All Flags

A flag changes how one claude run behaves. claude --help shows 66 flags. Another 64 work but stay hidden from the help text. This page lists both kinds in one place, grouped by what they do.

Pass a flag after claude, for example claude --model opus. Most flags last for one run only. A flag is on when you pass it, and some have a --no- twin, such as --no-chrome.

How to Read the Tables

• Documented means an official docs page names the flag, for example CLI Reference. The column In --help says whether the help text shows it.
• Undocumented means only the program has it. It can change or vanish in any release. Treat these flags as unstable.
• Hidden flags mostly serve the Agent SDK, the desktop app, IDE extensions, and teammate processes.

Example

# One answer, JSON out, with a spend cap
claude -p "Summarize README.md" --output-format json --max-budget-usd 1

# A careful planning session on a chosen model
claude --model opus --effort high --permission-mode plan

# Stop after three agent turns
claude -p --max-turns 3 "Fix the failing test"

Flags by Job

Start and Print Mode

Flags that choose how a run starts and how print mode reads input and writes output.

Flag	Status	In --help	What it does
-p, --print	Documented	✅	Prints the answer and exits, which suits pipes, and the trust dialog is skipped in this mode.
--output-format <format>	Documented	✅	Sets the output of print mode to text, json, or stream-json.
--input-format <format>	Documented	✅	Sets the input of print mode to text or stream-json.
--json-schema <schema>	Documented	✅	Checks the final answer against a JSON Schema. It works only with --print.
--include-hook-events	Documented	✅	Adds the hook events to the stream, with --output-format=stream-json.
--include-partial-messages	Documented	✅	Adds partial message chunks to the stream, with --print and stream-json.
--forward-subagent-text	Documented	✅	Adds the text of subagents to the stream as assistant and user messages, with --print and stream-json.
--replay-user-messages	Documented	✅	Sends the user messages from stdin back on stdout as an acknowledgment.
--max-budget-usd <amount>	Documented	✅	Stops at a spend limit in dollars, with --print only.
--max-turns <turns>	Documented	❌	Stops after a number of agent turns, with --print only.
--no-session-persistence	Documented	✅	Saves nothing to disk, so the session cannot be resumed, with --print only.
--prompt-suggestions [value]	Documented	✅	Turns on prompt suggestions, and print mode sends a prompt_suggestion message after each turn.
--init	Documented	❌	Runs Setup hooks with the init trigger and then goes on.
--init-only	Documented	❌	Runs Setup and SessionStart:startup hooks and then exits.
--maintenance	Documented	❌	Runs Setup hooks with the maintenance trigger and then goes on.
--session-mirror	Undocumented	❌	Writes transcript_mirror frames to stdout, for the SDK.
--await-claim	Undocumented	❌	Starts as a spare process for an SDK host and waits for a claim_session request.
--await-initialize	Undocumented	❌	Reads the initialize request from stdin at startup so its fields apply like flags.
--sdk-url <url>	Undocumented	❌	Uses a WebSocket endpoint for SDK input and output, with -p and stream-json.
--enable-auth-status	Undocumented	❌	Turns on auth status messages in SDK mode.
--workload <tag>	Undocumented	❌	Sets a workload tag for the attribution block, with --print only.
--client-data-url <url>	Undocumented	❌	Loads a signed configuration document from a URL and exits if it fails.
--allow-other-models	Undocumented	❌	With --client-data-url, starts on a model that the document does not list.
--managed-settings <json>	Undocumented	❌	Gives policy settings as JSON from a parent process, for the SDK.
--host-auth-is-org-sign-in	Undocumented	❌	Says that the token of the host is an organization sign-in, for desktop hosts.

Safe Modes, Settings, and Debug

Flags that strip your setup down, load extra settings, or turn on logs.

Flag	Status	In --help	What it does
-h, --help	Documented	✅	Shows the help text and exits.
-v, --version	Documented	✅	Prints the version number.
-d, --debug [filter]	Documented	✅	Turns on debug logs, and an optional filter picks the categories, such as --debug=hooks,mcp or --debug='!hooks,!mcp'.
-d2e, --debug-to-stderr	Undocumented	❌	Turns on debug mode and writes to stderr, and the flag is deprecated.
--debug-file <path>	Documented	✅	Writes the debug logs to a file and turns on debug mode.
--verbose	Documented	✅	Overrides the verbose setting of your config.
--bare	Documented	✅	Starts with a minimal setup: no hooks, no plugin sync, no auto memory, and no background work.
--safe-mode	Documented	✅	Starts with all customizations off, such as CLAUDE.md, skills, plugins, hooks, and MCP servers.
--restricted	Documented	✅	Removes the tools that run code and WebFetch, and ignores user, project, and local settings.
--ax-screen-reader	Documented	✅	Draws flat text with no borders or animations, for screen readers.
--settings <file-or-json>	Documented	✅	Loads more settings from a JSON file or a JSON string.
--setting-sources <sources>	Documented	✅	Chooses which setting sources load: user, project, and local.
--project-config-root <dir>	Undocumented	❌	Reads project settings and the .claude trees from another folder.

Permissions and Tools

Flags that set the permission mode and the tools the model may use. See Permissions.

Flag	Status	In --help	What it does
--permission-mode <mode>	Documented	✅	Sets the permission mode of the session.
--inherit-permission-mode <mode>	Undocumented	❌	Sets a permission mode from a parent session, used only when nothing else sets one.
⚠️ --dangerously-skip-permissions	Documented	✅	Skips all permission checks, and you should use it only in a sandbox with no internet access.
--allow-dangerously-skip-permissions	Documented	✅	Makes skipping all permission checks a choice, without turning it on.
--allowedTools, --allowed-tools <tools...>	Documented	✅	Lists the tools that run without a permission prompt.
--disallowedTools, --disallowed-tools <tools...>	Documented	✅	Lists the tools that the model cannot use.
--tools <tools...>	Documented	✅	Lists the built-in tools that the model can use. "default" uses the default set.
--permission-prompts <target>	Documented	✅	Sets who answers permission prompts: host or none.
--permission-prompt-tool <tool>	Documented	❌	Names an MCP tool that answers permission prompts, with --print only.
--plan-mode-instructions <instructions>	Undocumented	❌	Replaces the default steps of plan mode with your own text.

Model and Thinking

Flags that pick the model and set how hard it thinks.

Flag	Status	In --help	What it does
--model <model>	Documented	✅	Sets the model with an alias such as opus or with a full name.
--effort <level>	Documented	✅	Sets the effort level of the session.
--fallback-model <model>	Documented	✅	Sets one or more models to try when the main model is overloaded or not available.
--betas <betas...>	Documented	✅	Adds beta headers to API requests, for API key users only.
--advisor <model>	Documented	❌	Turns on the server-side advisor tool with the model that you name.
--thinking <mode>	Undocumented	❌	Sets the thinking mode: enabled (the same as adaptive) or disabled.
--thinking-display <display>	Undocumented	❌	Sets how the thinking text appears in the response.
⛔ --max-thinking-tokens <tokens>	Undocumented	❌	Sets a thinking token limit, and the flag is deprecated in favor of --thinking.
--task-budget <tokens>	Undocumented	❌	Sets a task budget in tokens on the API side (output_config.task_budget).
--fast-mode <on|off>	Undocumented	❌	Turns fast mode on or off.
--include-unsupported-models	Undocumented	❌	Offers models that your account can use and this version does not know.
--autocompact <auto|tokens>	Documented	✅	Sets the auto-compact window to auto or to 100k to 1M tokens.

System Prompt

Flags that replace or extend the system prompt. See System Prompt.

Flag	Status	In --help	What it does
--system-prompt <prompt>	Documented	✅	Replaces the system prompt for the session.
--system-prompt-file <file>	Documented	❌	Reads the system prompt from a file.
--append-system-prompt <prompt>	Documented	✅	Adds text to the end of the default system prompt.
--append-system-prompt-file <file>	Documented	❌	Reads text from a file and adds it to the end of the default system prompt.
--append-subagent-system-prompt <prompt>	Documented	❌	Adds text to the system prompt of every subagent, with --print only.
--append-subagent-system-prompt-file <file>	Documented	❌	Reads text from a file and adds it to every subagent prompt, with --print only.
--system-prompt-snapshot <on|off>	Documented	✅	Sets whether the system prompt is recorded once and reused on every request and resume.
--exclude-dynamic-system-prompt-sections	Documented	✅	Moves the data of the machine from the system prompt to the first user message for better cache reuse.

Sessions and Worktrees

Flags that continue, fork, name, or rewind a session, or run it in a worktree or the background. See Sessions and Storage.

Flag	Status	In --help	What it does
-c, --continue	Documented	✅	Continues the newest conversation of the current folder.
-r, --resume [value]	Documented	✅	Resumes a conversation by session ID or name, or opens the picker.
--fork-session	Documented	✅	Creates a new session ID when you resume, with --resume or --continue.
--from-pr [value]	Documented	✅	Resumes a session that links to a pull request.
--session-id <uuid>	Documented	✅	Uses a UUID that you give as the session ID.
-n, --name <name>	Documented	✅	Sets the display name of the session.
--resume-session-at <message id>	Undocumented	❌	Resumes only up to one message of the chain, with --resume in print mode.
--resume-drops-turn <message id>	Undocumented	❌	Names the turn that a truncating resume drops, and the resume fails if more would go.
--reply-on-resume	Undocumented	❌	Asks the model to reply at once when the transcript ends with a user message.
--rewind-files <user-message-id>	Undocumented	❌	Restores the files to the state at a user message and exits, and it needs --resume.
--bg, --background	Documented	✅	Starts the session in the background and prints its ID.
-w, --worktree [name]	Documented	✅	Creates a new git worktree for the session.
--tmux	Documented	✅	Creates a tmux session for the worktree, and it needs --worktree.

Agents, Plugins, MCP, and Context

Flags that add agents, plugins, MCP servers, folders, and files.

Flag	Status	In --help	What it does
--agent <agent>	Documented	✅	Sets the agent of the session and overrides the agent setting.
--agents <json-or-file>	Documented	✅	Defines custom agents with JSON or a file.
--mcp-config <configs...>	Documented	✅	Loads MCP servers from JSON files or strings.
--strict-mcp-config	Documented	✅	Uses only the MCP servers from --mcp-config.
--plugin-dir <path>	Documented	✅	Loads a plugin from a folder or a .zip file for this session only.
--plugin-url <url>	Documented	✅	Fetches a plugin .zip file from a URL for this session only.
--plugin-dir-no-mcp <path>	Undocumented	❌	Works like --plugin-dir but does not read the .mcp.json of the plugin.
--disable-slash-commands	Documented	✅	Turns off all skills and commands.
--add-dir <directories...>	Documented	✅	Gives the tools access to more folders.
--brief	Undocumented	✅	Turns on the SendUserMessage tool so the agent can message the user.
--file <specs...>	Undocumented	✅	Downloads file resources at startup, in the form file_id:relative_path.
--channels <servers...>	Documented	❌	Lists the MCP servers whose channel notifications may register this session.
⚠️ --dangerously-load-development-channels <servers...>	Documented	❌	Loads channel servers that are not on the allowlist, for local development only.
--watch-artifact <artifact>	Undocumented	❌	Watches a Claude artifact for new versions and comments.
--watch-artifact-no-autoreact <artifact>	Undocumented	❌	Watches an artifact and does not react to it by itself.
--messaging-socket-path <path>	Undocumented	❌	Sets the socket or pipe path of the cross-session messaging server.
--proactivity <level>	Undocumented	❌	Sets the level of proactivity, and the bundle gives no help text.

Teammate Flags

Flags that a lead agent passes to a teammate process it starts. You rarely set them by hand.

Flag	Status	In --help	What it does
--agent-id <id>	Undocumented	❌	Sets the ID of a teammate agent.
--agent-name <name>	Undocumented	❌	Sets the display name of a teammate.
--team-name <name>	Undocumented	❌	Sets the team name for teammate coordination.
--agent-color <color>	Undocumented	❌	Sets the UI color of a teammate.
--plan-mode-required	Undocumented	❌	Requires plan mode before the teammate implements anything.
--teammate-mode <mode>	Documented	❌	Sets the mode of a teammate, and the bundle gives no help text.
--agent-type <type>	Undocumented	❌	Sets a custom agent type for a teammate.
--parent-session-id <id>	Undocumented	❌	Gives the parent session ID for analytics.

IDE, Desktop, and Deep Links

Flags that connect to an IDE, the desktop app, or a deep link.

Flag	Status	In --help	What it does
--ide	Documented	✅	Connects to the IDE at startup when exactly one valid IDE is open.
--desktop	Documented	✅	Opens the session in the Claude Desktop app instead of the terminal.
--chrome	Documented	✅	Turns on the Claude in Chrome integration.
--no-chrome	Documented	✅	Turns off the Claude in Chrome integration.
--prefill <text>	Undocumented	❌	Fills the prompt box with text and does not send it.
--prefill-b64 <b64>	Undocumented	❌	Fills the prompt box with base64url text, for deep links.
--deep-link-origin	Undocumented	❌	Says that a deep link started this session.
--deep-link-repo <slug>	Undocumented	❌	Gives the repo slug that the deep link resolved to the current folder.
--deep-link-last-fetch <ms>	Undocumented	❌	Gives the time of the last fetch of the repo, in milliseconds.
--deep-link-cwd-b64 <b64>	Undocumented	❌	Gives the working folder as base64url text, for deep links.

Cloud and Remote Sessions

Flags for cloud sessions and Remote Control.

Flag	Status	In --help	What it does
--teleport [session]	Documented	✅	Resumes a teleport session.
--cloud [description|session_id|url]	Documented	✅	Creates a cloud session, or attaches to one by session ID or claude.ai/code URL.
--environment <environment_id>	Documented	✅	Creates a cloud session on a self-hosted environment.
--remote-control [name]	Documented	✅	Starts an interactive session with Remote Control turned on.
--rc [name]	Documented	❌	Is an alias for --remote-control.
--remote-control-session-name-prefix <prefix>	Documented	✅	Sets the prefix of the names that Remote Control makes, and the default is the host name.
⛔ --remote [description|session_id|url]	Documented	❌	Is a deprecated alias for --cloud.
⛔ --pool <pool_id>	Undocumented	❌	Is a deprecated alias for --environment.
--correlation-id <id>	Undocumented	❌	Sets an ID that goes back to the environment orchestrator, and it needs --environment.
--ref <ref>	Documented	❌	Sets the branch, tag, or SHA to check out in the remote session, and it needs --cloud or --environment.
--on-branch <branch>	Undocumented	❌	Works directly on a branch in the remote session, with checkout and push.
--forward-home-settings <true|false>	Undocumented	❌	Sets whether this machine sends its settings to the cloud session.
--attach-serve <session_id>	Undocumented	❌	Attaches a helper to a cloud session, and the desktop app starts it.

Related Pages

• CLI Overview
• Subcommands
• Environment Variables for variables that match many flags.
• Permission Modes for the values of --permission-mode.