A practical guide to running AI coding agents in tmux, and where it falls short

A guide from crystl, which makes a macOS terminal for AI agents, explains how to run Claude Code and other agents inside tmux. Key tips: recent Claude Code versions can create a git worktree plus a tmux session with `claude --worktree --tmux` (on iTerm2, add `--tmux=classic` if you want real tmux underneath). Keep one named session per project. Send a prompt to an agent with `send-keys -l`, wait a moment, then send Enter. Read its output back with `capture-pane -J`. The guide also lists tmux's limits for agents: it cannot tell you which agent is waiting for approval, has no signal for "done", and loses everything on reboot. The last section is a pitch for crystl. On Hacker News (168 points, 43 top-level comments) there is no consensus on tools. Many commenters said they had moved to other tools, with herdr named most often, or had built their own multiplexer; one remarked that nearly every top comment was about building their own. Two called the article an ad for crystl. Others said tmux is still worth learning, especially on remote servers, while a few said built-in options such as Claude Code's remote control or resume features now cover the need. Sorted take: the tmux commands are solid and free, so take those and skip the product pitch. If you only run one agent at a time, the built-in Claude Code flag is enough.