Gemini CLI

Gemini CLI for interactive and non-interactive AI tasks. Use -p for non-interactive mode.

Preferred Defaults

Unless otherwise specified, use these settings:

gemini -m gemini-3-pro-preview -p "your prompt"

• -m gemini-3-pro-preview - Preferred model
• -p - Non-interactive prompt mode

Basic Usage

# Non-interactive with prompt
gemini -p "your prompt here"

# Read prompt from stdin
echo "your prompt" | gemini -p ""

# Interactive mode
gemini

# Interactive with initial prompt
gemini -i "start with this prompt"

# Positional prompt (one-shot)
gemini "your query here"

Key Options

Model Selection

-m, --model <MODEL>        # Specify model (e.g., "gemini-3-pro-preview")

Approval Modes

-y, --yolo                 # Auto-approve all tool calls
--approval-mode <MODE>     # Set approval mode

Mode	Description
default	Prompt for approval
auto_edit	Auto-approve edit tools
yolo	Auto-approve all tools
plan	Read-only mode

Sandbox

-s, --sandbox              # Run in sandbox mode

Session Management

-r, --resume <ID>          # Resume session ("latest" or index number)
--list-sessions            # List available sessions
--delete-session <ID>      # Delete session by index

Output Format

-o, --output-format <FORMAT>

Format	Description
text	Plain text output (default)
json	JSON with session_id, response, and stats
stream-json	Streaming JSONL events

Workspace

--include-directories <DIR>  # Additional directories to include (comma-separated or multiple flags)

Extensions and Tools

-e, --extensions <EXT>       # Specific extensions to use (default: all)
-l, --list-extensions        # List available extensions
--allowed-tools <TOOLS>      # Tools allowed without confirmation
--allowed-mcp-server-names <NAMES>  # Allowed MCP server names

Other Options

-d, --debug                  # Debug mode (F12 opens debug console)
--screen-reader              # Enable screen reader accessibility

Output Formats

JSON Output (-o json)

Returns object with:

• session_id - Session UUID
• response - Final response text
• stats - Token usage and tool call stats per model

Stream JSON Output (-o stream-json)

Streaming JSONL events:

Event Type	Description
init	Session started, includes session_id and model
message	User or assistant message with role and content
result	Final result with status and stats

Common Patterns

Default Invocation

gemini -m gemini-3-pro-preview -p "your task"

With YOLO Mode

gemini -m gemini-3-pro-preview -y -p "your task"

With Output Capture

gemini -m gemini-3-pro-preview -p "your task" -o json > result.json

Resume Latest Session

gemini -r latest

Include Additional Directories

gemini -m gemini-3-pro-preview --include-directories /path/to/other -p "your task"

Subcommands

gemini mcp                  # Manage MCP servers
gemini extensions <cmd>     # Manage extensions
gemini skills <cmd>         # Manage agent skills
gemini hooks <cmd>          # Manage hooks

Notes

• Use -p for non-interactive/scripted usage
• Use -i to start interactive with an initial prompt
• Positional arguments are treated as one-shot queries
• Session resume uses index numbers from --list-sessions or "latest"