claude-cli（Claude CLI）

Claude CLI - Command Line Interface

You are a specialist in using the claude CLI (Claude Code). This skill provides comprehensive workflows, best practices, and common patterns for interactive sessions, scripting, MCP management, and automation.

What is Claude CLI?

The claude CLI is the command-line interface for Claude Code that provides:

• Interactive sessions - Default conversational mode for coding and problem-solving
• Print mode - Non-interactive output for pipes, scripts, and automation
• Session management - Continue, resume, and fork conversations
• MCP integration - Configure and manage Model Context Protocol servers
• Plugin system - Extend functionality with marketplace plugins
• Tool control - Fine-grained control over available tools
• Flexible output - Text, JSON, or streaming formats

Core Modes

Interactive Mode (Default)

# Start interactive session
claude

# Start with a prompt
claude "explain this codebase"

# Continue last conversation
claude --continue
claude -c

# Resume specific session
claude --resume
claude -r SESSION_ID

# Fork a session (new ID, same history)
claude --resume SESSION_ID --fork-session

Print Mode (Non-Interactive)

# Single response, then exit
claude --print "what is rust?"
claude -p "explain async/await"

# Pipe input
echo "explain this code" | claude -p

# Pipe output
claude -p "list all rust files" | grep "src/"

# JSON output for parsing
claude -p --output-format json "summarize this"

# Streaming JSON (realtime)
claude -p --output-format stream-json "analyze logs"

Session Management

Continue Last Conversation

# Continue most recent session
claude --continue
claude -c

# Continue and fork (new session ID)
claude -c --fork-session

Resume Specific Session

# Interactive selection
claude --resume
claude -r

# Specific session ID
claude -r abc123-def456-789

# Resume and fork
claude -r abc123 --fork-session

Custom Session ID

# Use specific UUID
claude --session-id 12345678-1234-1234-1234-123456789abc

Model Selection

Model Aliases

# Use Sonnet (default)
claude --model sonnet

# Use Opus
claude --model opus

# Use Haiku
claude --model haiku

Full Model Names

# Specific version
claude --model claude-sonnet-4-5-20250929
claude --model claude-opus-4-20250514
claude --model claude-haiku-4-20250430

Fallback Model (Print Mode Only)

# Fallback to Haiku if Sonnet overloaded
claude -p --fallback-model haiku "analyze this"

Tool Control

Specify Available Tools

# Only specific tools
claude -p --tools "Bash,Read,Edit" "refactor code"

# Default tools
claude -p --tools default "write function"

# No tools (chat only)
claude -p --tools "" "explain concept"

Allow/Disallow Tools

# Allow specific tool patterns
claude --allowed-tools "Bash(git:*)" "commit changes"
claude --allowed-tools "Bash(npm:*),Read,Edit" "update package"

# Disallow tools
claude --disallowed-tools "Bash(rm:*)" "clean directory"
claude --disallowed-tools "Write" "analyze code"

# Multiple patterns
claude --allowed-tools "Bash(git:*)" "Bash(cargo:*)" "rust workflow"

System Prompts

Set System Prompt

# Replace default system prompt
claude --system-prompt "You are a Rust expert. Focus on idiomatic code."

# Append to default
claude --append-system-prompt "Always explain your reasoning."

Permission Modes

# Auto-accept edits (no confirmation)
claude --permission-mode acceptEdits

# Bypass all permissions
claude --permission-mode bypassPermissions

# Plan mode (planning workflow)
claude --permission-mode plan

# Default (normal confirmations)
claude --permission-mode default

Dangerous Permissions

# Skip ALL permission checks (use in sandboxes only)
claude --dangerously-skip-permissions

# Allow skipping as an option (not default)
claude --allow-dangerously-skip-permissions

MCP Server Management

List MCP Servers

# List configured servers
claude mcp list

Add MCP Server

# Add HTTP server
claude mcp add --transport http sentry https://mcp.sentry.dev/mcp

# Add SSE server
claude mcp add --transport sse asana https://mcp.asana.com/sse

# Add stdio server
claude mcp add --transport stdio myserver npx -y my-mcp-server

# Add with environment variables
claude mcp add --transport stdio airtable \
  --env AIRTABLE_API_KEY=YOUR_KEY \
  -- npx -y airtable-mcp-server

Add from JSON

# Add stdio server from JSON
claude mcp add-json myserver '{"command":"node","args":["server.js"]}'

# Add SSE server from JSON
claude mcp add-json sse-server '{"url":"https://example.com/sse"}'

Import from Claude Desktop

# Import all MCP servers from Claude Desktop
claude mcp add-from-claude-desktop

# Import specific server
claude mcp add-from-claude-desktop --name myserver

Remove MCP Server

# Remove server
claude mcp remove myserver

Get MCP Server Details

# Show configuration
claude mcp get myserver

Reset Project MCP Choices

# Reset approved/rejected project-scoped servers
claude mcp reset-project-choices

MCP Configuration

# Load MCP from config file
claude --mcp-config config.json

# Load from JSON string
claude --mcp-config '{"myserver":{"command":"node","args":["server.js"]}}'

# Multiple configs
claude --mcp-config config1.json config2.json

# Strict mode (only use specified config)
claude --strict-mcp-config --mcp-config config.json

Project-Level MCP Configuration (.mcp.json)

IMPORTANT: Project-specific MCP servers belong in a .mcp.json file at the project root.

.mcp.json File Structure

Create .mcp.json in your project root with this structure:

{
  "mcpServers": {
    "server-name": {
      "command": "npx",
      "args": ["-y", "package-name"],
      "env": {
        "API_KEY": "value"
      }
    }
  }
}

Common Patterns

Node.js MCP server:

{
  "mcpServers": {
    "myserver": {
      "command": "node",
      "args": ["./dist/index.js"],
      "env": {
        "API_KEY": ""
      }
    }
  }
}

Python/uvx MCP server:

{
  "mcpServers": {
    "knowledge_base": {
      "command": "uvx",
      "args": [
        "chroma-mcp",
        "--client-type",
        "persistent",
        "--data-dir",
        "/path/to/data"
      ],
      "env": {
        "ANONYMIZED_TELEMETRY": "false"
      }
    }
  }
}

Language server integration:

{
  "mcpServers": {
    "gopls": {
      "command": "/path/to/mcp-language-server",
      "args": ["--workspace", ".", "--lsp", "gopls", "--", "-mode=stdio"],
      "transport": "stdio"
    }
  }
}

Docker/Podman container:

{
  "mcpServers": {
    "github": {
      "command": "podman",
      "args": [
        "run",
        "-i",
        "--rm",
        "-e",
        "GITHUB_PERSONAL_ACCESS_TOKEN",
        "ghcr.io/github/github-mcp-server"
      ],
      "env": {
        "GITHUB_PERSONAL_ACCESS_TOKEN": "your-token"
      }
    }
  }
}

Multiple servers in one project:

{
  "mcpServers": {
    "git": {
      "command": "uvx",
      "args": ["mcp-server-git"]
    },
    "sequential-thinking": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-sequential-thinking"]
    },
    "gdrive": {
      "command": "gdrive",
      "args": ["mcp", "stdio"]
    }
  }
}

Usage

When Claude CLI runs in a project directory:

# Automatically detects and loads .mcp.json
claude

# View which servers are loaded
claude mcp list

# Reset project server approvals
claude mcp reset-project-choices

Key Points:

• Place .mcp.json at project root
• Claude prompts to approve/reject project servers on first use
• Use for project-specific tools (language servers, project APIs, local databases)
• Separate from global user MCP configuration
• Safe to commit to version control (but NEVER commit secrets in env)

.gitignore Pattern for Secrets

If your .mcp.json contains secrets, create a template:

# .mcp.json.example (commit this)
{
  "mcpServers": {
    "myserver": {
      "command": "node",
      "args": ["./dist/index.js"],
      "env": {
        "API_KEY": "your-api-key-here"
      }
    }
  }
}

# .gitignore
.mcp.json

Then developers copy .mcp.json.example to .mcp.json and add their credentials.

Plugin Management

List Available Plugins

# Browse marketplace
claude plugin marketplace

Install Plugin

# Install from default marketplace
claude plugin install my-plugin

# Install from specific marketplace
claude plugin install my-plugin@my-marketplace

Uninstall Plugin

claude plugin uninstall my-plugin
claude plugin remove my-plugin

Enable/Disable Plugin

# Disable plugin
claude plugin disable my-plugin

# Enable plugin
claude plugin enable my-plugin

Validate Plugin

# Validate plugin manifest
claude plugin validate /path/to/plugin/

Load Plugin Directory

# Load plugins from custom directory
claude --plugin-dir /path/to/plugins/

# Multiple directories
claude --plugin-dir dir1/ dir2/

Settings and Configuration

Settings File

# Load from file
claude --settings /path/to/settings.json

# Load from JSON string
claude --settings '{"verbose":true,"model":"opus"}'

Setting Sources

# Only user settings
claude --setting-sources user

# User and project settings
claude --setting-sources user,project

# All sources (user, project, local)
claude --setting-sources user,project,local

Additional Directories

# Allow access to extra directories
claude --add-dir /path/to/data/

# Multiple directories
claude --add-dir /data/ /cache/ /logs/

Output Formats (Print Mode)

Text Output (Default)

# Plain text response
claude -p "summarize file.txt"

JSON Output

# Single JSON result
claude -p --output-format json "analyze code"

# Parse with jq
claude -p --output-format json "list issues" | jq '.response'

Streaming JSON

# Real-time streaming
claude -p --output-format stream-json "long analysis"

# Include partial messages
claude -p --output-format stream-json --include-partial-messages "task"

Input Formats

# Text input (default)
echo "prompt" | claude -p

# Streaming JSON input
cat input.jsonl | claude -p --input-format stream-json

# Replay user messages (for acknowledgment)
cat input.jsonl | claude -p \
  --input-format stream-json \
  --output-format stream-json \
  --replay-user-messages

Debug and Verbose Modes

Debug Mode

# Enable all debug output
claude --debug

# Filter debug categories
claude --debug "api,hooks"

# Exclude categories
claude --debug "!statsig,!file"

# Debug with print mode
claude -p --debug "test prompt"

Verbose Mode

# Override verbose setting
claude --verbose

IDE Integration

# Auto-connect to IDE on startup
claude --ide

Maintenance Commands

Check Health

# Run doctor to check auto-updater
claude doctor

Update

# Check for updates and install
claude update

Install Specific Version

# Install stable version
claude install stable

# Install latest version
claude install latest

# Install specific version
claude install 1.2.3

Setup Authentication

# Setup long-lived auth token
claude setup-token

Migrate Installer

# Migrate from global npm to local install
claude migrate-installer

Common Workflows

Workflow 1: Quick Analysis (Print Mode)

# Analyze file and exit
claude -p "summarize errors in app.log"

# Pipe to other tools
claude -p "extract function names from code.rs" | sort | uniq

Workflow 2: Scripted Automation

# JSON output for parsing
result=$(claude -p --output-format json "count TODO comments")
count=$(echo "$result" | jq -r '.response')
echo "Found $count TODOs"

Workflow 3: Restricted Tool Access

# Only allow git operations
claude --allowed-tools "Bash(git:*)" -p "show recent commits"

# Disallow destructive operations
claude --disallowed-tools "Bash(rm:*)" "Bash(mv:*)" "clean project"

Workflow 4: Custom System Prompt

# Specialized code reviewer
claude --system-prompt "You are a security-focused code reviewer. \
  Look for vulnerabilities, SQL injection, XSS, and auth issues." \
  "review authentication.py"

Workflow 5: Session Forking

# Resume session and fork for experimentation
claude -r abc123 --fork-session

# Try different approach without affecting original
claude -c --fork-session

Workflow 6: MCP Server Setup

# Add custom MCP server
claude mcp add --transport stdio mytools npx -y my-tools-server

# Test in session
claude "use mytools to analyze data"

# Remove if not needed
claude mcp remove mytools

Workflow 7: Project-Level MCP Configuration

# Create .mcp.json in project root
cat > .mcp.json <<'EOF'
{
  "mcpServers": {
    "git": {
      "command": "uvx",
      "args": ["mcp-server-git"]
    },
    "sequential-thinking": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-sequential-thinking"]
    }
  }
}
EOF

# Start claude - automatically detects .mcp.json
claude

# Approve project servers when prompted
# Servers are now available for this project only

# Reset approvals if needed
claude mcp reset-project-choices

Workflow 8: Plugin Workflow

# Install plugin
claude plugin install code-analyzer

# Use in session
claude "analyze code quality with code-analyzer"

# Disable temporarily
claude plugin disable code-analyzer

Workflow 9: Batch Processing with Stream JSON

# Process multiple prompts
cat prompts.jsonl | claude -p \
  --input-format stream-json \
  --output-format stream-json > results.jsonl

Workflow 10: Model Fallback for Reliability

# Try Sonnet, fall back to Haiku if overloaded
claude -p --model sonnet --fallback-model haiku "analyze data"

Workflow 11: Sandboxed Execution

# Safe sandbox with no permissions needed
claude -p --dangerously-skip-permissions --tools "Read" "analyze code"

Best Practices

1. Use Print Mode for Scripting

# Good - print mode for automation
claude -p "count errors" | wc -l

# Avoid - interactive for scripts
claude "count errors"  # Waits for user input

2. Specify Tools for Security

# Restrict tools in production scripts
claude -p --tools "Read,Grep" "find pattern"

# Disallow dangerous operations
claude --disallowed-tools "Bash(rm:*)" "cleanup task"

3. Use JSON for Parsing

# Easy to parse
claude -p --output-format json "analyze" | jq '.response'

# Harder to parse
claude -p "analyze"  # Mixed output

4. Fork Sessions for Experimentation

# Keep original session intact
claude -r session-id --fork-session

# Experiment without fear

5. Use Fallback Models for Reliability

# Production scripts should handle overload
claude -p --fallback-model haiku "critical task"

6. Debug with Filtering

# See relevant debug info only
claude --debug "api,mcp" "test mcp server"

# Exclude noisy categories
claude --debug "!statsig,!analytics" "run workflow"

7. Organize MCP Servers

# List before adding
claude mcp list

# Use clear names
claude mcp add --transport stdio db-tools npx -y db-mcp-server

# Remove unused servers
claude mcp remove old-server

8. Use System Prompts for Consistency

# Create specialized assistants
claude --append-system-prompt "Always explain your reasoning step by step."

Quick Reference

# Interactive mode
claude                                  # Start session
claude -c                               # Continue last session
claude -r                               # Resume session (interactive)
claude -r SESSION_ID                    # Resume specific session

# Print mode
claude -p "prompt"                      # Single response
claude -p --output-format json "prompt" # JSON output
claude -p --tools "Bash,Read" "prompt"  # Specific tools

# Model selection
claude --model opus                     # Use Opus
claude --fallback-model haiku           # Fallback model

# Tool control
claude --allowed-tools "Bash(git:*)"    # Allow only git
claude --disallowed-tools "Bash(rm:*)"  # Disallow rm

# MCP management
claude mcp list                         # List servers
claude mcp add --transport stdio name cmd # Add server
claude mcp remove name                  # Remove server

# Plugin management
claude plugin install name              # Install plugin
claude plugin uninstall name            # Uninstall plugin
claude plugin enable name               # Enable plugin
claude plugin disable name              # Disable plugin

# System prompts
claude --system-prompt "prompt"         # Set system prompt
claude --append-system-prompt "text"    # Append to system

# Debug
claude --debug                          # Enable debug
claude --debug "category,!exclude"      # Filter debug

# Maintenance
claude doctor                           # Check health
claude update                           # Update CLI
claude setup-token                      # Setup auth token

Common Patterns

Pattern 1: Safe Code Analysis

claude -p --tools "Read,Grep" "analyze security vulnerabilities in src/"

Pattern 2: Automated Git Workflow

claude -p --allowed-tools "Bash(git:*)" "review and summarize recent commits"

Pattern 3: JSON Parsing Pipeline

claude -p --output-format json "list all functions" | \
  jq -r '.response' | \
  grep "pub fn"

Pattern 4: Streaming Analysis

cat large_log.txt | \
  claude -p --output-format stream-json "find error patterns" | \
  jq -r 'select(.type == "response") | .content'

Pattern 5: Multi-Model Fallback

claude -p --model sonnet --fallback-model haiku "time-sensitive analysis"

Tips and Tricks

1. Session Management: Use --fork-session to experiment without affecting original conversation
2. Tool Safety: Always use --allowed-tools or --disallowed-tools for untrusted prompts
3. Automation: Prefer --output-format json for reliable parsing in scripts
4. Performance: Use --fallback-model haiku for faster responses when quality isn't critical
5. Debugging: Use --debug "!noisy,!category" to exclude verbose logs
6. MCP Servers: Test new servers with claude --strict-mcp-config to isolate issues
7. Plugins: Disable unused plugins to reduce startup time
8. Print Mode: Always use -p in scripts to avoid interactive prompts

Integration Examples

With Git

# Review changes before commit
claude -p --allowed-tools "Bash(git:*),Read" "review staged changes and suggest commit message"

With CI/CD

# Automated code review
claude -p --tools "Read,Grep" --output-format json "analyze code quality" > report.json

With Logs

# Real-time log analysis
tail -f app.log | claude -p --output-format stream-json "detect anomalies"

With Makefiles/Justfiles

# Analyze code quality
analyze:
    claude -p --tools "Read,Grep" "run code quality analysis"

# Generate documentation
docs:
    claude -p --output-format json "generate API docs from code" | jq -r '.response' > API.md

Summary

Primary use cases:

• Interactive coding sessions (default mode)
• Scripted automation (print mode)
• MCP server integration (global and project-level)
• Plugin management
• Tool-controlled execution

Key advantages:

• Flexible output formats (text, JSON, streaming)
• Fine-grained tool control
• Session persistence and forking
• MCP protocol support (use .mcp.json for project-specific servers)
• Extensible plugin system

Most common commands:

• claude - Start interactive session
• claude -p "prompt" - Print mode for scripting
• claude -c - Continue last conversation
• claude mcp list - Manage MCP servers
• claude --allowed-tools "pattern" "prompt" - Restrict tool access

Project-level MCP:

• Create .mcp.json in project root for project-specific MCP servers
• Automatically loaded when Claude runs in that directory
• Separate from global user configuration