Google has released Gemini CLI (command line interface), a terminal-based version of its AI assistant, with a generous free tier of up to 60 model requests per minute and 1,000 per day.

A CLI-based approach to AI coding has worked well for competitor Anthropic’s Claude Code, and Open AI has Codex CLI as an experimental project. It is a trend Google cannot afford to ignore, and Google’s Ryan Salva, senior director of product, told us in a briefing that Gemini CLI is distinctive because of its “unmatched usage limits,” which includes a 1 million token context windows, use of the Gemini 2.5 Pro LLM (large language model), and 1,000 model requests per day.

There is also a paid option which uses an AI Studio or Vertex API key and adds features including policy and governance capabilities, choice of models, running agents in parallel, and no requirement to use Gemini activity to improve Google’s products.

Advantages of the CLI include the ability to use coding assistance with any editor or IDE, rather than just those for which a plugin is available, and an option to run multiple instances concurrently. Some developers also find the CLI more efficient.

Gemini CLI runs on Mac, Linux (including ChromeOS) and Windows; and unlike Claude Code or Codex, the Windows version is native rather than requiring WSL (Windows Subsystem for Linux). Developers can configure it using a text file, gemini.md, at the root of a folder, setting context and other parameters. Gemini CLI will also automatically save context into gemini.md “when it finds details that should be longer lived,” said Tayor Mullen, senior staff engineer.

Capabilities of Gemini CLI include writing code, debugging issues, managing projects, querying documentation, and explaining code. The tool also has access to MCP (model context protocol) servers, enabling agentic AI. The utility is open source on GitHub under the Apache 2.0 license. It also integrates with the gcloud CLI, part of the Google Cloud SDK, and we presume that the company aims to monetize this product in part by steering developers towards deployment to Google Cloud. The demo we saw in a briefing was deployed automatically to Google Cloud Run. 

Google is also targeting general users including IT professionals, scientists, and anyone wanting to generate slides or images, for example.

Gemini CLI cannot yet be used with a local model, though Allen Hutchison, senior director AI developer, told us that the company hopes to use it with local models such as Gemma in future.

The company asserted that security is a key focus of Gemini CLI. Two key features are that actions are subject to approval via a prompt – with an option to “allow always” – and sandboxing. On Mac Gemini CLI uses native macOS sandboxing, also known as Seatbelt, while on other platforms it can use a container such as Podman or Docker. Security remains a concern though. MCP is known to lack sufficient security safeguards, and risks from prompt injection are inherently difficult to solve, particularly if less skilled users employ AI to write code or perform actions which they do not fully understand.

Gemini is available via an API as well as through Code Assist plugins for Visual Studio Code (VS Code) and JetBrains IntelliJ-based IDEs, and the browser-based AI Studio. Access to Gemini CLI includes a Code Assist license.

Early reaction to Gemini CLI is mainly focused on how extensive the free tier is. “Huge, and will put lots of pressure on Anthropic,” said one commentor.

That said, Gemini’s existing Code Assist plugins have a mixed reputation so far, judging by comments on the JetBrains and VS Code marketplace. The VS Code plugin has 900,000 installs but only a 2.5 star rating, with criticism such as “for code generation, this was a complete waste of time. The model behind the extension referenced functions and classes that didn’t exist, and even started generating tests for code that wasn’t part of my code base.”

An advantage of the CLI is that it no longer depends on working within VS Code, which Microsoft gears specifically towards GitHub Copilot.

Google’s Ryan J Salva’s pitch to the press at the briefing was: “we believe that these tools will dominate the way creators work in the next decade.” By giving away free AI compute the company hopes, perhaps, to achieve a strong position in this market from which it can later take advantage.