Flags
Launch-time options that configure a session before it even starts — which model, which tools, how verbose, how permissive.
Flags are options passed when you start Claude Code that shape that session's behavior: --model to choose which model runs, --allowedTools to control which tools are available without prompting, --verbose for detailed output of everything Claude does, and --dangerously-skip-permissions as the (discouraged) permission-free option. They're optional shortcuts, not requirements — you can configure all the same things interactively once a session has started.
Newer additions worth knowing: --worktree / -w to launch directly into an isolated git worktree (see Worktrees), and --bare, which skips auto-discovery of hooks, skills, custom commands, subagents, plugins, MCP servers and CLAUDE.md for a faster, more predictable start — useful for CI and scripted runs where you want the exact same behavior on every machine.
Real-life use cases
Consistent CI runs
A script that runs Claude Code in a CI pipeline uses --bare plus explicit --allowedTools, so the run behaves identically regardless of what's configured on the machine that happens to execute the pipeline.
Starting straight into a parallel task
Instead of starting a normal session and asking Claude to set up isolation afterward, a developer launches directly with claude --worktree bugfix-login to begin already isolated from their main checkout.
Source
- Source video: “Claude Code Explained” — 27 concepts in 27 minutes (YouTube)https://www.youtube.com/watch?v=ZlDnsf_DOzg
- CLI referencehttps://code.claude.com/docs/en/cli-reference