Gemini CLI Dead — Google's Open-Source Bait-and-Switch

If you ran a gemini command this morning and got HTTP 410 Gone, you are not alone and it is not a bug. As of today — June 18, 2026 — Google has killed Gemini CLI for every Google AI Pro, Ultra, and free-tier user. No grace period. No backward-compatibility shim. The replacement is Antigravity CLI, a closed-source Go binary called agy that breaks four things silently and caps your compute on a weekly cycle that nobody can inspect.

TL;DR

• What: Gemini CLI returns HTTP 410 Gone today for all non-enterprise users — hard cutover, no grace period
• Breaking changes: Default model, streaming format, state directory, and exit codes all changed silently
• Quota regression: Daily request limits replaced by opaque weekly compute cap — users report multi-day cooldowns
• Action: Audit every CI pipeline and script calling gemini right now; evaluate Claude Code, Aider, or OpenCode as alternatives

Gemini CLI Cutover — What Happened

Google announced Antigravity CLI at Google I/O on May 19, 2026, giving developers roughly 30 days to migrate before today’s hard cutoff. That timeline alone tells you how Google views this: not as a transition to manage carefully, but as a fait accompli to announce and enforce.

The backstory makes it worse. Gemini CLI was an Apache 2.0 project. Within roughly a year it accumulated 104,000+ GitHub stars and merged 6,000+ pull requests from external contributors. Those contributors built tooling, filed issues, fixed bugs, wrote integrations. They made Gemini CLI viable. And now their reward is a closed-source replacement where the GitHub repo contains a changelog, a README, and a GIF. No source code. No way to inspect what agy does with your prompts, your credentials, or your agent state.

Organizations on Gemini Code Assist Standard or Enterprise licenses and paid Gemini Enterprise Agent Platform API keys keep uninterrupted access to both tools. Everyone else gets the door. That enterprise carve-out is not a detail — it is the entire strategy. Google used open-source contributors to validate the product, then pulled the rug once the enterprise contracts were signed.

Here is the official migration path:

# Install Antigravity CLI via Go binary
go install google.dev/antigravity/cmd/agy@latest

Simple enough — until you realize a binary rename is the least of your problems.

Why This Matters

Four Silent Breaking Changes

If your team is treating this as “rename gemini to agy and move on,” you will break things. Here is what changed without a deprecation warning:

Default model shifted. Gemini CLI defaulted to gemini-1.5-pro. Antigravity CLI defaults to gemini-3-pro. Any script relying on implicit model selection now hits a different model with different behavior, different token limits, and different cost characteristics. If you hardcoded the model flag, you are fine. If you did not — and most quick scripts do not — your outputs just changed.

Streaming format changed. The old --stream flag emitted plain text deltas. Antigravity emits SSE-formatted events. Any parser consuming streaming output line-by-line is now broken. This one is particularly nasty in CI pipelines where streaming output feeds into downstream processing.

Agent state directory moved. State moved from ~/.gemini/agents/ to ~/.antigravity/agents/. Persistent agent configurations, conversation histories, and custom tool definitions do not automatically migrate. You need to copy them manually or symlink.

Exit codes changed semantics. Gemini CLI exited 0 on tool-use failures. Antigravity exits non-zero. If your CI pipeline uses exit codes to determine success — and it should — this is the one that will cause the most confusing failures, because the tool technically worked but your pipeline marks it as failed.

These four changes are not in the migration guide’s quick-start section. They are buried in the detailed changelog. If you are migrating under time pressure today, bookmark this list and check each one against your scripts.

The Quota Regression

This is the change that will hurt the most over time. Gemini CLI used daily request limits — though Google never formally published the exact cap. Antigravity CLI replaces that with a weekly compute-based cap. Google has not published exact limits here either — the official docs reference only “meaningful quota, refreshed weekly” for free and Pro users.

What we know from user reports paints a rough picture: multiple people have reported hitting their weekly quota quickly, in some cases after just a couple of requests. Pro users describe exhausting the weekly allowance after roughly two thousand lines of generated code, then facing cooldowns of up to 168 hours — a full week. These are community reports, not official Google metrics, but the pattern is consistent enough across sources to take seriously.

The shift from daily to weekly resets is consequential for anyone using CLI-based AI tools as part of their daily workflow. With daily limits, you knew your budget reset overnight. With weekly compute caps, a heavy Monday morning can lock you out until the following Monday.

If you are on the free or Pro tier, track your Antigravity usage aggressively in the first week. Log every request and its compute cost. You need to understand your burn rate before you hit the wall, not after.

Closed Source Means No Recourse

When Gemini CLI had a bug, you could read the source, understand the issue, and submit a fix. Six thousand contributors did exactly that. With Antigravity CLI, you file a GitHub issue on a repo that contains no code and wait. For teams running AI-assisted coding in production workflows, that loss of transparency is not philosophical — it is operational. You cannot audit what the tool sends to Google’s servers, you cannot patch a blocking bug on your timeline, and you cannot fork it if Google makes decisions you disagree with.

Developers exploring alternatives are gravitating toward a few options. Claude Code has not forced a binary migration or revoked access to a working tool, and its CI integration story is solid. Aider remains fully open-source with bring-your-own-key flexibility — developers in GitHub discussions and the Gemini CLI shutdown thread are actively sharing migration scripts and comparing alternatives. OpenCode ships under MIT with broad provider support. None of these are perfect substitutes — each has tradeoffs in model access, feature depth, and ecosystem maturity — but the common thread in community migration threads is clear: developers want a CLI layer that no single vendor can unilaterally revoke.

The Take

I have seen this playbook before, and calling it a “migration” is generous. Google accepted 6,000 community pull requests on an Apache 2.0 project, used that momentum to prove market demand, then shut the door and handed enterprises the keys. The 30-day window between announcement and cutoff was not a transition period — it was a countdown to dependency.

If your team had Gemini CLI calls in production and did not act before today, you are discovering this the hard way right now. If you are re-evaluating your CLI tooling stack, that is the right instinct. Not because Antigravity CLI is necessarily bad tooling — it may stabilize into something solid — but because Google just demonstrated exactly how much control they are willing to exercise over your workflow, and they did it with zero backward compatibility.

My concrete recommendation: do not just migrate to agy and call it done. Migrate your critical paths, then immediately start testing an open-source alternative for your highest-volume use cases. Aider for its BYOK flexibility, OpenCode for its provider breadth, Claude Code if you are already in Anthropic’s ecosystem. The goal is not to abandon Google’s models — they are good models — but to make sure your CLI layer is something no single vendor can take away from you on 30 days’ notice.

The 6,000 contributors who built Gemini CLI into something worth replacing learned this lesson for you. Do not waste it.