Gemini CLI

Original Article ↗ Hacker News Discussion ↗

Positioning and feature set

  • Seen as Google’s answer to Claude Code / Codex: an agentic CLI that can read/write files, run shell commands, use web search, and work over whole repos.
  • Uses Gemini 2.5 Pro with a 1M-token context window; can downgrade to 2.5 Flash when overloaded.
  • Open-source (Apache 2.0) and built around tools, MCP servers, and GEMINI.md-style project instructions; integrates with VS Code (Code Assist).

Pricing, limits, and product sprawl

  • Free tier via personal Google login: 60 requests/min and 1,000/day; many think this is extremely generous, others say it “doesn’t last long” in practice.
  • Higher, paid limits require Gemini API or Vertex / Gemini Code Assist Standard/Enterprise, but users find the matrix of plans (Gemini Pro, AI Pro, Code Assist, Workspace, Vertex, AI Studio) deeply confusing.
  • Workspace users are frustrated that existing Gemini entitlements don’t straightforwardly apply to the CLI and often still require GCP projects and extra billing.

Authentication and UX friction

  • Frequent pain around auth flows: headless/remote machines, Workspace accounts needing GOOGLE_CLOUD_PROJECT, and opaque error messages.
  • Some praise the interface, /help, and permission prompts; others dislike jokey spinners and vague “thinking” messages.
  • Many request a simple, consumer-style “Claude Max”-like subscription that covers CLI + app + API.

Practical performance and quality

  • Mixed results: some report “beast”‑level performance on large codebases, great refactors, and better-than-Claude debugging; others see flaky edits, broken imports, duplicated code, and misapplied patches.
  • Several users hit early rate limits, long stalls, and auto-downgrades to Flash, with CLI sessions burning millions of tokens and hours of API time.
  • Compared to Claude Code/Aider, Gemini is often described as harder to steer, more verbose, more “agentic” but prone to loops and overcomplicated changes.

Privacy and data use

  • Heavy concern about source code being sent to Google and used for training, especially on free/personal tiers.
  • Docs and terms are perceived as confusing and inconsistent; Google staff added a clarifying doc, but users still debate what is collected, retained, and trainable under each auth mode and whether any true opt-out exists.

Ecosystem, implementation, and reception

  • Implemented in Node/TypeScript with Ink; this sparks long debate about wanting single static binaries (Go/Rust) vs npm-based tooling.
  • Open source nature and MCP support are praised, including potential to swap in other or local models, but many still prefer model‑agnostic tools like Aider, OpenHands, or opencode.
  • Overall sentiment: strong interest and respect for Gemini 2.5 Pro, but disappointment with Google’s product fragmentation, pricing story, and early reliability of the CLI.