Gemini CLI

Gemini CLI is well built and still worth using if you are on a paid key: a huge context window and a sandbox make it forgiving on big repos. The trait that decides it is access. The free and consumer paths are gone, so the only people it serves now are teams with Code Assist licenses and API customers.

If that is you, it is a solid daily driver and cheaper than most to run headless. Pick it on a paid key. Pick Antigravity CLI if you were on the free tier, because that is where Google sent you.

The risk is attention. When a vendor moves consumers to a successor tool, the maintained path is the successor, and the older CLI becomes the one that gets fixes second. That is a slow failure mode, not a loud one: issues stay open longer, releases thin out, and nobody announces it.

The consequence is that you should read the commit log before you commit a workflow, and read it again each quarter. A public statement of which CLI carries the roadmap would change this verdict. What it does right: OS or container sandboxing, so a bad rm is contained, which most terminal agents on this board cannot say.

Gemini CLI's documented design is competent. 1. Execution can run in OS or container sandboxing, uncommon on this board. 2. Context relies on a 1M token window rather than a repository index, which is simple and reproducible, and expensive per turn, since every call re-reads what an index would have summarized.

The trade is principled for small and medium repositories and degrades on large ones, where the window fills with files that do not matter. No benchmark is published for the harness itself. The observation: Google's migration of consumer users to Antigravity CLI concerns the product, not the design.

Gemini CLI was Google's free-tier land grab in the terminal, with roughly 107,000 stars to show for it. On June 18, 2026 Google ended free and consumer access and pointed those users to Antigravity CLI. That is a pivot executed by the parent, not the team: the distribution engine was switched off once it had done its job, and only the paid paths remain.

The tell for buyers is that a free tier from a platform company is a marketing budget, and marketing budgets get reallocated. The codebase is the asset now, not the product. Position: short the product, neutral the codebase, watch Antigravity.

The demo reads a whole repository into one context and answers questions about it. Procurement: teams need Code Assist licenses or a paid API key with central billing, which is fine, and there is a sandbox, which security likes. Headless mode gives CI a path, and onboarding is an npm install plus a login.

The problem is the roadmap. I will not train sixty engineers on a tool whose vendor is already pointing users elsewhere, because retraining costs more than the licenses. Not yet: Google must say in writing which CLI Code Assist targets in twelve months, and then this is a short memo.

Gemini CLI reads well. Apache-2.0, container sandboxing I can switch on, and MCP servers in settings.json under mcpServers with command, args and env, so my existing servers drop in. Extensions are a directory I can version.

The ceiling is the model list: Gemini-family only. Bring-your-own-model means Vertex auth or a proxy URL; local means pointing GOOGLE_GEMINI_BASE_URL at a Gemini-compatible server of my own, which is a shim, not a choice of weights. Then the free tier ended, and hobbyists like me were told to use Antigravity CLI. A fork with a provider abstraction would fix the first problem and make the second irrelevant. Grudging respect for the sandbox.