Claude Code in tmux, and a tmux alternative for AI agents.

Give each agent its own named session. A name you can type beats a number you have to look up, and one session per task keeps a stuck run from hiding behind another.

tmux new -s api          # one session per agent or task
cd ~/code/api && claude  # or: codex
# press Ctrl-b then d to detach; the agent keeps running

tmux ls                  # list sessions
tmux attach -t api       # pick it back up

On a remote box, let SSH open the session for you. -A attaches if the session already exists and creates it if not, so the same command works the first time and every time after:

ssh -t devbox tmux new -A -s api

tmux keeps the process alive, not the machine. If the box sleeps, reboots or the tmux server is killed, the run stops. The conversation is still on disk, so claude --continue or codex resume --last in the same folder picks it up where it ended. The Claude Code resume guide and the Codex resume guide cover the picker and session IDs.

Out of the box, Claude Code inside tmux has two problems: Shift+Enter submits the prompt instead of adding a new line, and desktop notifications never reach your real terminal. The first three lines below come from Claude Code's own terminal docs and fix both. The rest are tmux options that help when an agent prints thousands of lines and you are not watching.

# Shift+Enter and notifications for Claude Code inside tmux
set -g allow-passthrough on
set -s extended-keys on
set -as terminal-features 'xterm*:extkeys'

# Agents print a lot; keep more scrollback (new windows only)
set -g history-limit 50000

# Flag a window in the status line when its agent rings the bell
setw -g monitor-bell on
set -g bell-action any

# Scroll and pick panes with the mouse
set -g mouse on

Run tmux source-file ~/.tmux.conf to apply it to a running server. history-limit only affects windows created after the change. If Claude Code's display flickers in a small pane, its fullscreen renderer (/tui fullscreen) scrolls inside Claude Code instead of relying on tmux scrollback. If you only need a new line now, Ctrl+J works in any terminal without setup.

An agent waiting on a permission prompt in a detached session does nothing until you look. Make it ring:

• Claude Code sends a desktop notification on its own only in Ghostty, Kitty and iTerm2. Anywhere else, set "preferredNotifChannel": "terminal_bell" in ~/.claude/settings.json. For a sound or your own script, add a Notification hook; it runs when Claude finishes or pauses for a permission prompt while you seem to be away.
• Codex emits terminal notifications through tui.notification_method (auto, osc9 or bel), and notify = ["python3", "/path/to/notify.py"] in ~/.codex/config.toml runs a program on every finished turn.

With monitor-bell on, the bell marks the window in tmux's status bar, so you can see which of five sessions wants you. The catch: a bell only reaches a terminal that is attached. If you closed the laptop, nothing reaches you. The Claude Code notifications guide has hook commands for macOS, Linux and Windows.

A worktree in its own tmux session. claude --worktree feature-x --tmux creates a git worktree for the session and opens it in a new tmux session. --tmux requires --worktree. In iTerm2 it uses native panes; --tmux=classic forces plain tmux.

Agent teams in split panes. Agent teams let one Claude Code session lead several teammates, each a separate Claude instance. They are experimental and off by default: set CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 in your environment or in the env block of settings.json. The teammateMode setting picks the display:

• "in-process" (the default): every teammate lives in your one terminal; arrow keys select one.
• "auto": a split pane per teammate when you are already inside tmux (or iTerm2 with the it2 CLI), in-process otherwise.
• "tmux": always split panes, tmux or iTerm2 detected from your terminal.

For one session, claude --teammate-mode auto does the same; the flag is experimental and does not appear in claude --help. Split panes do not work in the VS Code integrated terminal, Windows Terminal or Ghostty. Teammate permission prompts show up in the lead's pane, so you still answer them in one place. If a team's tmux session outlives Claude, tmux ls and tmux kill-session -t <name> clean it up. Each teammate is a full Claude session, so a team spends noticeably more tokens than one agent.

tmux is a multiplexer: it runs inside whatever terminal you have, on a server or over SSH, and keeps processes alive when you detach. cmux is a native Mac terminal app made for coding agents, with a sidebar of workspaces and a ring around any pane whose agent needs you. They solve different problems and many people run both. The full comparison, with cmux's platforms and pricing, is on Termdeck vs cmux.

Termdeck is a web UI for Claude Code, Codex and Grok. A small agent on each machine (macOS, Linux or Windows) dials out to termdeck.io, and every session on every machine shows up in one sidebar in your browser. Runs happen on the machine, so closing the tab does not stop them; transcripts stay in the engines' own files on your disk.

Its Terminals page is the part a tmux user will recognise: chats tiled edge to edge with a status bar along the bottom, one pane per chat, from any machine, with a tmux classic theme if you want it. You split a pane right or down, drag a chat from the sidebar onto an edge, and open more windows. Approval prompts appear in the pane as numbered buttons, and a window with one waiting gets a flashing !. Away from the desk, Web Push reaches your phone and the notification opens the chat. Edits render as diffs you can review and turn into a pull request. Terminals needs a wide screen; on a phone you use the normal chat view.

It works alongside tmux. A Claude Code session you started in a tmux pane appears in Termdeck too, view-only while the terminal holds it, so two writers never fork the transcript. Take over closes an idle terminal client, and the terminal can resume the session later. See the interface docs for Terminals and sessions for how the lock works.

If your agents run on one machine you are usually sitting at, tmux plus the config above is enough, and it costs nothing. The same goes if you live in SSH, run CLIs other than Claude Code, Codex and Grok, or do not want any hosted service in the loop: Termdeck needs an account on termdeck.io, even though your transcripts stay on your machine. Termdeck earns its place when runs outlive your attention: agents on several machines, approvals you want to answer from your phone, or more sessions than one status bar can show.