The Complete Claude Code Command Reference

A Claude Code command is one of three different things depending on where you type it, and no single source lists all three. This reference covers the shell subcommands, the CLI flags, and the in-session slash commands — and reconciles them against what the binary on this machine actually accepts, which turns out not to match its own help output. For what the features behind these commands actually do, the complete Claude Code guide is the place to start. Every figure below was measured against Claude Code v2.1.224 on 2026-08-07 with an internal check script, and the slash-command census against 2.1.259 on 2026-09-03.

Key takeaways

• claude --help is not a complete reference. On v2.1.224 it advertises 65 flags and 13 subcommands; we found 16 documented entries it never mentions but the binary accepts.
• Seven working subcommands are invisible to --help: attach, daemon, logs, respawn, rm, stop, and remote-control.
• One documented flag does not exist in this build. --exec returns unknown option '--exec'.
• The three surfaces are distinct: subcommands manage installation and sessions, flags configure a session at launch, slash commands act inside one.
• /help is the only authority for your machine, because the slash command set varies by plan, platform, and installed plugins. A probe of every name is the next best thing — see which slash commands work without a terminal.

What counts as a claude code command

Three surfaces, three different lifetimes.

The distinction matters because the three are documented in different places and behave differently. A flag applies for the life of one session. A subcommand usually does not start a session at all. A slash command only exists once you are inside one.

For authoring your own slash commands, skills and slash commands covers the merge that made them the same feature.

Shell subcommands

These run from your shell. Most do not start a session.

Three are worth calling out. claude doctor reads settings files in the current directory without a trust prompt, which makes it the right first move on a machine you have just set up. claude project purge deletes transcripts, tasks, file history, and the config entry for a project; it takes --dry-run, and you should use it.

claude mcp has its own subcommands — list, get, add, add-json, login, remove — and get is the one to reach for first, because it prints the scope a server was defined in as well as its status. When a server is configured and not working, that status line is usually the whole diagnosis: troubleshooting an MCP server that will not connect walks the states it can report and what each one actually means.

Seven commands the help output omits

Anthropic's CLI reference states plainly that claude --help does not list every flag and that absence from help does not mean a flag is unavailable. That is easy to read past. It is also measurable, so we measured it.

The probe exploits how the argument parser reports errors. Invoking a value-taking flag with no value produces option '--x <v>' argument missing when the binary knows the flag, and unknown option '--x' when it does not. Neither starts a session, so the probe costs nothing and touches no API:

claude --max-turns
# → error: option '--max-turns <turns>' argument missing
claude --exec
# → error: unknown option '--exec'

Running that across 40 documented flags and 20 documented subcommands with our internal check script produced this:

npm run check:cli
# → claude 2.1.224 (Claude Code)
# → --help advertises 65 flags and 13 commands.
# →
# → Documented and accepted, but absent from --help (16)
# →   --advisor
# →   --append-subagent-system-prompt
# →   --append-system-prompt-file
# →   --channels
# →   --max-turns
# →   --permission-prompt-tool
# →   --ref
# →   --system-prompt-file
# →   --teammate-mode
# →   attach
# →   daemon
# →   logs
# →   remote-control
# →   respawn
# →   rm
# →   stop
# →
# → Documented but rejected by this build (1)
# →   --exec
# →
# → 16/60 documented entries are invisible to --help (27%).

The seven hidden subcommands are a coherent set — they are the background-session controls:

All seven print real usage when asked. claude respawn --help returns Usage: claude respawn <id>|--all, which is a flag combination the docs do not show either.

--exec is the one that goes the other way. The CLI reference documents it as running a shell command as a PTY-backed background job, paired with --bg. On v2.1.224 it is rejected outright. Either it landed after this build or it was withdrawn; the docs give no version gate, so treat documented-but-unlisted and documented-but-absent as two different risks.

Flags grouped by what they change

The full set is 65 in --help alone. These are the ones that change behaviour you will notice.

Starting and resuming. -c continues the most recent conversation in this directory; -r resumes a specific one by ID or name. --fork-session gives a resumed session a new ID instead of writing back to the original. -n names a session so /resume shows something readable.

Permissions and tools.

Model and effort. --model takes an alias (sonnet, opus, haiku, fable) or a full ID. --effort takes low through max. --fallback-model accepts a comma-separated list tried in order when the primary is overloaded, and only works with --print.

System prompt. --system-prompt replaces the default entirely; --append-system-prompt adds to it. Both have -file variants that are absent from --help. This is the escape hatch when a rule must reach the model at system-prompt level rather than through a Claude Code instructions file, which is delivered as context and carries no compliance guarantee.

Print mode. -p prints and exits. A cluster of flags only works with it: --output-format (text, json, stream-json), --input-format, --json-schema for validated structured output, --max-turns, --max-budget-usd, and --forward-subagent-text.

Troubleshooting. Two flags disable configuration, and the difference matters.

Reach for --safe-mode when a session behaves oddly and you suspect your own configuration. --bare is for scripted runs where you want to supply every input explicitly.

Slash commands inside a session

There are well over a hundred, and the set depends on your plan, platform, and plugins. /help lists what your session actually has. Everything below is from Anthropic's commands reference.

Correction, 2026-09-03. This section previously said /which <command> names the handler for a command. On Claude Code 2.1.259 /which answers Unknown command, as do /todos and /pr-comments — three names retired in the four weeks since this article was measured against v2.1.224. The evidence, and the three different replies the CLI gives, are in which slash commands work without a terminal.

Two of those slash commands are big enough to have their own measurement: /code-review and its effort levels, and /security-review and the empty report it can return.

Context and session state.

The first three are a catalogue entry here and a workflow elsewhere: what each one actually reclaims is in Claude Code context window management, with a full benchmark of token reduction and prompt cache mechanics in our Claude Code /compact command guide. /clear is the bluntest of them and the most misread — what clearing context actually removes measures the floor it leaves you on and the bytes it does not touch. /rewind reads from a store you can audit directly; what rewind can and cannot restore counts it, including the 7,081 shell calls it never covered. For resuming, continuing, and forking sessions from the command line, see Claude Code session management.

Configuration.

Delegation and parallel work.

Bundled skills. These are skills rather than built-ins, which is why they accept arguments and can be chained. /code-review, /security-review, /simplify, /test, /search, /debug, /doctor, and /web-search all run this way, and up to six can be chained in one message: /skill-a /skill-b do XYZ passes do XYZ to both.

Which slash commands work without a terminal

/help knows your plan, platform and plugins, but not which commands survive outside the terminal. So on Claude Code 2.1.259, on 2026-09-03, we ran 84 command names through the CLI one process per name, stdout not a TTY, and sorted the replies by the CLI's own wording:

claude -p "/context" --model haiku
# → ## Context Usage
# → **Model:** claude-opus-5[1m]
# → **Tokens:** 22k / 1m (2%)

claude -p "/status" --model haiku
# → /status isn't available in this environment.

claude -p "/which" --model haiku
# → Unknown command: /which

Three replies, three meanings: it ran, it exists but needs the terminal UI, there is no such command. The count: 23 returned real output, 52 need the TUI, and 9 are not registered at all.

/context is the standout: the same table the interactive command prints, no terminal needed, and the cheapest way to see from a script which MCP servers are eating your prompt. /doctor took 69 seconds and /insights 78, because both hand the turn to the model; everything else returned in under five. Two replies are neither a run nor an error — /agents answers "The /agents wizard has been removed" and /extra-usage answers "/extra-usage is now /usage-credits" — so a script that checks only the exit code will not notice its command was retired.

The 52 terminal-only names include most of what you reach for interactively: /help, /status, /skills, /hooks, /permissions, /memory, /plan, /diff, /rewind, /resume, /export, /tasks, /theme, /vim, /output-style, /statusline, /keybindings, /add-dir, /release-notes, /terminal-setup, /ide, /chrome and /plugin. Assume a command is TUI-only until a probe says otherwise.

/version   /todos   /alias   /which   /pr-comments
/daemon    /migrate-installer   /todo   /agent

/agent gets a suggestion — Unknown command: /agent. Did you mean /agents? — and /version is defined in the binary only as a terminal variant, so a command that plainly exists interactively reads as unknown here. That is the limit of the method: the result means this build answers to it in this environment, not this command exists.

The names come from two sources. A conservative parse of the 2.1.259 binary found 71 command definitions covering at least 64 distinct names — a floor, because that parse drops anything it cannot resolve cleanly. The session also loaded 45 skills, which are invoked with a slash exactly like built-ins: 11 from user settings, 21 from one installed plugin and 13 bundled with the build.

Verify the reference against your own build

Any command reference is a snapshot. Claude Code shipped 30 versions between v2.1.195 and v2.1.224, and at least three slash command names stopped resolving in the four weeks between v2.1.224 and 2.1.259. We re-derive the tables in this article with internal check scripts rather than maintaining them by hand, and the slash-command half of that probe is short enough to run yourself:

for name in context usage cost model mcp config effort autocompact; do
  printf '%-14s ' "/$name"
  claude -p "/$name" --model haiku 2>&1 | head -1
done

Grep the output for isn't available in this environment (terminal-only) and Unknown command (not registered). Leave out anything that signs you out, updates the binary, spends cloud credit, writes into the repository or opens $EDITOR — our census excluded 32 such names, including /logout, /update, /install, /ultrareview, /fork and /init, rather than counting them as failures.

The flag-and-subcommand script parses claude --help, probes every documented flag and subcommand, and reports two categories: accepted but undocumented in help, and documented but rejected. It exits non-zero on the second, because that is genuine drift between the docs and the binary.

Two bugs in it are worth repeating, because both produced confident wrong answers first. The flag parser used a lowercase-only pattern, so --allowedTools and --disallowedTools were reported as hidden when they are in --help in camelCase. The command parser required two spaces after the command name, so every row written as agents [options] was missed and the count came out at 8 instead of 13. A reference built on either would have been wrong in a way no reader could catch.

The script also runs three control probes before trusting any result: a known flag must be recognised, an invented flag must be rejected, and an invented subcommand must fall through to root help. If any control fails, it aborts rather than reporting numbers.

Common mistakes

• Treating claude --help as the reference. It omits 16 of the 60 documented entries we probed, including every background-session control.
• Assuming the docs match your binary. --exec is documented and absent on v2.1.224. Check before you script against a flag.
• Using print-mode flags interactively. --output-format, --max-turns, and --max-budget-usd are silently irrelevant without -p.
• Reaching for --dangerously-skip-permissions when --permission-mode would do. The mode flag has six values, and dontAsk or acceptEdits covers most of what people want bypass for.
• Confusing --safe-mode with --bare. Both disable customisation, but --bare also drops OAuth and keychain auth, so an interactive session started with it may not authenticate.
• Treating "Unknown command" and "isn't available" as the same failure. The first means the name is not registered; the second means it exists and needs a terminal.
• Looking for a claudecode binary. The command is claude. Neither claudecode nor the transposed cluadecode resolves to anything.

Conclusion

Use /help for slash commands, because only your session knows what your plan and plugins provide, and probe the names you intend to script against — the census above shows how, and which nine names have already stopped resolving. For the chords that drive the same surface, see Claude Code keyboard shortcuts. Use the CLI reference rather than claude --help for flags, since help omits more than a quarter of them — and verify anything you are about to script against, because at least one documented flag is missing from the current build. If you only remember three: /context when a session slows down or misbehaves, --safe-mode when your configuration does, and claude doctor before either. /context reports the session in front of you; for whether a release actually changed anything, the version-by-version join over your own transcripts is the measurement /context cannot give you. For the features these commands drive rather than their spelling, our full walkthrough of the tool is the better next read.

Frequently asked questions

Muhammad Kashif

Founder and editor of Devventa, covering AI coding assistants, Next.js and the modern AI development stack.