Claude in scripts

Headless mode is what makes Claude Code scriptable. A single command like

claude -p "Summarize this log in one line" --output-format json

runs Claude without an interactive session, prints a result, and exits. That makes it a normal building block you can drop into a shell script, a cron job, or a CI pipeline.

Parse the output

With --output-format json the reply comes back as structured JSON, so you can pull out exactly the field you want with jq:

claude -p "Classify this line as INFO or ERROR: $LINE" \
  --output-format json | jq -r '.result'

Now the answer is a plain string your script can act on: write it to a file, branch on it, or feed it to the next step.

Rules for unattended runs

When nothing is watching the terminal, a few habits keep automation safe and predictable:

• Grant permissions up front. A scheduled run cannot answer a permission prompt, so decide what it may do before it starts. Use a scoped allowlist in settings, or a permission mode like acceptEdits when it needs to write files. Reach for bypassPermissions only in a throwaway sandbox, never against anything real.
• Keep prompts tight. One clear task per call. Say what output you want and in what shape.
• Log what it did. Redirect output to a file with a timestamp so you can see later what ran and what it decided.

It does not wait for you

The point of headless mode is that it runs non-interactively. Nothing pauses for a human to type y. That is exactly what you want for automation, and also exactly why you scope its permissions carefully before you schedule it.

Your lab setup

This VM comes with Claude Code preinstalled and signed in - just run claude to start.