Claude Code CLI

Claude Code Agent Development

Build autonomous agents using Claude Code CLI. Language-agnostic patterns for agent orchestration and programmatic integration.

Quick Reference

Authentication

API Key

export ANTHROPIC_API_KEY="sk-ant-..."

OAuth Token (Pro/Team Subscription)

claude setup-token  # Interactive setup
export CLAUDE_CODE_OAUTH_TOKEN="your-token"

Running Agents

Headless Mode (-p)

# Basic prompt
claude -p "Implement a REST API"

# Piped input
echo "Fix the bug" | claude -p

# Full autonomy (sandboxed environments only)
claude -p --dangerously-skip-permissions "Build the feature"

Output Formats

For detailed parsing examples: See references/response-parsing.md

Structured Output

echo "What is 2+2?" | claude -p --output-format json \
  --json-schema '{"type":"object","properties":{"answer":{"type":"number"}},"required":["answer"]}'

Result: {"type": "result", "structured_output": {"answer": 4}, ...}

Tool Restrictions

# Read-only agent
claude -p --tools "Read,Glob,Grep" "Analyze code"

# Specific bash commands only
claude -p --allowed-tools "Bash(npm:*) Bash(git:*)" "Build project"

# Deny dangerous operations
claude -p --disallowed-tools "Bash(rm:*)" "Clean up"

For project configuration: See references/configuration.md

Multi-Agent Patterns

Sequential Pipeline

analysis=$(claude -p --output-format json "Analyze codebase")
claude -p "Implement based on: $analysis"

Parallel Agents

claude -p "Review auth/" > auth.txt &
claude -p "Review api/" > api.txt &
wait
claude -p "Summarize: $(cat *.txt)"

Specialist Agents

claude --agents '{"security": {"prompt": "You are a security expert"}}' \
  --agent security -p "Review auth"

For complete orchestration patterns: See references/orchestration.md

Session Management

# Persistent session
claude --session-id "task-123" -p "Start feature"
claude --session-id "task-123" -c -p "Continue"

# Stateless
claude -p --no-session-persistence "One-off task"

Project Configuration

.claude/settings.json

{
  "permissions": {
    "allow": ["Bash(npm:*)", "Edit", "Read"],
    "deny": ["Bash(rm -rf:*)"]
  }
}

CLAUDE.md

Project instructions auto-loaded by Claude.

For CI/CD and MCP: See references/configuration.md

Parsing JSON Output

Quick jq Examples

# Get result
claude -p --output-format json "prompt" | jq -r '.[-1].result'

# Check success
claude -p --output-format json "prompt" | jq '.[-1].is_error'

# Get cost
claude -p --output-format json "prompt" | jq '.[-1].total_cost_usd'

Message Types

For complete parsing guide: See references/response-parsing.md

Best Practices

1. Budget limits: Always --max-budget-usd for autonomous agents
2. Tool restrictions: Minimal toolset with --tools
3. Structured output: --json-schema for reliable parsing
4. Sandboxing: --dangerously-skip-permissions only in isolated environments
5. Error handling: Check is_error and exit codes