Google killed Gemini CLI yesterday. Not a soft deprecation — a hard shutdown that broke CI/CD pipelines, shell scripts, and IDE integrations for thousands of dev teams in a single morning.
If you haven't heard about it yet, you're probably about to.
What happened
Gemini CLI was an open-source coding agent from Google: 105,000 GitHub stars, Apache 2.0 license, 6,000 merged pull requests from hundreds of community contributors. It came with a free-tier allowance of 1,000 AI requests per day. Teams built CI pipelines around it. Devs wired it into their shell workflows. It was genuinely useful and genuinely free.
On May 19 — exactly 30 days ago — Google announced at Google I/O they were killing it. The replacement: Antigravity CLI, a closed-source Go binary.
Free-tier allowance: roughly 20 requests per day.
The math: you just lost 98% of your free quota. On 30 days' notice.
And on June 18 — yesterday — the switch flipped. Every CI/CD pipeline, every shell script, every IDE integration calling the gemini command stopped receiving responses. The Antigravity CLI also ships with an MCP config change that silently breaks existing pipelines — your workflow might fail before you even know what hit you.
The GitHub thread where Google announced all of this: 143 thumbs-down reactions. Four cheers.
The part that's actually enraging the community
It's not just that Google killed a free tool. It's how.
Those 6,000 merged pull requests? Community developers built Gemini CLI. They filed bugs, wrote features, tested edge cases — all under Apache 2.0, all under the assumption the project would stay open.
Google cited those contributions as evidence of the project's success. Then used that success to justify consolidating the tool into a closed-source, enterprise-gated product.
The contributors' code didn't disappear. It became the foundation of Antigravity CLI — a product they can no longer influence, fork, or use freely.
Arnaud Le Hors of IBM, who co-authored the Linux Foundation's Model Openness Tool framework, called out the Gemini CLI case as the framework's first public exhibit of a tool that scores well on "code availability" but fails completely on "operational independence" — meaning: you helped build it, but you can't run it on your own terms.
The open-source community has a word for this pattern. It's not printable here.
Why this matters if you're running a 10-50 person org
The pattern is predictable.
Big tech company launches a free AI tool. It's good. You integrate it into workflows, CI/CD pipelines, daily team habits. You get dependent. Then — 30 days' notice, sometimes less — the tap closes. The free tier gets gutted, the tool goes closed-source, or it becomes enterprise-only.
This has happened with Google more times than we can count. But it's not just Google. Microsoft has been tightening GitHub Copilot's free tier. OpenAI has monetized every meaningful API capability it once offered for free.
The problem isn't that these companies are evil. It's that small orgs build critical workflows on tools they have no contractual right to keep using. You're not a customer on a free tier — you're a resource. You have no SLA. You have no guaranteed notice period beyond what the company decides to give you.
Gemini CLI gave you 30 days. Some of these shutdowns give you less.
What to actually do
There are two credible paths for small dev teams right now.
Path 1: Open-source tools you actually own
OpenCode (165,000+ stars, MIT license) is the most-starred open-source coding agent right now. It supports 75+ LLM providers and local models through Ollama. You can run it entirely locally — no API keys, no usage caps, no cloud vendor between you and your dev workflow. Aider (46,000+ stars, Apache 2.0) is the other battle-tested option, particularly solid for git-integrated workflows.
Both tools work with local models. If you're already running Ollama, point them at Qwen3, DeepSeek, or Gemma 4 and pay zero per-token fees. Your free tier becomes: forever, as long as you have electricity.
Path 2: Pay for what you use, with a contract
If you need cloud AI inference for your tooling, pay for it properly. Even $20–50/month for an actual API key gives you pricing guarantees and a real customer relationship. At that scale, every major provider has change-notification policies and SLAs. You're a customer, not a charity case.
The thing to avoid: building anything critical on free tiers from platforms where you are the feature.
What we audit for
In the last few months we've helped several small teams map their "invisible AI dependencies" — the places where a free tool slipped into a critical workflow and nobody noticed until it broke. The audit is usually fast. The fixes are usually simple. The alternative is finding out about the dependency the day the tool goes dark.
Gemini CLI will not be the last time this happens. Google has more AI tools in various stages of "community warmth followed by enterprise monetization." Microsoft is navigating the same arc with GitHub Copilot. OpenAI has done it repeatedly.
The lesson isn't to avoid AI tools. It's to know the difference between tools you own and tools you've been lent — and to make sure the ones keeping your pipelines alive are the former.
If you want to run a quick audit of where your team is exposed, that's exactly the kind of sprint we run.