Running Claude Code in a workspace instead of one terminal

The plain version

You can run Claude Code in any terminal, in any directory, today. Nothing on this page is a reason you cannot. What changes when more than one is running is bookkeeping: which session is in which checkout, which diff is ready, which one is stuck on a test that will never pass.

Nix Agents keeps that bookkeeping on screen. Each Claude Code session gets its own git worktree, its own real terminal pane and its own line in the diff view, so "what is running and where" is a glance rather than `git worktree list` in three windows.

What the workspace adds

• One worktree per session, created from the app — no branch juggling before you start.
• Real terminals, including Windows PTY support, so the CLI behaves the same as it does in your shell.
• A diff view per worktree, so a session can be reviewed before it is merged rather than after.
• Loop mode — a task re-run until it passes, with an exit condition you set instead of "try again".
• Swarm mode — several agents across several worktrees with one merge queue at the end.

Where it stays out of the way

Nix Agents does not proxy your model calls and does not ask for an API key. Claude Code authenticates with Anthropic exactly as it does now; the app runs the binary you installed. Your provider relationship, your billing, your settings.

It also does not hide the terminal. Every agent pane is a terminal you can type into, which means the escape hatch is always one click away — hand the session a command yourself, or finish the job by hand.