Your first crtr session

You can start an agent in a terminal, talk with it in a tmux pane, then close that pane without stopping the agent. Use a root node for a conversation that you will drive yourself.

Install and start

Install crouter globally, then run crtr from the directory where you want the agent to work.

npm install -g @north-light/cli
crtr

The first time you run crtr on a machine, it opens the setup installer. The installer signs you in to a model provider, then offers companion packages (including the official plugin marketplace) and system dependencies such as tmux. On a Mac, a last step explains how to make Option act as Alt in your terminal (crouter's menu is Alt+C), and can write the setting for you in Ghostty or VS Code after you confirm; until it is set, /menu opens the same menu. Next it asks which profile this directory belongs to, then opens a conversation with an agent in the same terminal. That first agent offers a short tour; say yes to learn the viewer, the inbox and document review, or just tell it what you're working on.

The conversation viewer is a tmux pane. If you ran crtr outside tmux, it starts tmux for you: after setup, the profile choice and sign-in, it opens a tmux session named crouter (or adds a window to it, if it already exists) and runs your same command there. Your other tmux sessions are left alone. If tmux is not installed, crtr tells you how to install it. The first conversation in a session that crtr started shows how to leave and come back: Ctrl+B d leaves tmux while the agent keeps running, and Ctrl+B n / Ctrl+B p switch windows. Run without a terminal (piped or in a script), crtr never starts tmux.

To reopen the installer later, for example to sign in to another provider or add a package, run crtr sys setup.

Start a conversation

Every later crtr opens a new conversation in the current directory. You can pass a directory and a first message:

crtr                                   # a conversation here
crtr ~/code/app "Find the failing tests."  # a conversation in ~/code/app with a first message
crtr -c                                # return to the most recent conversation here
crtr -r                                # choose a conversation here to return to

Type in the editor at the bottom of the viewer. Enter sends the message. Alt+Enter (Option+Enter on a Mac) or Shift+Enter starts a new line.

To start an agent from a script or another agent without taking over the terminal, create a root node instead:

crtr node new --root "Inspect this repository and report the failing tests."

--root creates an independent node rather than a child of another agent. Inside tmux it opens a viewer pane for that node; outside tmux it prints the node id, and you open the viewer with:

crtr surface attach to <node-id>

The viewer is not the agent process. The broker keeps the node's conversation and session running separately from the pane, so closing the terminal or leaving the viewer does not discard the work. A root node's final report is not sent back to the terminal that created it; open the node again when you want to see its result.

Return to a node

When you leave a conversation (Ctrl+D), crtr prints that the agent is still running and the exact command to reopen it. crtr -c and crtr -r reopen conversations started in the current directory. To reopen a specific conversation by id, from any shell (outside tmux it starts tmux first, as above):

crtr -r <node-id>

crtr -r resolves the id first and uses that conversation's own profile and directory, so it works from any directory. An id prefix works too (including the 8-character form pickers show); when it matches several running conversations, crtr lists their full ids. crtr -r with no id lists this directory's conversations to pick from. Inside a conversation, Alt+C → Find node (or /resume) searches every conversation. From inside tmux you can also use:

crtr surface node focus <node-id>

For a running node, this brings its viewer into view. For a parked resident node, it opens the saved conversation; sending a message wakes it. For finished work, it opens a read-only saved transcript instead of starting another agent process.

For the controls inside that viewer, continue with The viewer. For a view of every node, use Canvas and dashboard.