Non-Interactive & Headless Mode

Base flag: -p

The -p flag (or --print) activates non-interactive mode. Claude returns the response in the terminal and exits immediately without waiting for user input.

# Direct prompt
claude -p "Analyze this file and list the bugs"

# With pipe stdin (content becomes the input)
git diff | claude -p "Explain these changes"
cat error.log | claude -p "Identify the root cause"

Output formats

--output-format controls the response structure, which is essential for integration into scripts.

# JSON output for automatic parsing
git status --short | claude -p "Categorize the changes" \
  --output-format json | jq -r '.result'

# Stream JSON for long operations
cat report.txt | claude -p "Summarize" --output-format stream-json --verbose

Flow control in pipes

In print mode, claude -p waits for the complete response and writes it to stdout in a single block, which simplifies pipes with tools that expect complete input. Real-time streaming is opt-in via --output-format stream-json.

claude -p "Generate a report" > report.md

CI/CD usage

In an isolated container, --dangerously-skip-permissions suppresses all confirmation requests. Reserve exclusively for sandboxed environments where Claude cannot access sensitive resources.

# GitHub Actions (isolated container)
claude -p "Run the tests and fix failures" \
  --dangerously-skip-permissions \
  --output-format json

Common patterns

# Automated log analysis
tail -n 200 app.log | claude -p "Critical alerts only" \
  --output-format json

# PR review (package.json integration)
git diff main...HEAD | claude -p "Security review, JSON format" \
  --output-format json > review.json

Interactive vs non-interactive

Best practice: limit pipe size to avoid exceeding the context window. Filter with head, grep or --name-only before sending to Claude.