Google Just Killed Its 100k-Star CLI. The Harness Is the Product Now

A year ago, Gemini CLI was one of the most beloved things Google had shipped to developers in a decade: an open-source terminal agent that pulled in more than 100,000 GitHub stars and over 6,000 merged pull requests from a community that actually liked using it. On June 18, 2026, Google switched off its lights for everyone who isn't paying. Gemini CLI and the Gemini Code Assist IDE extensions stopped serving consumer requests — AI Pro, Ultra, and the free tier — and new GitHub installations ceased. A tool people built workflows around got a one-month eviction notice, and the month is up.

The replacement is Antigravity CLI, and Google's framing of it is the tell. This isn't presented as a faster, cleaner Gemini CLI. It's the terminal face of Antigravity 2.0, Google's agent-first development platform — a Go-based command-line interface and a desktop app that share one unified agent harness built for "complex, multi-agent orchestration." The pitch is no longer here is a smart model you can talk to in your terminal. It's here is a system that runs fleets of agents, and the CLI is one window into it.

What you gain, and what you give up

Google did the migration politely, which is to say it kept the parts developers would riot over losing. Agent Skills, Hooks, Subagents, and Extensions all carry over — the Extensions are now reborn as Antigravity plugins. You can still get a quick answer, scaffold a project, or manage cloud infrastructure from the prompt. The headline performance claim is that Antigravity is co-optimized for Gemini 3.5 Flash, which Google says runs up to 12× faster inside the harness thanks to inference tricks that only work when the model and the orchestration layer are designed together.

That last detail is the whole strategy in miniature. The speed doesn't come from a better model in isolation; it comes from welding the model to the harness so tightly that the combination outruns the parts. It's also why Google can afford to retire a 100k-star project: the moat was never the CLI. The moat is the orchestration layer that the CLI now plugs into.

The giving-up side is real, and Google admits it. Antigravity CLI does not have full feature parity with Gemini CLI at launch. Anyone who scripted their daily work against the old tool's exact behavior is now testing a new surface that mostly — but not entirely — matches. For a community that adopted Gemini CLI precisely because it was hackable and stable, "mostly" is a tax. The one escape hatch: organizations holding paid Gemini Code Assist licenses keep full Gemini CLI access unchanged, which neatly tells you who Google is willing to inconvenience and who it isn't.

A deprecation cluster, not a one-off

June 18 isn't landing in isolation. Google is pruning the same week on multiple fronts: its image-preview models go dark on June 25, and video-generation models go offline on June 30. Read together, the moves look less like routine housekeeping and more like a deliberate consolidation — Google forcing developers off a sprawl of individually-addressable endpoints and onto a smaller set of platforms it controls end to end. The era of "here's an API for each capability" is giving way to "here's the platform; the capabilities live inside it."

For developers, the practical message is blunt. If your CI pipeline, your local scripts, or your team's tooling called Gemini CLI on a free or Pro tier, those calls started failing on June 18. The migration path runs through Antigravity CLI or Antigravity 2.0, with documentation and video walkthroughs Google has been publishing since the transition was first announced on May 19. Nobody was ambushed — but a month is not long to re-plumb infrastructure you'd assumed was permanent, which is the recurring lesson of building on someone else's free tier.

The bigger shift Google is betting on

Strip away the brand names and Antigravity is Google staking out the same ground the rest of the industry is racing toward: the harness as the product. The last two years trained developers to think the model was the thing — pick the smartest one, prompt it well, ship. The frontier labs have quietly moved past that. What increasingly decides whether an agent is useful isn't raw model IQ; it's the orchestration around it — how it spawns subagents, recovers from failure, holds context across a long task, calls tools, and runs many jobs in parallel without stepping on itself. That layer is where the differentiation now lives, and it's far stickier than a model, because models get leapfrogged every few months while a harness people have built their workflows on does not.

Google calling its platform Antigravity and pushing developers toward asynchronous multi-agent workflows is a statement that it intends to compete on that layer, not just on Gemini's benchmark scores. The cost is the 100k-star CLI it had to kill to get there, and the goodwill of a community that has to relearn a tool it already loved. Whether the trade pays off depends on a question every platform shift eventually faces: did developers stay for the model, or for the harness? Google just bet the harness — and made sure there's no longer an easy way back to find out.