AI Agents

Gemini CLI was retired — what to use instead

On 18 June 2026 Gemini CLI stopped serving requests. If you are here because a script that worked last week now does not, this page is the short version: what happened, what the replacement actually is, and which agents are still yours to run.

What happened

Google announced at I/O on 19 May 2026 that Gemini CLI and the Gemini Code Assist IDE extensions would stop serving requests on 18 June 2026 — for Google AI Pro and Ultra subscribers and for free Code Assist users alike. The replacement is Antigravity CLI, a closed-source rewrite in Go.

There was no soft deprecation, no grace period on the free tier, and no automatic migration. The Apache-2.0 source of the old CLI still exists; the service it talked to does not.

Why people took it badly

Gemini CLI was Apache-2.0 and had passed 100,000 GitHub stars. Google merged more than 6,000 pull requests from outside contributors into it, and its own announcement cited that community as evidence of the project’s success — before consolidating that success into a closed-source replacement.

Whether Antigravity CLI is a better tool is a separate question, and it may well be. The complaint is not about quality. It is that thousands of hours of donated work went into something that was then closed, and that the people who donated it got a shut-off date rather than a migration path.

The transferable lesson

An open-source client is not the same thing as an open stack. Gemini CLI’s code was open; the model endpoint it depended on was not, and that is the half that got switched off. When you pick a replacement, the question that matters is not “is the repo open?” but “if this vendor changed its mind tomorrow, would my setup still run?” That is only true when you can point the agent at a provider of your choosing — and, ideally, run it somewhere you control.

The open-source agents still standing

There is no drop-in replacement, and anyone telling you otherwise is selling something. These are the terminal agents worth looking at, all of them model-agnostic — you bring a key rather than inheriting a vendor’s free tier:

AgentLicenseLanguageStarsOne-click here
OpenCodeMITTypeScript201,935
GooseApache-2.0Rust53,577
PiMITTypeScript98,265
OpenClawMITTypeScript379,000Yes
Hermes AgentMITPython195,000Yes
Grok BuildApache-2.0Rust26,140

Star counts and licences read from each project’s repository on 27 August 2026.

OpenCode is the largest by a distance and installs from almost any package manager. goose is Apache-2.0, was built at Block and now sits with the Linux Foundation, which is a meaningfully different governance story after what just happened. Pi is smaller and sharper — a toolkit rather than a product.

What running one yourself actually costs

“Self-host it” sounds heavier than it is. We measured the agents we host on a 1 vCPU / 1 GB server — the cheapest tier we could provision — taking readings thirty seconds after a live model request:

AgentMemory at idleInstall time
OpenClaw278 MBunder a minute
Hermes Agent120 MB5 minutes
Claude Codenothing residentunder a minute
Codexnothing residentunder a minute

Claude Code and Codex show “nothing resident” because they are invoked per session rather than run as a service — an idle box carries only the operating system. OpenClaw and Hermes hold a gateway process open, which is what lets them keep working while you are away.

Full numbers and method are in our VPS hosting for AI agents comparison, where every infrastructure figure was measured on a real server rather than read off a README.

If you are migrating today

  1. Find what actually broke. CI jobs and cron scripts calling gemini fail loudly; a shell alias fails quietly and you will notice it a week later.
  2. Pick on governance, not features. After this, “who can switch it off?” deserves more weight than a feature checklist.
  3. Bring your own key. A model-agnostic agent turns a vendor decision into a config change instead of a rewrite.
  4. Put it somewhere that is not your laptop if it runs on a schedule or answers webhooks. A 1 GB server is enough — see the numbers above.

Related