TL;DR: Gemini CLI is a genuinely Apache 2.0-licensed terminal coding agent from Google — 107K GitHub stars, weekly releases, and a free tier (1,000 requests/day) that no other frontier-model agent matches. The catch: it has no first-party local model support, so every prompt, file, and diff goes to Google’s servers. Free and open-licensed is not the same as private.

	Gemini CLI	OpenCode + Ollama	Aider + local model
Best for	Free frontier-model agent, cloud OK	Private agentic coding, zero per-token cost	Git-centric pair programming, local or cloud
Price / Cost	$0 up to 1,000 req/day, then paid API	$0 (your hardware, ~12-24GB VRAM)	$0 local, or BYOK cloud pricing
The catch	All code goes to Google; no native local backend	Local 7B-32B models lag Gemini 3 on hard refactors	No parallel tool loop; leaner feature set

Honest take: If you don’t mind Google reading your code, Gemini CLI is the best free coding agent available in September 2026. If you’re on this site because you do mind, wire up OpenCode with Ollama instead and accept the model quality gap.

What does Gemini CLI actually do?

Gemini CLI is a full agentic loop in your terminal, not a one-shot prompt runner. As of v0.61.0 (released September 23, 2026 — the project ships stable releases weekly), the agent can read and modify files in your project, execute shell commands, search and fetch from the web, and connect to MCP (Model Context Protocol) servers for custom tools. You give it a task in plain English; it plans, edits, runs, reads the errors, and iterates — the same working pattern as Claude Code or OpenCode.

The backend is Google’s Gemini 3 model family with a 1M-token context window. That context size matters more than it sounds for agent work: a 1M window means the agent can hold a mid-sized codebase, the conversation history, and tool outputs simultaneously without aggressive truncation. Local models running on a 24GB card typically operate at 32K-128K usable context — a real handicap on multi-file refactors, which is exactly where agents earn their keep.

Adoption is not in question. The repository sits at 107.1K stars and 14.6K forks as of September 2026, making it one of the fastest-growing developer tools Google has ever shipped. It installs through npm:

$ npm install -g @google/gemini-cli
$ gemini
# First run opens a browser window for Google account sign-in,
# then drops you into the interactive agent prompt:
#
# ███ GEMINI
# Tips for getting started:
# 1. Ask questions, edit files, or run commands.
# 2. Be specific for the best results.

Sign in with a personal Google account and you’re on the free tier immediately — no card, no API key.

Is Gemini CLI really open source?

Yes — the code is Apache 2.0 with no strings attached. The LICENSE file in google-gemini/gemini-cli is stock Apache License 2.0: no usage-volume clause, no network-use restriction, no “community license” carve-outs of the kind Meta attached to Llama. You can fork it, embed it in a commercial product, ship it in a Docker image for your team’s CI, or build a competing tool from its agent loop. Several projects have done exactly that (more on those below).

But be precise about what the license covers: the client is open source; the model is not. Gemini 3 weights are proprietary and API-only. This is the inverse of the usual self-hosting trade: with Qwen or DeepSeek you get open weights and write your own tooling; with Gemini CLI you get polished open tooling permanently tethered to a closed cloud model. Apache 2.0 on the CLI means you can audit exactly what gets sent to Google — and the answer is: your prompts, the files the agent reads, shell output, and telemetry (which is configurable). It does not mean any of that stays on your machine.

How much can you use Gemini CLI for free?

The free tier is 60 requests per minute and 1,000 requests per day with a personal Google account — verified from the project README, September 2026. A free Gemini API key gets the same 1,000 requests/day. That is dramatically more generous than anything comparable:

Access path	Limit	Cost
Google account sign-in (OAuth)	60 req/min, 1,000 req/day	$0
Free Gemini API key	1,000 req/day	$0
Paid Gemini API key	Pay per token	Usage-based
Claude Code (comparison)	Plan quota	From $20/mo
OpenCode + Ollama (comparison)	Unlimited	$0 + your hardware

One caution on reading that table: 1,000 requests/day is not 1,000 tasks. An agentic loop burns multiple requests per task — plan, edit, run, read output, retry. A heavy refactoring session can chew through 100+ requests. In practice, the free tier comfortably covers a solo developer’s daily driver usage, but a team hammering it through CI will hit the ceiling and need a paid key. Rate-limit specifics also change; check Google’s rate limit docs before building a workflow around today’s numbers.

There is no free lunch mystery here: the free tier is a data and mindshare play. Your usage helps Google tune its models and locks the workflow to its API. Whether that trade bothers you is the entire decision.

Can Gemini CLI use a local model like Ollama?

No — not natively, and this is the finding that matters most for this site’s readers. As of v0.61.0 there is no supported flag or config to point Gemini CLI at an Ollama, llama.cpp, or vLLM endpoint. The feature has been requested since 2025 (issue #5938, and discussion #24166 proposing universal OpenAI-compatible provider support) and remains open. A community PR that added native Ollama integration was not merged.

Three workarounds exist, in ascending order of commitment:

1. The undocumented base URL override. The underlying @google/genai SDK respects a GOOGLE_GEMINI_BASE_URL environment variable, so you can redirect the CLI’s API traffic to a proxy that speaks the Gemini API shape. It is undocumented in Gemini CLI itself and can break on any weekly release. Don’t build on it.

2. A LiteLLM proxy. Recent LiteLLM releases explicitly support fronting Gemini CLI. You run LiteLLM locally, register your Ollama models in its config, and point the CLI at the proxy:

# litellm-config.yaml
model_list:
  - model_name: gemini-2.5-pro   # what the CLI asks for
    litellm_params:
      model: ollama/qwen3-coder  # what actually serves it
      api_base: http://localhost:11434

This works, but you’re translating between two API dialects, and tool calling — the thing that makes an agent an agent — only survives the translation if your local model’s chat template supports it properly.

3. The forks. Because the license is Apache 2.0, the community route around Google is simply forking: ollama-code, easy-llm-cli, and open-gemini-cli all take Gemini CLI’s agent loop and rewire it for OpenAI-compatible providers, including local Ollama. If you want this codebase with local models, a maintained fork is the honest path — at that point, though, compare the forks against OpenCode, which was designed provider-agnostic from day one rather than patched into it.

The strategic read: Google has had over a year of loud demand for local backend support and has not shipped it. The free tier is the product; a local escape hatch would defeat its purpose. Assume the privacy ceiling is permanent.

How does Gemini CLI compare to OpenCode and Aider?

Gemini CLI wins on model quality per dollar; OpenCode wins on privacy and flexibility; Aider wins on git discipline and predictability. Numbers as of September 2026:

	Gemini CLI	OpenCode	Aider
License	Apache 2.0	MIT	Apache 2.0
Local model support	No (proxy/fork only)	Native, any OpenAI-compat endpoint	Native (Ollama, llama.cpp)
Default model	Gemini 3 (cloud)	Bring your own	Bring your own
Free usage	1,000 req/day	Unlimited w/ local	Unlimited w/ local
MCP support	Yes	Yes	Limited
Data leaves your machine	Always	Only if you choose a cloud model	Only if you choose a cloud model

The deeper pattern — covered in our state of open-source coding agents overview — is that the agent loop itself is now commodity. File editing, shell execution, MCP: every serious tool has them. What differentiates is (a) which models you’re allowed to use and (b) where your code goes. Gemini CLI is the strongest entry in the “free but cloud-locked” quadrant; it has no entry at all in the “private” quadrant.

If you’re weighing local hardware for the OpenCode/Aider path instead: a used RTX 3090’s 24GB runs Qwen3-class 32B coder models at Q4 — see runaihome.com’s local AI hardware guides for current pricing (GPU street prices moved a lot in 2026; don’t trust MSRPs). For how Gemini CLI stacks against Cursor and Claude Code in a pure productivity frame, aicoderscope.com covers AI coding tools from the cloud-first angle.

When NOT to use Gemini CLI

• Any code you can’t send to Google. Client work under NDA, proprietary employer code without explicit approval, healthcare/finance code under compliance regimes. There is no configuration that keeps prompts local — the workarounds above route around the product, not through it.
• Air-gapped or offline environments. No cloud API, no agent. OpenCode or Aider with a local model is the only option.
• Workflows that need provider portability. If you might switch models — cloud or local — building muscle memory and scripts around a single-provider CLI is technical debt. The 2025-26 model leaderboard has flipped repeatedly.
• Sustained heavy usage on the free tier. CI pipelines and team usage will hit 1,000 req/day fast, and then you’re on metered API pricing you should compare against alternatives before, not after, adopting.

Verdict

Gemini CLI is the best free coding agent of 2026 and a legitimately open-licensed one — that combination is rare and worth acknowledging plainly. For a student, a hobbyist on cloud-tolerant side projects, or anyone who wants to feel what frontier-model agentic coding is like without paying, it’s the obvious first install. For the self-hosting audience, it’s a well-built tool pointed the wrong way: Apache 2.0 code wrapped around a mandatory data pipe to Google. Use it where privacy doesn’t matter; keep OpenCode + Ollama for everything else, and revisit if Google ever merges local backend support. Don’t hold your breath.

FAQ

Is Gemini CLI free for commercial use? The software is Apache 2.0, so yes, without restriction. The free API tier’s terms are a separate question — Google’s terms of service govern API usage, and free-tier prompts may be used to improve Google’s services. For commercial code, a paid API key with Google’s data-handling commitments is the defensible setup.

Can Gemini CLI work fully offline? No. The CLI is open source but requires Google’s cloud API for inference. Offline agentic coding requires a locally served model — see our local coding agent comparison for tools that support it natively.

What are Gemini CLI’s free tier limits in 2026? 60 requests per minute and 1,000 requests per day with a Google account sign-in, or 1,000 requests per day with a free Gemini API key (verified September 2026). An agent loop uses several requests per task, so budget accordingly.

Sources

• google-gemini/gemini-cli — GitHub repository and README (license, features, free tier figures; accessed 2026-09-24)
• Gemini CLI releases — v0.61.0 (2026-09-23)
• Discussion #24166: Universal Local Model Support via Ollama/OpenAI-Compatible Providers
• Issue #5938: Add support for local/offline language models
• Gemini API rate limits — official docs

What self-hosting actually costs

Real cost breakdowns for self-hosted AI: hardware floors, power, maintenance hours, and the honest comparison against paying for it. No spam, unsubscribe anytime.

Email address Subscribe