Introducing Gemini CLI

Google’s Gemini team is rolling out Gemini CLI, a command-line interface that turns Gemini Code Assist into an agent-like workflow for editing files, running commands, and integrating tools—while offering a surprisingly generous free tier. The core pitch is straightforward: developers can log in with a personal Google account, get a free Gemini Code Assist license, and then use Gemini 2.5 Pro with a 1 million context window for 60 requests per minute and 1,000 requests per day at no charge. For teams already using command-line LLM tools (and frustrated by fast-billing limits), that free rate structure is the headline.

Gemini CLI builds on the earlier Gemini Code Assist offering and adds a workflow centered on a repo-local “Gemini MD” file that acts like a rules/context document. After installing via npx, users authenticate either with a Google login or a Gemini API key from Vertex AI / Gemini AI Studio. Once authorized, the CLI can generate and modify project files directly. In a live walkthrough, it creates a Tailwind-based HTML/JS file, then iterates on a more substantial request: a landing page for a cat café in San Francisco with menu content and an “about” page. The system updates files on the fly, asks for permission for actions (with options like “allow once” or “always allow”), and tracks context usage as the project grows—dropping to roughly 98% remaining context during the expansion.

A key differentiator is tool integration through MCPs (Model Context Protocols). The CLI supports adding MCP servers so the model can query external capabilities at the command line level. The walkthrough adds the Hugging Face MCP server, then uses it to search for model components like rerankers and to locate Spaces. The results include multiple variants (e.g., Onyx and GGUF forms), and the CLI can also surface model details and paper search. This makes the CLI less about “chatting” and more about orchestrating concrete actions across services.

The workflow also includes a memory feature that saves user-provided facts into the repo’s Gemini MD context, so later prompts can reuse details (such as a port number). When the user asks for a Flask backend, the CLI creates directories, installs dependencies, sets up a virtual environment, runs the app, and then reports execution metrics like input/output tokens and other internal token counts. It also supports rollback of actions, giving users a safety valve when automated changes go in the wrong direction.

Overall, Gemini CLI positions command-line agent workflows—file editing, command execution, and MCP tool access—inside a Google-backed stack with a free tier that’s unusually large for developers who want to experiment without immediately hitting cost ceilings. The practical takeaway: with Gemini CLI, a single authenticated session can generate a working project, integrate external knowledge via MCPs, and keep state in a repo-local rules/context file, all while staying within clearly defined daily and per-minute limits.