Running an agent

This guide assumes the Quickstart: CLI installed, signed in, daemon running.

Start from the terminal

Run the agent’s subcommand in the directory you want it to work in:

cd ~/code/my-repo
consortium claude

The agent starts exactly as it would natively — same auth, same configuration — and the session registers with your account. It shows up in the app within a second or two.

Start from the app

On any machine whose daemon is online:

1. Open the machine (or New session).
2. Pick the directory, the agent, and optionally a model or provider account.
3. Send the first prompt.

Sessions started this way run under the daemon on that machine — the terminal-less equivalent of the flow above. You can also spawn into a git worktree to keep parallel runs from stepping on each other.

Hand control back and forth

• Terminal → app: just send the session a message from the app. The local process hands over and keeps running in remote mode under the daemon.
• App → terminal: press any key in the original terminal (double-tap space for Consortium Code sessions). You’re back at the native prompt with full context.

Permission requests (file writes, shell commands) surface in the app as push-style approvals while the session is remote.

Stop and resume

• Stop: Ctrl-C in an attached terminal, or Stop in the app.
• Resume: reopen the session from the app or start the agent again in the same project — sessions are deduplicated per directory tag, so you continue rather than fork.
• A machine that goes offline marks its sessions inactive; they pick up when the daemon reconnects.

Working patterns

• One directory, many sessions — parallel investigations stay isolated; use worktrees when they must write.
• Small models for long tool loops — see Supported models for the cheap variants.
• Caps — set a spending cap (Pricing) if a managed session might run unattended.