Claude Code cheatsheet

Every command on this page was read out of the installed binary (claude --help, v2.1.234) or the official docs - not from another cheat sheet. That distinction matters: the popular community sheets circulating GitHub list flags like --production-mode, --cache-results, and --security-enabled that do not exist, and mark --max-turns as a favorite without noting it vanished from --help output in 2.1.x (it still works - print mode only - and is marked docs where it appears below).

Learn the commands where mistakes are free:

A simulation, not the real binary - but every command below behaves the way Claude Code v2.1.234 does. Type / for the command menu, chat to fill the window, Shift+Tab to cycle modes, Esc Esc to rewind.

claude · demo-app — simulated

fableeffort high34%

Context window · 67k / 200k

System prompt3k
Tool definitions12k
Rules (CLAUDE.md)2k
MCP schemas (deferred)4k
Conversation46k

The seeded session left 46k of reads and test output in the window. Squeeze it (/compact), undo it (/rewind), or drop it (/clear).

Everything the sandbox knows, in reference form - search it (press / anywhere on this page to jump to the search box), or jump by section:

108 commands · verified against the installed binary v2.1.234 and the official docs · checked 2026-08-18 source source source

Install & keep current

Getting it on the machine and keeping it there.

Command	What it does	Notes
`curl -fsSL https://claude.ai/install.sh \| bash` docs	Native install for macOS, Linux, and WSL (the recommended path).	Windows PowerShell: `irm https://claude.ai/install.ps1 \| iex`
`brew install --cask claude-code` docs	Install via Homebrew on macOS.	Windows: `winget install Anthropic.ClaudeCode`
`npm install -g @anthropic-ai/claude-code` docs	Install via npm - a supported path, installs the same native binary.	Do not `sudo`; upgrade with `@latest`, not `npm update -g`.
`claude update`	Check for updates and install if available.	Alias: `claude upgrade`
`claude install <stable\|latest\|version>`	Install or pin a specific native build, e.g. `claude install 2.1.118`.
`claude doctor`	Check installation health without a trust prompt.	For a checkup that can also fix issues, run `/doctor` inside a session.
`claude --version`	Print the installed version.

Accounts, tokens, and enterprise auth.

Command	What it does	Notes
`claude auth login`	Sign in to your Anthropic account (opens a browser flow).
`claude auth status`	Show authentication status.
`claude auth logout`	Sign out.
`claude setup-token`	Set up a long-lived auth token (requires a Claude subscription).
`claude gateway`	Run the enterprise auth/telemetry gateway.

Start & steer a session

Interactive use: launching with the right shape of workspace.

Command	What it does	Notes
`claude`	Start an interactive session in the current directory.
`claude "summarize this project"`	Start with an initial prompt.
`claude -n <name>`	Name the session at startup (shows in the prompt bar, /resume picker, terminal title).
`claude --add-dir ../apps ../lib`	Grant tool access to additional working directories.chapter →
`claude -w <name>`	Create a git worktree for the session (add --tmux to open it in tmux/iTerm2 panes).
`claude --ide`	Auto-connect to an IDE on startup when exactly one is available.
`claude --bg`	Start the session as a background agent and return immediately; manage with `claude agents`.
`claude --safe-mode`	Start with all customizations (CLAUDE.md, skills, plugins, hooks, MCP) disabled - for troubleshooting a broken config.
`claude --bare`	Minimal mode: skips hooks, plugins, auto-memory, CLAUDE.md discovery; auth via API key only.
`claude import <source>`	Import config from another AI coding agent into Claude Code.

Get back to a session

Continuing, resuming, forking, and rewinding.

Command	What it does	Notes
`claude -c`	Continue the most recent conversation in this directory.	Long form: `claude --continue`
`claude -r <session>`	Resume by session ID or name; opens the picker with no argument.	Long form: `claude --resume`
`claude --fork-session`	When resuming, create a new session ID instead of reusing the original.	Use with `--resume` or `--continue`.
`claude --from-pr 123`	Resume the session linked to a pull request (by number or URL).
`claude --session-id <uuid>`	Start the conversation with a specific session ID.
`claude --teleport`	Pull a Claude Code on the web session into this terminal.
`/rename <name>`	Rename the current session; auto-generates a name with no argument.	Also works in `-p` mode from v2.1.205.
`/rewind`	Rewind conversation and/or code to a previous checkpoint - an undo, not a summary.chapter →	Aliases: `/checkpoint`, `/undo`. Also double-tap `Esc`.
`/recap`	One-line summary of the current session on demand.

Model & effort

Picking the brain and the throttle for the run.

Command	What it does	Notes
`claude --model <alias\|name>`	Model for the session - an alias (`opus`, `sonnet`) or a full name (`claude-fable-5`).chapter →
`claude --effort <level>`	Reasoning effort for the session: low, medium, high, xhigh, or max.
`claude --fallback-model <list>`	Comma-separated fallback models when the default is overloaded (print mode only).
`claude --max-budget-usd <amount>`	Cap API spend for the run (print mode only).
`/model <model>`	Switch model mid-session.
`/effort <level\|auto>`	Adjust reasoning effort mid-session.
`/fast <on\|off>`	Toggle fast mode.

Permissions & modes

What the agent may do without asking.

Command	What it does	Notes
`claude --permission-mode <mode>`	Start in a mode: acceptEdits, auto, bypassPermissions, manual, dontAsk, or plan.chapter →
`claude --allowedTools "Bash(git log:*)" "Read"`	Allow specific tools (with optional patterns) without prompting.	Hyphen spelling `--allowed-tools` also works.
`claude --disallowedTools "Bash(rm:)" "Bash(sudo:)"`	Deny specific tools for the run.
`claude --tools "Bash,Edit,Read"`	Restrict the session to a subset of built-in tools; `""` disables all.
`claude --dangerously-skip-permissions`	Bypass all permission checks - for sandboxes with no internet access, nothing else.chapter →	Safer pairing: `--allow-dangerously-skip-permissions` enables bypass as an option without defaulting to it.
`/permissions`	Manage allow/ask/deny rules by scope in an interactive dialog.
`/fewer-permission-prompts`	Scan transcripts for common read-only calls and add a safe allowlist to project settings.
`/sandbox`	Toggle OS-level sandbox mode (supported platforms only).chapter →
`Shift+Tab` docs	Cycle permission modes in-session: default → acceptEdits → plan (plus auto/bypass if enabled).chapter →	Key reference: the keyboard shortcuts chapter.
`/plan <description>`	Enter plan mode - read/search work, edits blocked, agent produces a plan.chapter →

Watch the window

Context is the resource; these are the gauges and the release valves.

Command	What it does	Notes
`/context [all]`	Visualize what occupies the window right now: system prompt, rules, tools, MCP schemas, conversation.chapter →
`/usage`	Session cost, plan usage limits, and activity stats - including spend by skill, subagent, plugin, and MCP server.	Aliases: `/cost`, `/stats`.
`/compact <instructions>` `/compact Keep the API changes and the failing test output`	Summarize the conversation to free context, optionally steering what survives.chapter →
`/clear <name>`	Start a fresh conversation; the previous session stays resumable.	Rename first (`/rename`) if you will want to find it later.
`/autocompact <auto\|tokens>`	Tune auto-compaction: auto, or a window size from 100k to 1M tokens.
`claude --autocompact <auto\|tokens>`	Set the auto-compact window size at startup.
`/memory`	Edit CLAUDE.md files and manage auto memory.chapter →
`/init`	Initialize the project with a CLAUDE.md guide; interactive flow also walks through skills and hooks.chapter →

Everyday slash commands

The in-session set most people actually use. Roughly a hundred exist; these earn their keystrokes.

Command	What it does	Notes
`/help`	Show help and available commands.
`/exit`	Exit. In an attached background session, detaches and the session keeps running.	Alias: `/quit`
`/status`	Version, model, account, connectivity.
`/config <key=value>`	Open the settings panel, or set values directly.chapter →
`/diff`	Interactive diff viewer: uncommitted changes and per-turn diffs.
`/export <filename>`	Export the current conversation.
`/tasks`	View and manage background work in the session, including finished subagents.	Alias: `/bashes`
`/mcp <reconnect\|enable\|disable>`	Manage MCP servers in-session.chapter →
`/skills`	List skills; filter by name, sort by token cost, cycle visibility.chapter →
`/hooks`	View hook configurations for tool events.chapter →
`/agents`	Create or manage subagents - v2.1.198+ asks Claude to do it, or edit `.claude/agents/` directly.chapter →
`/review <level> [--fix]` `/review medium --fix 1234`	Multi-agent review of the current diff, a PR number, branch, or path.
`/security-review`	Security review of your branch against origin default (needs an `origin` remote).
`/simplify <target>`	Bundled skill: four parallel agents review changed code for cleanup, then apply fixes.
`/debug <description>`	Bundled debugging skill.
`/deep-research <question>`	Bundled skill: multi-search research pass with citations.
`/loop <interval> <prompt>`	Run a prompt on a schedule in-session.
`/batch <instruction>`	Bundled skill: run one instruction across many files.
`/focus`	Toggle focus view: last prompt, tool-call summary, final response.
`/terminal-setup`	Configure the terminal for Shift+Enter (only shown in terminals that need it).

Script it (print mode)

Pipes, CI, and the SDK surface.

Command	What it does	Notes
`claude -p "<prompt>"`	Print mode: run the agent loop to a final answer, print, exit - built for pipelines.chapter →	Long form: `claude --print`. Skips the workspace trust dialog - only use in trusted directories.
`cat error.log \| claude -p "find the root cause"`	Process piped content as input.
`claude -p "<query>" --output-format <format>`	Structured output: text (default), json (single result), or stream-json (realtime).
`claude -p --input-format stream-json`	Streaming input for programmatic two-way sessions.
`claude -p --max-turns 3 "<query>"` docs	Limit agentic turns in print mode; exits with an error at the limit.	Real but hidden from `claude --help` in 2.1.x - verified in the official CLI reference.
`claude -p --json-schema <schema> "<query>"`	Validate structured output against a JSON Schema.
`claude -p --include-partial-messages "<query>"`	Emit partial message chunks as they arrive (with stream-json output).	Companions: `--include-hook-events`, `--forward-subagent-text`.
`claude -p --no-session-persistence "<query>"`	Do not save the session to disk (print mode only).
`claude --append-system-prompt "<text>"`	Append to the default system prompt; or replace it with `--system-prompt`.
`claude ultrareview <target>`	Cloud-hosted multi-agent review of the current branch or a PR.
`claude-agent-sdk (npm / PyPI)` docs	Embeddable SDK for programmatic control; the GitHub Action is the supported CI path.chapter →	npm package was renamed from `@anthropic-ai/claude-code` in late 2025.

MCP servers

Mounting tools and data sources.

Command	What it does	Notes
`claude mcp add --transport http <name> <url>` `claude mcp add --transport http sentry https://mcp.sentry.dev/mcp`	Add an HTTP MCP server; `--header` for auth headers.chapter →
`claude mcp add <name> -- <command>` `claude mcp add my-server -e API_KEY=xxx -- npx my-mcp-server`	Add a stdio server; `-e KEY=val` passes environment variables.
`claude mcp add-json <name> <json>`	Add a server from a JSON config string (stdio or SSE).
`claude mcp list`	List configured servers with health checks (pending ones shown as unapproved).
`claude mcp remove <name>`	Remove a server.
`claude mcp login <name>`	Authenticate with an MCP server (OAuth); `logout` clears credentials.
`claude --mcp-config <file.json>`	Load MCP servers from JSON files for this run; `--strict-mcp-config` ignores all other config.
`claude mcp serve`	Run Claude Code itself as an MCP server.

Configuration files

Where settings live and how they stack.

Command	What it does	Notes
`.claude/settings.json` docs	Project-shared settings (checked in); `.claude/settings.local.json` is the gitignored local override.chapter →	User scope: `~/.claude/settings.json`. Orgs can pin a managed policy above everything.
`claude --settings <file\|json>`	Load additional settings from a file or inline JSON for this run.
`claude --setting-sources <list>`	Limit which setting sources load: user, project, local.
`CLAUDE.md` docs	Standing instructions, loaded every session - the rules file.chapter →	There are no named profiles; a scope's settings.json is its only config at a time.
`claude --verbose` `claude -d "api,hooks"`	Verbose logging for the session; `-d` enables debug mode with category filters.

Skills, agents & plugins

Teaching the tool new tricks.

Command	What it does	Notes
`.claude/skills/<name>/SKILL.md` docs	A skill becomes a slash command: type `/<name>` to invoke.chapter →	Legacy `.claude/commands/<name>.md` still works; skills are the recommended path.
`$ARGUMENTS` docs	Placeholder in skill content, replaced with whatever follows the command.chapter →
`disable-model-invocation: true` docs	Skill frontmatter that hides the skill from Claude - only explicit `/<name>` works. For side-effecting skills.
`claude plugin`	Manage plugins; plugin slash commands are namespaced like `/my-plugin:review`.chapter →
`claude --plugin-dir <path>`	Load a plugin from a directory or .zip for this session only; `--plugin-url` fetches from a URL.
`claude --agents <json>`	Define custom agents inline for the session; `--agent <name>` starts on a specific one.chapter →
`/reload-skills`	Re-scan skill directories so on-disk changes are available without restarting.	Companion: `/reload-plugins --force`.
`/list-agents`	List subagents and sessions Claude can message, with the name to use for each.