devpit

Each project has a single terminal that receives attention. Switch to another project and the previous session keeps running in the background, out of sight but not stopped. You stop juggling tabs and start watching one thing at a time.

Terminals are tmux panes, or psmux panes on Windows. Close devpit, reopen it later, and the processes are still there with their scrollback. The same applies during an update, so a running agent does not die because the app restarted.

Every project gets a Kanban board with columns you name, reorder and extend. A column either does nothing or holds one step of three kinds: an agent call, a session that takes over the target terminal, or a shell command. Moving a card resolves the step into one invocation with a fixed agent, model, context slice and expected answer shape.

Each step carries its own ceiling in dollars, and the card shows the real cost when the step finishes. That combination makes an agent workflow auditable: you know what was allowed before it ran and what it actually cost afterwards.

A file tree and diff view sit next to the terminal, so reviewing an agent's change does not mean switching apps. A browser pane opens the project's localhost server, keeps sessions separate per pane, and can pull a signed-in session from Chrome, Firefox or Safari for a specific domain.

Agent definitions are markdown files with frontmatter in ~/.devpit/agents/, plus whatever your tooling already installed. devpit references them instead of copying them, so the catalogue stays where it belongs. Nothing needs recompiling or registering.

The orchestrator is a chat you drive. It reads every board you link, lists the sessions running on your Claude Code account, and can start a new session on a card in that card's own worktree. It never moves a card into a lane that runs a step, never acts on a timer, and cannot approve anything in another session.