Claude Headless

Programmatic usage of Claude Code CLI. Three modes available:

Quick Start

One-shot — single prompt, get result, exit:

claude -p "Explain this error" --output-format json

Stream JSON — bidirectional real-time NDJSON over stdin/stdout:

claude -p --output-format stream-json --input-format stream-json --verbose
# Send: {"type":"user","message":{"role":"user","content":"hello"}}
# Recv: {"type":"assistant","message":{...}}
# Recv: {"type":"result","subtype":"success",...}

Resume — continue a previous session:

claude -p "continue the review" --resume <session-id> --output-format json

Key Flags

Flag	Purpose
-p / --print	Non-interactive mode
--output-format stream-json	NDJSON event stream on stdout
--input-format stream-json	NDJSON messages on stdin (multi-turn)
--bare	Skip CLAUDE.md, hooks, skills, plugins, MCP, auto-memory, OAuth/keychain (fast CI; see auth.md)
--allowedTools "Bash,Read"	Pre-approve specific tools
--max-budget-usd N	Cost cap
--json-schema '{...}'	Structured output extraction

For complete flag reference, read references/cli-flags.md.

Deep References

Load these on demand based on what the user needs:

• references/cli-flags.md — Complete CLI flags: headless, permissions, budget, session, system prompt, subcommands
• references/auth.md — Authentication: subscription OAuth, API key, apiKeyHelper, --bare caveat, token storage
• references/env-vars.md — Complete environment variables: auth, models, thinking, bash/network timeouts, telemetry, TLS, session, plugins, MCP, detection
• references/stream-json.md — Stream JSON protocol: stdout events, stdin messages, pipe-chaining
• references/agent-view.md — Background sessions & claude agents: --bg, supervisor, shell management commands, peek/attach, worktree isolation, vs claude-dm
• references/permissions.md — Headless permission modes, --allowedTools, --permission-prompt-tool MCP
• references/sdks.md — TypeScript and Python Agent SDKs: API, options, streaming, multi-turn
• references/patterns.md — Real-world patterns: CI/CD, bot bridges, structured extraction, Agent Teams headless