Claude Code CLI Hidden Flags

Claude Code CLI Hidden Flags

A hidden flag works, but claude --help does not show it. Most of them serve the SDK, the desktop app, the IDE extensions, or tests. For example, claude -p --max-turns 3 "Fix the failing test" stops after three agent turns, and the docs list --max-turns although the help does not.

About the Table

A row has the mark ✅ when a docs page names the flag, even though the help hides it. A row has the mark 🔍 when only the bundle has the flag, or when only the changelog and the Agent SDK pages name it. Treat a 🔍 flag as unstable, because Anthropic can change it without notice.

Flag	Mark	What it does
-d2e, --debug-to-stderr	🔍	Turns on debug mode and writes to stderr, and the flag is deprecated.
--init	✅	Runs Setup hooks with the init trigger and then goes on.
--init-only	✅	Runs Setup and SessionStart:startup hooks and then exits.
--maintenance	✅	Runs Setup hooks with the maintenance trigger and then goes on.
--session-mirror	🔍	Writes transcript_mirror frames to stdout, for the SDK.
--await-claim	🔍	Starts as a spare process for an SDK host and waits for a claim_session request.
--await-initialize	🔍	Reads the initialize request from stdin at startup so its fields apply like flags.
--thinking <mode>	🔍	Sets the thinking mode: enabled (the same as adaptive) or disabled.
--thinking-display <display>	🔍	Sets how the thinking text appears in the response.
--max-thinking-tokens <tokens>	🔍	Sets a thinking token limit, and the flag is deprecated in favor of --thinking.
--max-turns <turns>	✅	Stops after a number of agent turns, with --print only.
--task-budget <tokens>	🔍	Sets a task budget in tokens on the API side (output_config.task_budget).
--enable-auth-status	🔍	Turns on auth status messages in SDK mode.
--permission-prompt-tool <tool>	✅	Names an MCP tool that answers permission prompts, with --print only.
--system-prompt-file <file>	✅	Reads the system prompt from a file.
--append-system-prompt-file <file>	✅	Reads text from a file and adds it to the end of the default system prompt.
--append-subagent-system-prompt <prompt>	✅	Adds text to the system prompt of every subagent, with --print only.
--append-subagent-system-prompt-file <file>	✅	Reads text from a file and adds it to every subagent prompt, with --print only.
--plan-mode-instructions <instructions>	🔍	Replaces the default steps of plan mode with your own text.
--inherit-permission-mode <mode>	🔍	Sets a permission mode from a parent session, used only when nothing else sets one.
--watch-artifact <artifact>	🔍	Watches a Claude artifact for new versions and comments.
--watch-artifact-no-autoreact <artifact>	🔍	Watches an artifact and does not react to it by itself.
--prefill <text>	🔍	Fills the prompt box with text and does not send it.
--deep-link-origin	🔍	Says that a deep link started this session.
--deep-link-repo <slug>	🔍	Gives the repo slug that the deep link resolved to the current folder.
--deep-link-last-fetch <ms>	🔍	Gives the time of the last fetch of the repo, in milliseconds.
--prefill-b64 <b64>	🔍	Fills the prompt box with base64url text, for deep links.
--deep-link-cwd-b64 <b64>	🔍	Gives the working folder as base64url text, for deep links.
--resume-session-at <message id>	🔍	Resumes only up to one message of the chain, with --resume in print mode.
--resume-drops-turn <message id>	🔍	Names the turn that a truncating resume drops, and the resume fails if more would go.
--reply-on-resume	🔍	Asks the model to reply at once when the transcript ends with a user message.
--rewind-files <user-message-id>	🔍	Restores the files to the state at a user message and exits, and it needs --resume.
--include-unsupported-models	🔍	Offers models that your account can use and this version does not know.
--fast-mode <on|off>	🔍	Turns fast mode on or off.
--workload <tag>	🔍	Sets a workload tag for the attribution block, with --print only.
--client-data-url <url>	🔍	Loads a signed configuration document from a URL and exits if it fails.
--allow-other-models	🔍	With --client-data-url, starts on a model that the document does not list.
--managed-settings <json>	🔍	Gives policy settings as JSON from a parent process, for the SDK.
--host-auth-is-org-sign-in	🔍	Says that the token of the host is an organization sign-in, for desktop hosts.
--project-config-root <dir>	🔍	Reads project settings and the .claude trees from another folder.
--plugin-dir-no-mcp <path>	🔍	Works like --plugin-dir but does not read the .mcp.json of the plugin.
--advisor <model>	✅	Turns on the server-side advisor tool with the model that you name.
--enable-auto-mode	✅	Opts in to auto mode, and the flag is deprecated.
--proactivity <level>	🔍	Sets the level of proactivity, and the bundle gives no help text.
--messaging-socket-path <path>	🔍	Sets the socket or pipe path of the cross-session messaging server.
--channels <servers...>	✅	Lists the MCP servers whose channel notifications may register this session.
--dangerously-load-development-channels <servers...>	✅	Loads channel servers that are not on the allowlist, for local development only.
--agent-id <id>	🔍	Sets the ID of a teammate agent.
--agent-name <name>	🔍	Sets the display name of a teammate.
--team-name <name>	🔍	Sets the team name for teammate coordination.
--agent-color <color>	🔍	Sets the UI color of a teammate.
--plan-mode-required	🔍	Requires plan mode before the teammate implements anything.
--parent-session-id <id>	🔍	Gives the parent session ID for analytics.
--teammate-mode <mode>	✅	Sets the mode of a teammate, and the bundle gives no help text.
--agent-type <type>	🔍	Sets a custom agent type for a teammate.
--sdk-url <url>	🔍	Uses a WebSocket endpoint for SDK input and output, with -p and stream-json.
--forward-home-settings <true|false>	🔍	Sets whether this machine sends its settings to the cloud session.
--remote [description|session_id|url]	✅	Is a deprecated alias for --cloud.
--attach-serve <session_id>	🔍	Attaches a helper to a cloud session, and the desktop app starts it.
--pool <pool_id>	🔍	Is a deprecated alias for --environment.
--correlation-id <id>	🔍	Sets an ID that goes back to the environment orchestrator, and it needs --environment.
--ref <ref>	✅	Sets the branch, tag, or SHA to check out in the remote session, and it needs --cloud or --environment.
--on-branch <branch>	🔍	Works directly on a branch in the remote session, with checkout and push.
--rc [name]	✅	Is an alias for --remote-control.

See Also

• CLI
• CLI Flags