for teams that run agents for hours, not minutes.
Apache 2.0 answers a narrow question: may I read, modify, and redistribute this source? It does not answer whether tomorrow's inference URL still accepts my token, whether training data is disclosed, or whether weights can be audited. That gap is why the Linux Foundation highlighted isitopen.ai at Open Source Summit in May 2026: a public checklist mindset for AI products that separates license badges from operational openness.
Applied to Gemini CLI, the checklist fails on service continuity for non-enterprise users even when it passes on source availability. You can fork the client, but if the model route you depended on vanishes, your fork is a research artifact until you plug in another backend. Community commentary grouped this with older "open washing" patterns: star counts and contributor graphs used as marketing proof, followed by a commercial product that captures the habit without capturing the governance.
Enterprise teams were never the audience for the 1,000-requests-per-day hobby tier. They already route through assist SKUs, VPC constraints, and invoiced API keys. For them, June 18 is a communications event, not a cliff. For everyone else, the lesson is to document two dependencies in every agent design: the client (can we fork it?) and the inference contract (will the endpoint exist at our price point next quarter?).
# Minimal sanity check before you bet a team workflow on a CLI
# 1) Where does the binary send traffic by default?
strings "$(which gemini)" 2>/dev/null | grep -E 'https?://' | head
# 2) Can you override with YOUR key and YOUR base URL?
export GEMINI_API_KEY="${GEMINI_API_KEY:-}"
export ANTHROPIC_API_KEY="${ANTHROPIC_API_KEY:-}"
# Document fallbacks in repo CONTRIBUTING.md — not only in a private wiki
# 3) Log quota headers or billing alerts daily
# If free tier drops 98%, your CI will fail silently until someone notices
Tip: When evaluating Antigravity, Claude Code, or Codex CLI side by side, score them on endpoint portability and quota observability, not on launch keynote demos. Demos always have budget; your cron job does not.
Policy fights on social media are loud; capacity planning is quiet. These three figures are the ones worth citing in an architecture review or a postmortem doc because they change throughput math, not just sentiment.
Translate the numbers into workload terms. A medium repo refactor might take dozens of model round-trips when you include planning, file reads, test runs, and fix-up passes. Twenty requests per day is not a slower refactor; it is a different product category. Teams that conflate "we use a CLI" with "we have agent capacity" will feel that difference on the first Monday after cutoff.
Hardware follows the same arithmetic. Terminal agents keep state in long sessions; iOS CI keeps DerivedData hot; both want a machine that does not sleep. A laptop sharing one unified memory pool between Xcode, Chrome, and two agent sessions is where swap shows up first. Moving agents to a dedicated node is less about Google drama and more about decoupling uptime from your daily driver.
Use this sequence as a migration runbook. Order matters: inventory before rip-and-replace, keys before CI, stable runtime before optimizing model choice.
Inventory every Gemini CLI entry point. Search dotfiles, CI YAML, internal templates, and onboarding docs for default binaries, env vars, and commented "quick start" blocks. Mark anything that assumes Google's non-enterprise hosted route.
Classify workloads by quota sensitivity. Separate interactive coding (burst requests) from scheduled agents (steady burn). Map each to Antigravity, enterprise assist, paid Gemini keys, or a move to Claude Code / Codex CLI with explicit spend caps.
Bring your own keys and document base URLs. Store secrets in CI vaults, not repos. For each client, record which endpoint it hits, who pays, and what happens when a key rotates. Treat the Apache repo as client code, not as a SLA.
Add quota and billing observability. Daily checks on remaining free tier, invoice anomalies, and 429 rate limits. After a 98% cut, silent failure is the default mode.
Run terminal agents on a stable runtime. Long sessions, large git trees, and parallel tool calls punish sleep, thermal throttling, and VPN drops. Prefer a dedicated Mac with SSH session isolation over a laptop that travels.
Re-score vendors with isitopen.ai-style questions. Before the next bet, ask: Can we change inference without rewriting the client? Can we audit data handling? What happens to CI on day one of a policy change?
After step six, the weaknesses of "keep everything on the laptop with whatever CLI was free last month" are obvious. Sleep interrupts agent state; thermal throttling stretches iOS builds; switching providers mid-sprint without an isolated environment invites port and cache collisions. Relying on another vendor's teaser free tier repeats the same trap with a different logo. For teams that need CLI agents and iOS CI to stay up across policy shocks, putting a high-performance Mac on the network as a dedicated node is usually cheaper than emergency migration. NodeMini Mac Mini cloud rental fits that role: predictable SSH access, isolated disks for DerivedData, and hardware that does not close when the lid shuts. It pairs with the agent hosting patterns in remote Mac AI coding and session isolation, iOS CI on SSH-first Mac nodes, and transparent rental rates. Setup and regions are documented in the help center.
Yes for the code: Apache 2.0 still applies and the repository remains public. No for the old default experience: Google's hosted Gemini CLI API path for non-enterprise users stops on June 18, 2026. You must attach your own paid Gemini API key, move to an enterprise assist contract, or point a fork at another model backend. Treat "clone and run" docs written in 2025 as stale until someone updates the endpoint section.
Not in the open-source sense. Antigravity is a closed-source Google product with a much smaller free tier (about twenty requests per day versus roughly one thousand). It may cover similar UX goals, but you cannot inspect or fork it the way you could Gemini CLI. Teams needing predictable throughput should budget paid tiers or migrate terminal workflows to Claude Code or Codex CLI with explicit API keys and monitoring.
Run them on hardware that stays awake, with SSH session isolation so a dropped laptop Wi-Fi does not kill a four-hour agent run. Pair that with the same node for iOS CI if you ship Apple platforms. NodeMini dedicated cloud Macs are built for that pattern; compare specs on the rental rates page and follow SSH provisioning in the help center. For architecture background, see CLI agent SSH session isolation on a remote Mac.