CLI flags set the behavior of a specific run and split by purpose just like everything else in Claude Code. First you distinguish the run forms: an interactive session with the claude command, a non-interactive run claude -p and subcommands like mcp, doctor, agents, project, auth, plugin and update. The form determines which flags make sense at all. This reference groups the most-used flags; the full and exact list for your version is in the official CLI reference and via claude --help.

The context and permission flags are the most important for security. --permission-mode sets the mode for the run, --allowedTools and --disallowedTools give pointed tool control with the permission-rule syntax, --add-dir gives access to an additional directory, --dangerously-skip-permissions enables bypass (and is named dangerous deliberately). --settings, --setting-sources, --mcp-config, --agents, --plugin-dir and --plugin-url control which configuration is picked up - especially together with --bare, which disables auto-discovery, and --safe-mode for diagnostics.

The automation flags determine determinism. --output-format chooses text, json or stream-json; --json-schema constrains structured output; --max-turns and --max-budget-usd cap turns and cost; --continue and --resume control session continuity; --verbose and --debug reveal diagnostics. Separately stand --worktree for an isolated tree, --background for a background session and --remote-control for remote control of local work. Below is a map of the key flags and subcommands by group.

As with commands, a flag's presence in the reference does not replace a check. Flags, models and their behavior change between versions, so before wiring a flag into a script or CI you check claude --help and the official reference - especially those with the words dangerously, skip or bypass. The exact --allowedTools syntax matches the permission rules from the corresponding chapter; when in doubt it is better to verify the actual behavior once with a narrow run than to build automation on an assumption.