Headless mode

Run Claude Code non-interactively for scripts, CI pipelines, and programmatic usage.

Headless mode lets you run Claude Code without an interactive terminal session. Pass --print (or -p) and Claude outputs its response to stdout then exits — no UI, no prompts, no spinner.

Basic usage

# Inline prompt
claude -p "Summarise the changes in the last git commit"

# Prompt as positional argument
claude --print "List all TODO comments in src/"

# Pipe from stdin
echo "What does this regex match?" | claude -p

The workspace trust dialog is skipped when using -p. Only run Claude Code with --print in directories you trust.

Output formats

Control how Claude’s response is serialised with --output-format. This flag only takes effect when --print is also set.

text (default)

claude -p "What is 2 + 2?" --output-format text
# Outputs: 4

JSON output structure

With --output-format json, Claude Code writes a single JSON object to stdout after the run completes:

{
  "type": "result",
  "subtype": "success",
  "result": "The final assistant response text",
  "num_turns": 3,
  "total_cost_usd": 0.0042,
  "duration_ms": 8200,
  "is_error": false,
  "stop_reason": "end_turn",
  "usage": {
    "input_tokens": 1500,
    "output_tokens": 420,
    "cache_read_input_tokens": 800,
    "cache_creation_input_tokens": 0
  }
}

On error, subtype is one of error_during_execution, error_max_turns, or error_max_budget_usd, and an errors array is included.

stream-json event stream

With --output-format stream-json, Claude Code emits newline-delimited JSON events as they happen:

{"type":"system","subtype":"init","model":"claude-sonnet-4-6","tools":["Bash","Read","Write"],"session_id":"..."}
{"type":"assistant","message":{...},"session_id":"..."}
{"type":"tool_progress","tool_name":"Bash","tool_use_id":"...","session_id":"..."}
{"type":"result","subtype":"success","result":"Done.","total_cost_usd":0.002,"session_id":"..."}

Pass --include-partial-messages to also receive partial assistant message chunks as they stream in.

Tool control

--allowedTools

Restrict which tools Claude can use. Accepts a space- or comma-separated list. Supports permission rule syntax.

claude -p "Run the tests" --allowedTools "Bash(npm test)" "Read"

--disallowedTools

Block specific tools entirely.

claude -p "Review this code" --disallowedTools "Bash" "Write"

--tools

Specify the complete set of built-in tools available. Use "" to disable all tools, "default" to restore the full set.

# Only file reading tools
claude -p "Explain the architecture" --tools "Read,Glob,Grep"

# No tools at all (pure text conversation)
claude -p "What is a monad?" --tools ""

Model and limits

# Use a specific model
claude -p "Write a haiku" --model claude-sonnet-4-6

# Limit agentic turns (exits early after N turns)
claude -p "Fix all lint errors" --max-turns 10

# Cap spend
claude -p "Large refactoring task" --max-budget-usd 0.50

MCP servers

Load MCP server configuration for a headless run with --mcp-config. Accepts one or more JSON file paths or inline JSON strings.

# From a file
claude -p "Query the database" --mcp-config ./mcp-servers.json

# Inline JSON
claude -p "Use the custom tool" --mcp-config '{"myserver":{"type":"stdio","command":"node","args":["server.js"]}}'

# Use only the servers from --mcp-config, ignoring all other MCP config
claude -p "Query the database" --mcp-config ./mcp-servers.json --strict-mcp-config

Using in scripts

Shell script

#!/usr/bin/env bash
set -euo pipefail

RESPONSE=$(claude -p "What files changed in the last commit?" --output-format text)
echo "Claude says: $RESPONSE"

Using in CI pipelines

GitHub Actions

- name: Review PR changes
  run: |
    git diff HEAD~1 | claude -p "Review these changes for bugs and suggest improvements" \
      --output-format json \
      --allowedTools "" \
      --max-turns 5 \
      > review.json

    cat review.json | jq -r '.result'
  env:
    ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}

Additional flags

Flag	Description
--system-prompt <text>	Override the system prompt
--append-system-prompt <text>	Append to the default system prompt
--permission-mode <mode>	Set permission mode: default, acceptEdits, plan, bypassPermissions, or dontAsk
--no-session-persistence	Disable saving the session to disk
--session-id <uuid>	Use a specific session UUID
-c, --continue	Continue the most recent session in the current directory
-r, --resume <id>	Resume a specific session by ID
--settings <file-or-json>	Load additional settings from a file or inline JSON
--add-dir <directories...>	Grant tool access to additional directories
--bare	Minimal mode — skip hooks, LSP, plugins, keychain reads, and CLAUDE.md discovery