Gemini CLI - FREE? Claude Code by Google | First Look and NextJS RAG App Test

Gemini CLI lands as a free, open-source “developer-terminal” layer for Google’s Gemini Code Assist, pairing a ChatGPT-like coding workflow with a generous usage tier and built-in tooling such as MCP. In a hands-on build, it quickly scaffolds a NextJS + TypeScript RAG-style app that can upload documents, index them for retrieval, and answer questions grounded in the uploaded text—despite rough edges around speed, configuration, and styling.

The core pitch is straightforward: Gemini CLI brings Gemini directly into a terminal experience, positioned as a competitor to tools like “code” assistants. Google’s free tier is described as 60 model requests per minute and 1k model requests per day, with the project released as open source. Under the hood, Gemini CLI uses Gemini Code Assist, which—on the free plan—collects prompts, related code, generated output, code edits, and usage/feedback data to improve Google products. Users can opt out via the Gemini Code Assist for individuals setup.

Installation is presented as simple: install the required packages and point to the Gemini CLI GitHub repository. Once authenticated, the interface resembles familiar coding assistants: it detects a Gemini MD file, shows available “magic commands,” and runs tools such as Google Search to pull documentation. The workflow also includes agent memory management (showing current context from Gemini files), usage stats (input/output tokens and duration), and MCP enablement—meaning the CLI can integrate with external tool servers.

In the build demo, the user targets a document-chat application: a RAG system designed to be easy to deploy, using Ollama for local model serving and a NextJS/React/TypeScript stack with Tailwind CSS and Shadcn UI. The CLI produces a step-by-step implementation plan, then executes it: creating package.json, scaffolding the NextJS structure, generating TypeScript config (including editor support for VS Code and Cursor), adding Tailwind configuration, and wiring up file upload plus chat logic.

The session isn’t smooth. The CLI hits errors around missing imports (notably Formidable), and Tailwind configuration issues lead to a running app with little to no styling. There are also signs of throttling and model switching—responses slow down and the system shifts from “Gemini 2.5 Pro” to “Gemini 2.5 Flash,” which the narrator attributes to free-tier constraints. Even so, the application comes up and functions: a text file uploads successfully to the project’s uploads/public area, and the chat answers a question using the uploaded document. The demo response frames the answer as grounded in the provided text, showing the RAG loop working end-to-end.

By the end, the main takeaway is pragmatic. Gemini CLI can generate a working RAG web app quickly, with MCP and tool-use capabilities already integrated. But the experience still needs polish: configuration alignment with Tailwind 4, fewer abrupt model switches, faster execution, and more reliable build/test steps would make the workflow more dependable. The demo also flags confusing context/token accounting (high “context left” despite heavy usage) and a long session time for the amount of work completed.