Gemini CLI

Google-backed terminal coding agent with sandboxing, approval modes, hierarchical GEMINI.md context, MCP, skills, and workspace trust controls.

Type: CLI

Best for: Google-backed terminal workflows, vendor diversification, sandboxed execution, and teams that want reusable context and skills

Official docs: https://geminicli.com/docs/

Last verified: 2026-08-09

Status: Documentation verified

Verification scope: Official Gemini CLI installation, CLI cheatsheet, GEMINI.md, configuration, sandboxing, trusted folders, skills, and MCP documentation were reviewed. Commands were not executed locally in this repository.

60-Second Cheat Sheet

Need Command or file Notes
Install with npm npm install -g @google/gemini-cli Stable release by default
Run without installing npx @google/gemini-cli Useful for quick evaluation
Install with Homebrew brew install gemini-cli macOS/Linux
Start interactive CLI gemini Main terminal entrypoint
Run one prompt gemini -p "summarize README.md" Non-interactive output
Prompt and stay interactive gemini -i "explain this project" Runs prompt, then continues session
Resume latest session gemini -r "latest" Continue previous work
Start sandboxed gemini -s Isolates shell/file operations
Start in plan mode gemini --approval-mode=plan Analyze before acting
Inspect context /memory show Shows concatenated active context
Reload context /memory reload Reloads GEMINI.md files
List skills /skills list Shows discovered Agent Skills
Shared repo context GEMINI.md Hierarchical project instructions
Workspace skills .gemini/skills/ or .agents/skills/ Shareable specialized workflows
User settings ~/.gemini/settings.json User configuration
Project settings .gemini/settings.json Repository-specific configuration

Safe default: use plan mode for broad changes and -s when command/file isolation matters. Workspace trust is a separate optional security control and is disabled by default unless you enable it in settings.

Five Prompts Worth Bookmarking

Understand a repository

Explain this repository's architecture, entry points, build/test commands, and highest-risk areas. Do not modify files.

Plan a change

Read the relevant files and active GEMINI.md context, then propose a step-by-step implementation plan with tests, risks, and validation. Do not edit until approved.

Review code

Review the current changes for correctness, regressions, security issues, and missing tests. List actionable findings by severity and cite the relevant files.

Debug safely

Reproduce the failing command or test, identify the first root-cause failure, explain it, and propose the smallest fix before changing files.

Create reusable project context

Review this repository and draft a concise GEMINI.md containing architecture, build/test commands, coding conventions, risky paths, and completion criteria. Include only facts confirmed from the repo.

What Gemini CLI Is Best At

  • terminal-first workflows backed by Google/Gemini accounts and models
  • teams that want a secondary major-vendor coding agent
  • sandboxed local execution
  • explicit approval modes including plan-first workflows
  • hierarchical project context with GEMINI.md
  • reusable Agent Skills with progressive activation
  • MCP integrations and extension-based customization

Installation

npm

npm install -g @google/gemini-cli

No-install evaluation

npx @google/gemini-cli

Homebrew

brew install gemini-cli

The current installation docs also list MacPorts, Anaconda, containerized execution, and source-based paths.

Gemini CLI publishes stable, preview, and nightly channels. Use stable for team defaults unless you have a deliberate testing reason to choose another channel.

First Session

cd /path/to/your/repository
gemini

A good first request is:

Explain this project, identify the build and test commands, and recommend one low-risk improvement. Do not modify files.

For higher-risk or unfamiliar work, start with:

gemini --approval-mode=plan

or use sandboxing:

gemini -s

Daily CLI Commands

gemini
gemini -p "summarize README.md"
gemini -i "explain this project"
gemini -r "latest"
gemini -s
gemini --approval-mode=plan
gemini update

Useful in-session commands include:

/memory show
/memory reload
/skills list
/settings
/trust

Use the current CLI cheatsheet and /help for the exact command surface on your installed version.

Approval Modes

Current Gemini CLI docs list these approval modes:

  • default — standard confirmation behavior
  • auto_edit — automatically approves edit operations while retaining broader controls
  • yolo — automatically approves actions; broad and high risk
  • plan — plans and analyzes before implementation

Use plan for unfamiliar or cross-cutting work. Avoid yolo as a normal team default.

Sandboxing

Start sandboxed execution with:

gemini -s

Sandboxing is designed to isolate shell commands and file modifications from the host system and can help limit file access to the project boundary.

Use sandboxing together with approval policy rather than treating one as a replacement for the other.

Repository Instructions with GEMINI.md

Gemini CLI uses hierarchical context files whose default name is:

GEMINI.md

Current official context discovery includes:

  1. global context at ~/.gemini/GEMINI.md
  2. project/workspace context from configured workspace directories and parents
  3. just-in-time context discovered near files the agent accesses

Inspect the exact active context with:

/memory show

After changing context files during a session, reload them with:

/memory reload

You can configure alternate context filenames in settings.json, including cross-agent names such as AGENTS.md.

Trusted Folders

Trusted Folders is an optional security feature that controls whether project-specific configuration is loaded with full capabilities.

Important correction for team rollout: folder trust is disabled by default. Enable it deliberately in user settings if you want trust prompts and trust-root behavior to be part of your security model.

Do not use --skip-trust as a routine shortcut if your team relies on folder trust as a boundary.

Agent Skills

Gemini CLI supports Agent Skills as on-demand specialized workflows.

Discovery tiers include:

  • built-in skills
  • extension skills
  • user skills at ~/.gemini/skills/ or ~/.agents/skills/
  • workspace skills at .gemini/skills/ or .agents/skills/

Useful commands:

/skills list
/skills reload

and from the terminal:

gemini skills list --all

Workspace skills can be committed with the repository, but review bundled scripts and assets before approving them.

MCP and External Tools

Gemini CLI supports MCP servers and can restrict the allowed server set through CLI/configuration options.

Treat each server as a separate approval for:

  • authentication
  • data exposure
  • tool permissions
  • write/destructive capability
  • network access

Team Adoption Checklist

  • [ ] standardize stable versus preview/nightly release channels
  • [ ] define whether GEMINI.md is the canonical project context file
  • [ ] decide whether folder trust will be enabled across the team
  • [ ] use plan mode for broad or unfamiliar work
  • [ ] use sandboxing for higher-risk local execution
  • [ ] approve skills and MCP servers separately from base CLI access
  • [ ] keep yolo mode out of normal developer defaults

Security Notes

  • approval mode and sandboxing are separate controls
  • trusted folders do nothing unless the feature is enabled
  • workspace GEMINI.md files shape behavior but do not enforce policy
  • Agent Skills can include scripts and bundled assets; review before sharing
  • MCP servers expand the tool and data boundary
  • keep Google/API credentials outside repository files

Alternatives

  • Claude Code for terminal-first work with mature project-memory, permission, and MCP patterns
  • OpenAI Codex for OpenAI-centered terminal, review, and automation workflows
  • GitHub Copilot for GitHub-native editor, CLI, review, and repository workflows
  • Cursor for a strongly editor-first environment with a terminal agent

See the Comparison Matrix for a side-by-side shortlist.

Verification Status

  • Status: Documentation verified
  • Last verified: 2026-08-09
  • Scope: installation, release channels, CLI commands, approval modes, sandboxing, GEMINI.md, folder trust, skills, and MCP reviewed against current official Gemini CLI docs
  • Not locally tested: npm/Homebrew installation, Google sign-in, sandbox execution, skills installation, and enterprise account behavior

Sources