The most revealing line on Google’s gemini-cli repository is not a feature. It is a warning. Two of them, in fact, pinned near the top of the README: unpaid tier and Google One users were “replaced by Antigravity CLI on June 18th, 2026,” and “Gemini CLI has not been optimized for Gemini 4 Argon. To use Gemini 4 Argon, please use Antigravity.”

Read that twice. Google shipped a flagship model, Gemini 4 Argon, and its own open-source terminal agent is not tuned for it. The tool that was supposed to be “the most direct path from your prompt to our model,” in the README’s phrasing, now routes you somewhere else for the newest model. That is a strange thing to leave standing on a repository with tens of thousands of stars. It is also, read correctly, a fairly honest signal about where Google thinks the agent fight is being won.

What the repo actually still is

Strip out the deprecation banners and Gemini CLI remains a genuinely capable piece of software. It is Apache 2.0, installable via npx @google/gemini-cli, Homebrew, MacPorts, or a conda environment for restricted setups. It runs three release channels: nightly builds at 00:00 UTC daily, preview releases Tuesdays at 23:59 UTC, and stable promotions Tuesdays at 20:00 UTC. That cadence is disciplined, and it is the kind of thing you only build for a tool you expect people to depend on.

The feature surface is broad. Multimodal generation from PDFs, images, or sketches. MCP server integration for pulling in tools like Imagen, Veo, and Lyria. Non-interactive scripting via --output-format json and stream-json, which matters for anyone wiring an agent into CI. Three auth paths: Google OAuth with a GOOGLE_CLOUD_PROJECT for paid Code Assist licenses, a GEMINI_API_KEY for usage-based billing, and Vertex AI for enterprise teams that need the compliance story. Context files via GEMINI.md, checkpointing, token caching, sandboxing, trusted folders.

None of that is the news. The news is that all of it now sits behind a sign pointing at the exit.

The tell is in the model gating

Here is the part worth sitting with. A terminal agent is not a hard product to keep current. The model endpoint changes, you update a default, you ship a stable release on Tuesday. Google has the release machinery to do exactly this. So the decision not to optimize Gemini CLI for Gemini 4 Argon is not a resourcing accident. It is a choice to consolidate.

That consolidation has a name: Antigravity. The README mentions it four times, always as the recommended path. Free and Google One users were migrated in June. Argon users are told to go there. The message to the open-source community is that the Apache 2.0 repo is being kept alive as a reference implementation and a compatibility surface, not as the place Google wants its best coding experience to live.

There is a defensible argument for this. Running one agent product well beats running two half-well, and Antigravity presumably bundles the harness, the tooling, and the model tuning that a bare CLI cannot. But it comes at a real cost. The people who adopted Gemini CLI did so because it was open, scriptable, and theirs to fork. A closed successor collects none of that loyalty automatically, and the README’s own framing (“not optimized,” “replaced”) tells those users their investment is now maintenance.

What it means for builders

The practical read for anyone with Gemini CLI in a pipeline: it still works, and the Apache 2.0 license means it keeps working as long as the API behind it does. The stream-json output and MCP integration are not going anywhere on your side of the fence. But you are now building on a surface Google has publicly deprioritized, and the model you most want to call, Argon, is explicitly out of scope. That is a migration decision you should make on your own timeline rather than waiting for a banner to make it for you.

The broader signal is about how the big labs are treating the terminal. For two years the CLI was the proving ground for agentic coding: cheap to ship, easy to instrument, beloved by the developers who would evangelize the model. Google is now saying the proving ground has served its purpose and the real product is the integrated environment. That is a bet that developers will follow the best experience rather than the most open one. It is the same bet every incumbent makes, and it is not always right.

Watch what happens to the commit cadence on the repo over the next quarter. If the Tuesday stable releases keep landing with real feature work, the deprecation is softer than it looks. If they thin out to dependency bumps and security patches, the README banners were the whole announcement, and the open-source terminal agent at Google is effectively in hospice.