claude-code-cli

name	claude-code-cli
description	Claude Code CLI usage for automated/headless execution in Daytona sandboxes. Use when running Claude Code commands, configuring authentication, or setting up non-interactive mode.

Claude Code CLI

Source: Claude Code CLI Reference, Headless Usage

Installation

# Native (recommended)
curl -fsSL https://claude.ai/install.sh | bash

# NPM (deprecated)
npm install -g @anthropic-ai/claude-code

Verify: claude --version

Authentication

Environment Variables

Variable	Description	Priority
ANTHROPIC_API_KEY	API key from console.anthropic.com	High (overrides OAuth)
CLAUDE_CODE_OAUTH_TOKEN	OAuth token for subscription auth	Normal

Warning: Setting ANTHROPIC_API_KEY bypasses subscription and uses pay-as-you-go API rates.

For Headless/Sandbox Use

Since Claude Code requires browser-based OAuth, use one of:

1. API Key (recommended for CI/sandboxes):

   export ANTHROPIC_API_KEY="sk-ant-..."
   claude -p "your prompt"
   

2. OAuth Token Transfer (from authenticated machine):

   # On local machine: authenticate normally, then copy token
   # Location: ~/.config/claude-code/auth.json
   # Transfer to sandbox via environment variable
   export CLAUDE_CODE_OAUTH_TOKEN="token-from-auth-json"
   

Check Current Auth

In interactive mode: /status

Non-Interactive Mode (-p / --print)

For automated execution, use -p flag:

claude -p "your prompt here"

This runs the prompt and exits (no REPL).

Essential Flags

Core Flags

Flag	Description	Example
-p, --print	Non-interactive mode, exit after response	claude -p "explain this"
-c, --continue	Continue most recent conversation	claude -c -p "now fix it"
-r, --resume <id>	Resume specific session	claude -r "abc123" "continue"
--model <name>	Set model (sonnet, opus, or full name)	claude --model opus -p "..."

Permission & Security

Flag	Description
--dangerously-skip-permissions	Skip ALL permission prompts (use with caution)
--allowedTools <tools>	Auto-approve specific tools without prompting
--disallowedTools <tools>	Block specific tools entirely
--permission-mode <mode>	Start in specific mode (e.g., plan)

Tool Control

# Allow specific tools (comma-separated)
claude -p "fix the bug" --allowedTools "Read,Edit,Bash"

# Allow with prefix matching (e.g., any git command)
claude -p "commit changes" --allowedTools "Bash(git:*)"

# Disable all tools
claude -p "explain this" --tools ""

# Use only specific tools
claude -p "review code" --tools "Read,Grep,Glob"

Output Formats

Flag	Value	Description
--output-format	text	Plain text (default)
	json	JSON with metadata
	stream-json	Newline-delimited JSON streaming

# Get JSON output
claude -p "summarize" --output-format json

# Parse with jq
claude -p "summarize" --output-format json | jq -r '.result'

# Structured output with schema
claude -p "list functions" --output-format json \
  --json-schema '{"type":"object","properties":{"functions":{"type":"array","items":{"type":"string"}}}}'

System Prompt

Flag	Description
--system-prompt <text>	Replace entire system prompt
--append-system-prompt <text>	Add to default prompt (recommended)
--system-prompt-file <path>	Load replacement from file
--append-system-prompt-file <path>	Append from file

# Append custom instructions (keeps defaults)
claude -p "review this" --append-system-prompt "Focus on security vulnerabilities"

# Replace entire prompt
claude -p "task" --system-prompt "You are a Python expert"

Budget & Limits

Flag	Description
--max-budget-usd <amount>	Maximum spend (print mode only)
--max-turns <n>	Limit agentic turns (print mode only)

claude -p "refactor this file" --max-budget-usd 5.00 --max-turns 10

Session Management

# Get session ID from output
session_id=$(claude -p "start review" --output-format json | jq -r '.session_id')

# Resume later
claude -p "continue" --resume "$session_id"

# Continue most recent
claude -p "what was I doing?" --continue

Recommended Sandbox Configuration

For running in Daytona sandboxes:

# Full automated execution with tool access
ANTHROPIC_API_KEY="sk-ant-..." claude -p "Create a hello world Python script" \
  --dangerously-skip-permissions \
  --output-format json \
  --max-turns 20

# Or with specific tool allowlist
ANTHROPIC_API_KEY="sk-ant-..." claude -p "Fix the bug in auth.py" \
  --allowedTools "Read,Edit,Bash,Grep,Glob" \
  --output-format json

# With OAuth token instead
CLAUDE_CODE_OAUTH_TOKEN="token..." claude -p "Explain this codebase" \
  --dangerously-skip-permissions

Exit Codes

Code	Meaning
0	Success
1	Error (check stderr)

Piped Input

# Pipe file content
cat src/main.ts | claude -p "explain this code"

# Pipe command output
git diff | claude -p "review these changes"

# Pipe multiple files
cat src/*.ts | claude -p "find security issues"

Common Patterns

Code Review

git diff HEAD~1 | claude -p "Review these changes for bugs and improvements" \
  --append-system-prompt "Focus on security and performance" \
  --output-format json

Bug Fix

claude -p "Fix the failing test in tests/auth.test.ts" \
  --allowedTools "Read,Edit,Bash(npm test:*)" \
  --max-turns 15

Code Generation

claude -p "Create a REST API endpoint for user registration" \
  --dangerously-skip-permissions \
  --max-turns 30

Explanation Only (No Tools)

cat complex-file.ts | claude -p "Explain how this works" --tools ""

Error Messages to Watch For

In sandbox output, watch for:

• "Authentication failed" - Invalid token/key
• "OAuth token has expired" - Need fresh token
• "Invalid API key" - Check ANTHROPIC_API_KEY
• "Rate limit exceeded" - Back off and retry
• "Context length exceeded" - Reduce input size

Notes

• Skills/slash commands (like /commit) only work in interactive mode, not with -p
• -p mode exits after completion; use --continue for multi-step tasks
• --dangerously-skip-permissions gives full tool access - use in trusted environments only
• Output buffering: Claude streams output, buffer at ~100ms for IPC efficiency