Claude Code CLI Flags Reference

Stop digging through --help every time

The Claude Code CLI is small on purpose, but the handful of flags it does expose change how a whole session behaves - which model you get, whether it asks permission, whether it runs once and exits or opens an interactive loop. Most people learn three flags and never discover the ones that would save them the most time. This reference puts the real, established flags in one searchable place with a working example for each, so you can build the exact command you need without leaving the page.

Every flag listed here is a stable, documented option. Where a detail is version-specific or could drift, this page deliberately points you back to `claude --help`, which is always the authoritative source for your installed version. The goal is a fast, accurate cheat sheet - not a guess.

The flags you will actually reach for

• -p / --print - run a prompt once, print the answer, and exit. This is the single most useful flag for scripting and CI, because it turns Claude Code into a normal command-line program you can pipe into and out of.
• -c / --continue and -r / --resume - pick up where you left off. --continue grabs the latest session in the current directory; --resume lets you choose from your history.
• --model - switch between Opus, Sonnet, and Haiku without leaving the terminal. Default to Sonnet, escalate to Opus for hard reasoning, drop to Haiku for cheap bulk work.
• --permission-mode - start in a specific safety posture, such as plan mode when you only want Claude to read and think before touching anything.
• --add-dir - grant access to sibling directories so Claude can work across a monorepo or a shared library without you copying files around.

Print mode is the power user's secret

The interactive session gets all the attention, but `claude -p` is where serious automation lives. Because it reads from stdin and writes to stdout, you can chain it into any pipeline: generate a commit message from a diff, summarize a log file, transform a data blob, or answer a question inside a shell script. Pair it with --output-format json when you need structured output that another program will parse, and --input-format stream-json when you are feeding structured input in.

Permission flags: convenience versus safety

The permission flags are the ones to understand before you use them. --allowedTools pre-approves a specific set of tools so a session runs without stopping to ask - great for trusted, repetitive work. --disallowedTools does the opposite, hard-blocking tools you never want to fire. And --dangerously-skip-permissions removes the safety net entirely. That last one is named the way it is for a reason: only use it inside a sandbox or a throwaway container where a mistaken command cannot hurt anything real.

1. Prefer --permission-mode plan when you want Claude to investigate and propose without editing.
2. Use --allowedTools to scope trust narrowly, like allowing only git commands.
3. Reserve --dangerously-skip-permissions for disposable, isolated environments.
4. Layer --disallowedTools on top to explicitly forbid destructive operations.

MCP and configuration from the command line

The `claude mcp` subcommands let you add, list, and remove MCP servers so Claude Code can talk to your databases, browsers, and internal tools. If you keep a per-project MCP setup, --mcp-config points the session at a specific JSON file. On the config side, `claude config` reads and writes settings without hand-editing files, and `claude doctor` diagnoses a broken install faster than trial and error. Every MCP server you load costs tokens on each turn, so keep the set lean.

How to keep this reference honest

Command-line tools evolve, and Claude Code ships often. Treat this page as a fast lookup, then trust `claude --help` and `claude --version` as the final word for whatever build you have installed. If a flag here behaves differently than described, your installed version is the source of truth - update with `claude update` and re-check the help output. That habit keeps you accurate and never caught out by a rename.