Skip to content
arcturn

CLI reference

Every command-line flag and every slash command in one place, with what each is actually for.

On this page

Everything Arcturn accepts, in one place. The pages linked from each entry are where the reasoning lives; this is the index you come back to when you know the feature exists and want the exact spelling.

Invocation

arcturn                              # interactive TUI in the current directory
arcturn "fix the flaky retry test"   # TUI, seeded with a first prompt
arcturn -p "summarise src/auth.ts"   # run once, print the answer, exit
cat notes.md | arcturn -p            # piped stdin becomes the prompt
cat ctx.txt | arcturn -p "explain"   # piped stdin becomes leading context

With -p and a prompt argument, Arcturn waits briefly for piped input and then proceeds without it. That matters when a parent process — CI, a Makefile, subprocess.run — leaves an inherited stdin open that never closes; waiting for EOF there would hang forever.

Flags

Flag What it does
-p, --print Non-interactive: run to completion, print the final message, exit
--output-format <fmt> With --print: text (default) or json (NDJSON of every agent event)
-m, --model <id> Model to use — see --list-models and Model providers
-c, --continue Resume the most recent session in this directory
-r, --resume <id> Resume a specific session — see Sessions
--permission-mode <m> default, acceptEdits, plan or yolo — see Permissions
--cwd <dir> Working directory for tools, config and sessions
--no-mcp Do not start any configured MCP servers — see MCP
--max-turns <n> Stop a run after n model turns
--max-cost <usd> Abort the run once it has cost this much — see Cost & audit
--dry-run Send file edits to a shadow copy; review with /diff — see Dry run
--trace Write one JSON line per finished telemetry span to stderr — see Telemetry
--list-models Print the model catalog and exit
--list-providers Print every provider and preset endpoint, and exit
-h, --help · -v, --version Usage, version

--host, --port and --token apply to arcturn serve; --cassette applies to arcturn bisect.

--max-cost and --max-turns are the two worth reaching for by default. An agent that loops on a failure it cannot fix spends real money doing it, and a ceiling turns "I checked my bill on Friday" into "the run stopped."

Subcommands

Command What it does
arcturn auth login <provider> Sign in to a provider with an OAuth subscription
arcturn auth logout <provider> · auth status Forget credentials; show who is signed in
arcturn mcp-serve Expose Arcturn as an MCP server — see MCP server
arcturn serve WebSocket server over the NDJSON protocol — see Server mode
arcturn completions <shell> Print a bash, zsh or fish completion script
arcturn blame · arcturn replay · arcturn bisect Provenance and replay — see Provenance and Replay & bisect

Slash commands

Typed inside the interactive TUI.

Session and context

Command What it does
/help List the available commands
/clear Start a fresh session
/compact Summarise the conversation to free up context — see Context management
/sessions Resume an earlier session in this directory — see Sessions
/rewind Restore to an earlier turn; /rewind <query> jumps by intent — see Checkpoints
/export Export the conversation to markdown or HTML
/todos Show the current todo list
/exit Quit Arcturn

/rewind restores files and forks the conversation, rather than truncating it. The session is a tree, so the branch you rewound away from is still there.

Models, cost and appearance

Command What it does
/model Switch the model; /model refresh re-queries each provider's own list
/cost Show usage and cost; also limit <usd>, preview [steps]
/theme Switch the colour theme
/permissions Show rules and mode; also suggest
/mcp Show MCP server status

Changes

Command What it does
/diff Show pending dry-run changes
/apply Apply pending dry-run changes to the workspace
/discard Throw away pending dry-run changes

These three are the dry-run loop: with --dry-run, edits land in a shadow copy until you have read them.

Delegation

Command What it does
/bg <task> Run a task in the background; /bg, logs|cancel|adopt <id> — see Teams
/team <goal> Run a team of agents on one goal; status|cancel|merge|discard [id]
/scout <a> | <b> Explore approaches in parallel worktrees — see Scouts
/workflow <name> [args] Run a scripted multi-step workflow — see Workflows
/workflow list List discovered workflows
/workflow status [runId] What a run reached, its spend and turns, and why it stopped
/workflow resume <runId> [answer] Re-enter an interrupted run, or answer an ORG-ASK:
/org memory Inspect per-role lessons injected into later runs — see Agent organizations
/org memory add|propose|approve|revoke|rm Edit that store

The distinction between /bg, /team, /scout and /workflow is what each is for: /bg is one task off-thread, /team is several agents on one goal with a merge step, /scout is competing approaches you pick between, and /workflow is a scripted pipeline whose shape you decided in advance. Teams compares the first three; Workflows covers the last.

/org memory entries are inert until a person approves them — a proposed entry is never injected into a role's prompt. That gate exists because a memory entry becomes standing instruction text in later runs, which is not something a model should be able to grant itself.

Edit this page on GitHub