Gemini CLI isn’t new — Google launched it in June 2025. But v0.36.0, released on April 1, 2026, marks a maturity milestone where the tool deserves serious attention if you haven’t put it on your radar yet.
Nearly ten months of weekly open source development. A free tier that remains the most generous on the market. And now subagents with native sandboxing for macOS and Windows. If you dismissed it when it came out for being “too raw,” it’s time to take another look.
What it is and how it works
Gemini CLI is an open source terminal agent (Apache 2.0) that runs a ReAct loop (Reason and Act) — think, plan, execute, evaluate, and repeat until the task is complete. The default model today is Gemini 3; users with a Gemini Code Assist license have access to Gemini 3.1 Pro (in Preview).
It comes with built-in tools from day one:
- Shell execution — runs commands within the agent’s loop
- File operations — reads, writes, and modifies files in your project
- Google Search grounding — connects answers with current web information
- Web fetch — reads external URLs for analysis or additional context
- MCP — full support for Model Context Protocol
And a technical advantage that’s still hard to ignore: 1 million context tokens. You can pass it large repositories without worrying about limits.
The free tier: the real differentiator
With a personal Google account, no credit card required:
60 requests per minute — 1,000 requests per day
For context: Google’s team says that 1,000 requests per day is double what an average developer uses in a normal workday. In practice, the free tier is enough for most individual projects.
# Installation
npm install -g @google/gemini-cli
# Without installing
npx @google/gemini-cli
# Start
gemini
# → Asks you to login with your Google account → activates your free license
What it can do today: real-world use cases
Understanding a new codebase
gemini -p "Analyze this project and give me a summary of the architecture, the main data flows, and any technical debt you can identify"
With 1M context tokens, you can feed a large repo entirely and have the agent truly understand it.
Feature generation
> Add input validation to the registration form. Required fields: name, email (format validation), password (min 8 chars). Show inline error messages without reloading.
The agent identifies relevant files, makes multi-file changes, and shows you the diff before applying it.
Plan Mode — built in since v0.29.0
One of the most useful additions in recent months: /plan generates a detailed execution plan before touching a single file.
> /plan Refactor the authentication module to support OAuth2 in addition to email/password
The agent presents the complete plan — files to modify, steps in order, dependencies — and waits for your approval before starting. Same pattern Claude Code uses with its planning mode.
CI/CD in non-interactive mode
# Automatically generate release notes
gemini -p "Generate release notes from the git log of the last 7 days" --output-format json
# Analysis in pipeline
gemini -p "Check this diff for security vulnerabilities" --output-format stream-json
The --output-format json flag is key for programmatic integrations — it returns the response in structured format, not free text.
GEMINI.md: you already know the pattern
If you use Claude Code, you already know how this works. You create a GEMINI.md file at your project root with the context the agent needs to know:
# GEMINI.md
## Stack
- Node.js 22, Express 5, PostgreSQL 16, Prisma ORM
- Tests: Vitest
## Conventions
- Strict TypeScript, no `any`
- Only async/await, never callbacks
- Don't modify /legacy without notifying first
You can combine a global ~/.gemini/GEMINI.md (personal preferences) with a local one per project. Identical to how CLAUDE.md works.
MCP: the servers you already have configured, work
// ~/.gemini/settings.json
{
"mcpServers": {
"github": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-github"],
"env": { "GITHUB_TOKEN": "your-token" }
},
"postgres": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-postgres",
"postgresql://localhost/my_db"]
}
}
}
Any MCP server you already have configured for Claude Code works here with the same configuration.
Direct comparison: Gemini CLI vs Claude Code
| Aspect | Gemini CLI | Claude Code |
|---|---|---|
| Base price | Free (1,000 req/day) | Requires paid API key |
| Default model | Gemini 3 | Claude Sonnet / Opus |
| Context | 1M tokens | 200K tokens |
| Open source |  Apache 2.0 |
 Closed |
| Google Search built-in | Native |
Via external MCP |
| Plan Mode | /plan |
Built-in |
| MCP | |
|
| Config file | GEMINI.md | CLAUDE.md |
| Non-interactive CI/CD | |
|
| Plugin ecosystem | Growing | More mature |
The most important point in that table: Claude Code requires an Anthropic API key with per-token costs. Gemini CLI with a personal Google account is free for individual use — no surprises on the bill.
When does it make sense to try it?
Use Gemini CLI if:
- You want an AI terminal agent without compromising your budget
- You work with large repositories (the 1M token context is hard to ignore)
- Your workflow benefits from natively integrated Google Search
- You want to see the agent’s source code and understand how it works
- You’re evaluating options before committing to something
Stick with Claude Code if:
- You have an established workflow with configured plugins and hooks
- You prefer the more mature ecosystem with more community resources
- You work with Anthropic’s models in other parts of your stack
The honest truth: it’s not a binary decision. You can have both configured and choose based on the task.
Conclusion
Gemini CLI launched in June 2025 and has ten months of weekly open source development under its belt. With v0.36.0 in hand, Gemini 3 as the default model, Plan Mode, subagents with sandboxing, and the most generous free tier on the market — there’s no excuse for not having tried it.
Do you have it installed? Are you using it alongside Claude Code or did you migrate? Let us know in the comments 