Claude Code as a Unix-Style Tool

Claude Code is not only an interactive terminal assistant. In print mode, it can behave like a command-line tool: read stdin, process a prompt, write stdout, and exit.

This is useful for logs, diffs, release notes, CI checks, and small repeatable analyses.

Basic Pattern

git log --oneline -20 | claude -p "Summarize these recent commits"

The -p flag is short for print mode. Use it when you want a non-interactive result instead of an ongoing terminal session.

Good Uses

Use case	Example
Commit summary	`git log —oneline -20
Build failure explanation	`npm test 2>&1
Release note draft	`git diff v1.2.0..HEAD
PR risk scan	`git diff main
Structured extraction	`claude -p "extract changed modules" --output-format json`

Use `--bare` For Scripts

For reproducible scripts and CI jobs, prefer --bare when you do not want local memory, hooks, skills, plugins, or MCP configuration to affect the run.

git diff main | claude --bare -p \
  "Review this diff for typo-level issues only. Return filename and issue."

Use normal print mode without --bare when the run should honor project instructions, local tools, skills, or memory.

Structured Output

For scripts, plain text is fragile. Prefer JSON when another command needs to parse the result.

git diff main | claude --bare -p \
  "Summarize the diff in one paragraph" \
  --output-format json | jq -r '.result'

For stricter automation, use a JSON schema if your installed Claude Code version supports it.

Tool Permissions

If all context is piped in, Claude may not need permission to run commands.

If you ask it to inspect files or run tests, allow only the tools and commands needed:

claude -p "Run the focused auth tests and explain any failures" \
  --allowedTools "Read,Bash(npm test *)"

Avoid broad command allowlists in unattended scripts. For cloud CLIs, deployment tools, and database commands, prefer read-only commands, dry runs, and least-privilege credentials.

When Not To Use Print Mode

Use an interactive session instead when:

the task needs back-and-forth clarification
the agent may need to edit many files
permissions require human judgment
you need to review a plan before implementation
the input is too large for stdin and should be referenced as files