Three days from now, on June 18, Gemini CLI stops working for most of the people who use it.
Gemini CLI is a TypeScript-based AI coding agent Google released under the Apache 2.0 license in June 2025. Over the following year it accumulated over 100,000 GitHub stars and more than 6,000 merged pull requests from external contributors - a level of community investment that Google's own announcement cited as evidence of the project's success.
Then, on May 19, 2026, Google announced it was done.
On June 18, Google will stop serving requests for free users, Google AI Pro and Ultra subscribers, and individual Gemini Code Assist users. Enterprise customers with paid Gemini Code Assist Standard or Enterprise licenses are explicitly exempted: they keep Gemini CLI, keep model updates, and keep security patches. The community that supplied the labor does not.
The replacement is called Antigravity CLI. It is closed-source. If you visit the Antigravity CLI GitHub repository, you will find a changelog, a readme, and a demonstration GIF. No source code.
This is not a story about a deprecated tool. It is a story about what "open source" means when the vendor controls the inference backend.
The license said Apache 2.0. The backend said otherwise.
When Google released Gemini CLI, it made a sensible-sounding promise: permissive license, community contributions welcome, inspectable code. The tool was designed to act as a local AI assistant supporting complex developer workflows - code refactoring, documentation generation, shell commands, file editing. It joined a growing ecosystem of AI developer tools, standing apart by placing openness and composability at the center.
The catch, obvious in retrospect, is that the "open" part was the TypeScript wrapper. The thing that made it useful - model inference - was always Google's to give or take back. Without a Google backend serving requests, the codebase becomes an archive. Running a working fork would require independent access to a compatible large language model - a non-trivial engineering undertaking that effectively ends community use of the tool as a practical matter.
Group product manager Dmitry Lyalin acknowledged in his GitHub post that the Gemini CLI repository "remains available to the community as an Apache 2.0 licensed repository with no changes." That is technically accurate. It is also the kind of statement that tells you exactly what kind of open source this was.
What the community actually built
Gemini CLI loses free-user API access on June 18, replaced by the closed-source Antigravity CLI. Developers who contributed 6,000 merged pull requests are calling it a bait-and-switch - community labor absorbed into an enterprise product.
Developer Andrea Alberti had a 27-commit pull request merged to Gemini CLI on the same day as the shutdown announcement. That is not a footnote. That is the kind of thing that makes developers recalibrate whether to contribute to the next Apache 2.0 release from a vendor who also sells enterprise software.
Rather than treating community investment as a reason to maintain the open-source commitment, Google used it as justification for the transition: the project had proven itself, and now it was time to consolidate it into a proprietary platform.
The fork path is real, but narrow
The Apache 2.0 license means the Gemini CLI codebase can be forked and continued by the community. Whether someone steps up to maintain an independent backend is an open question, but the conversation is happening.
The practical question for a team on a deadline this week is: what do you point the client at? Gemini CLI was model-agnostic in principle but wired to Google's inference in practice. A working fork needs an OpenAI-compatible endpoint, an Ollama instance, or some other backend that speaks the same protocol. That is doable for a team with the infrastructure already in place. It is a Thursday-afternoon project for a team that does not.
The alternatives that do not have this problem are tools built model-agnostic from day one. The best open-source coding agents are model-agnostic by design - you bring an API key for OpenAI, Anthropic, a local Ollama instance, or any OpenAI-compatible endpoint. This lets you optimize cost, latency, and privacy independently. Aider has operated this way since 2023. Goose was released in January 2025 by Block under Apache 2.0 and has since moved to the Agentic AI Foundation under the Linux Foundation - a real governance signal for teams that worry about a single vendor abandoning a tool.
What this means beyond Gemini CLI
The specific timeline - four weeks, no feature-equivalent replacement, enterprise carve-out - is unusual. Most vendor deprecations are slower and quieter. But the underlying structure is common: vendor opens a client layer to attract contributors, captures mindshare, then prices or closes the backend once the audience is established.
This situation is a stress test for open-source trust in the AI tooling era. When AI companies release developer tools under permissive licenses, there is an implicit contract: we are building this together, and you can rely on it. When that contract breaks - on a four-week timeline with no feature-equivalent replacement - developers have to recalibrate how much they trust the next Apache 2.0 release from a company that also sells enterprise software.
For teams that route work through AI coding agents - in the terminal, in CI, or piped into a tool like Beagle for downstream context - the lesson is the same one that applies to any piece of infrastructure: know what you cannot replace in a weekend. If the answer is "the inference backend," then the license of the client code tells you less than you think.
The fork conversation on GitHub is worth watching. But June 18 arrives in 72 hours.