Using Claude Code CLI

Patterns for programmatically invoking Claude Code CLI from orchestrators, scripts, and automation pipelines.

Contents

• Quick Start Checklist
• Essential Commands
• Core CLI Flags
• Sandbox Mode
• Validation
• Troubleshooting
• When Not to Use

Quick Start Checklist

Copy and track as you work:

CLI Automation Setup:
- [ ] 1. Identify required tools, add to --allowedTools
- [ ] 2. List directories Claude needs, add to --add-dir
- [ ] 3. Choose pattern: sync, async, or parallel (see references/)
- [ ] 4. Configure hooks if logging/monitoring needed
- [ ] 5. Enable sandbox mode for untrusted operations
- [ ] 6. Add fallback strategy if using OpenCode backup
- [ ] 7. Test with simple prompt before full automation
- [ ] 8. Validate setup (see Validation section)

Essential Commands

Minimal automation:

bash
claude -p "Your prompt" --allowedTools Write Read Edit --max-turns 5

Full automation:

bash
claude \
  --model sonnet \
  --add-dir /path/to/skills \
  --add-dir /path/to/templates \
  --allowedTools Write Read Edit Bash Task \
  --output-format json \
  -p "Your automation prompt"

With settings JSON:

bash
claude --settings '{"hooks":{},"sandbox":{"enabled":true}}' -p "prompt"

Core CLI Flags

| Flag | Purpose | Example | |------|---------|---------| | -p | Non-interactive mode (required) | claude -p "query" | | --allowedTools | Pre-approve tools | --allowedTools Write Read Bash | | --add-dir | Grant directory access | --add-dir ../lib | | --settings | Pass hooks/permissions JSON | --settings ./settings.json | | --model | Select model | --model sonnet | | --output-format | Output: text, json, stream-json | --output-format json | | --max-turns | Limit agentic turns | --max-turns 5 |

Recommended tools for automation:

• Core: Write, Read, Edit, Glob, Grep, LS
• Automation: Bash, Task, Skill
• Web: WebFetch, WebSearch

MCP tools require full path:

bash
--allowedTools "mcp__perplexity-ask__perplexity_ask"
--allowedTools "mcp__context7__get-library-docs"

See references/cli-reference.md for complete flag documentation.

Sandbox Mode

Enable sandbox for safe file operations:

Via settings JSON:

bash
claude --settings '{"sandbox":{"enabled":true,"autoAllowBashIfSandboxed":true}}' -p "prompt"

Via stdin (pass /sandbox before prompt):

python
stdin_content = "/sandbox\nYour actual prompt"
subprocess.run(["claude", "-p"], input=stdin_content, text=True)

See references/hooks-examples.md for full sandbox options.

Validation

Before full automation, verify setup works:

bash
# 1. Test tool pre-approval (should complete without prompts)
claude -p "Create file test.txt with 'hello'" --allowedTools Write

# 2. Test directory access
claude -p "List files in /path/to/dir" --add-dir /path/to/dir --allowedTools LS

# 3. Test JSON output
claude -p "Return {\"status\": \"ok\"}" --output-format json | jq .

# 4. Test hooks (check log file after)
claude --settings '{"hooks":{...}}' -p "Simple task" --allowedTools Read

If any step fails, see Troubleshooting.

Troubleshooting

| Problem | Solution | |---------|----------| | Permission prompts appearing | Verify tool names match exactly (case-sensitive); MCP tools need full path | | Hooks not firing | Check script is executable (chmod +x), has shebang, timeout not exceeded | | Subprocess hangs | Add timeout; check if waiting for permission (missing --allowedTools) | | Working directory issues | Use absolute paths in --add-dir; set cwd in subprocess | | JSON parse errors | Use --output-format json; see json-extraction.md |

When Not to Use

This skill is for subprocess invocation. Do not use for:

• Interactive Claude sessions (use Claude directly)
• API integration (use Claude SDK/API)
• MCP server setup (use claude mcp command)

Reference Files

| File | Contents | |------|----------| | cli-reference.md | Complete CLI commands and flags | | subprocess-patterns.md | Python sync/async/parallel patterns | | hooks-examples.md | Hook scripts and settings JSON | | json-extraction.md | Parsing JSON from CLI output | | orchestrator_example.py | Complete Python orchestrator class |

OpenCode CLI Fallback

When using OpenCode as fallback, note key differences:

| Feature | Claude CLI | OpenCode CLI | |---------|------------|--------------| | Headless | claude -p "prompt" | opencode run "prompt" | | Model | --model sonnet | --model provider/model | | Settings | --settings JSON | Not supported | | Add dirs | --add-dir /path | Not supported |

Try-Claude-then-OpenCode pattern:

python
import subprocess

def invoke_with_fallback(prompt: str, timeout: int = 300) -> str:
    """Try Claude CLI first, fall back to OpenCode on failure."""
    try:
        result = subprocess.run(
            ["claude", "-p", prompt, "--allowedTools", "Write", "Read"],
            capture_output=True, text=True, timeout=timeout
        )
        if result.returncode == 0:
            return result.stdout
    except (subprocess.TimeoutExpired, FileNotFoundError):
        pass  # Fall through to OpenCode

    result = subprocess.run(
        ["opencode", "run", prompt],
        capture_output=True, text=True, timeout=timeout
    )
    return result.stdout

See subprocess-patterns.md for advanced patterns.