Tmux Keeps AI Coding Agents Running for Days After You Disconnect

Tmux, the terminal multiplexer Nicholas Marriott first published online in late 2007, shipped version 3.6b on May 20, 2026, according to its GitHub release page. The nearly 19-year-old open-source program has found a second use among developers who run AI coding agents. Tools such as Anthropic's Claude Code and OpenAI's Codex CLI now run for hours, sometimes days, inside tmux sessions on a rented server, where the work continues after the laptop that started it disconnects.

Marriott, a software developer from Northern Ireland, told the OpenBSD Journal in July 2009 he had been "vaguely unhappy" with GNU Screen, which he described as carrying poor documentation, a strange configuration file, and an unintuitive command-line interface. He wanted source he could read. The first version went online in late 2007. OpenBSD developer Paul Irofti raised it at a hackathon. The system folded tmux into its base distribution with the 4.6 release of 2009. Two years later, Marriott told LinuxTag he spent about half an hour to an hour a day on it. User requests drove most of the features he added, among them session-selection menus and better copy support.

The repository now carries about 46,000 stars on GitHub and remains under Marriott's maintenance, with 3.6b the latest of more than a dozen public releases. Its use among people running agents rests on one capability: a tmux session keeps running on the server after the SSH connection that launched it drops, and it can be reattached later from a different machine.

Named developers have written about the workflow. Peter Steinberger, the creator of the open-source agent framework OpenClaw, described in an October 2025 post how he gave up on a built-in background-task feature and turned back to the older utility: "I now use tmux," he wrote, a way "to run CLIs in persistent sessions in the background." Simon Willison, writing the same month, said he keeps several terminal windows open running Claude Code and Codex CLI "in YOLO mode (no approvals)," some of the work launched from his phone. Steve Yegge's Gas Town, a Go-based orchestration system, pushes the pattern to 20 or 30 Claude Code agents coordinated across tmux panes. Boris Cherny, who built Claude Code at Anthropic, told the Latent Space podcast that the tool leaves parallel-session management to other software: "a lot of people use [Claude] Code with ... Tmux ... to manage a bunch of windows and a bunch of sessions happening in parallel," he said. David Ondrej, in a May 25 video titled "Build Anything with Tmux," started a session on a low-cost virtual private server, split it into panes for several agents, then detached and reconnected from a phone while the work kept running. Implicator has covered the same persistence trick for coding from an iPhone and in a roundup of CLI tools for AI-assisted work.

That workflow has drawn newer tools built for it. cmux, released as open source in February 2026, hit the number-two spot on Hacker News and drew thousands of GitHub stars quickly, according to the developer site SoloTerm. It runs only on macOS. There it wraps panes in notification rings and adds an embedded browser for developers watching several agents at once. Termdock takes a different form, an Electron desktop application that replaces the terminal with a graphical window and ships with built-in connections to OpenAI, Anthropic, Google, and xAI. Zellij, written in Rust and released in 2021, competes by showing users which keys perform which action. None of the three targets tmux's main job. Termdock and cmux run locally on a Mac, and while Zellij works over SSH, the SoloTerm comparison still recommends tmux on the remote server and cmux for local agent work.

The setup carries a specific risk. Running agents unattended means turning off their permission prompts, a mode both Claude Code and Codex expose, which lets an agent edit files and run shell commands for hours with no human reviewing each step. Even developers who rely on the mode hedge it: Willison said he limits his no-approval runs to "tasks where I'm confident malicious instructions can't sneak into the context," and wants to move his agents into Docker containers "to further limit the blast radius." The program also has limits. It runs on Linux, macOS, BSD, and Solaris, but there is no native Windows build, so Windows users run it inside the Windows Subsystem for Linux. Newcomers face the prefix-key system, the C-b chord that has to come before every command. Red Hat has recommended tmux over GNU Screen since RHEL 8.

On remote servers reached over SSH, tmux remains the default, and the Mac applications built around agents have not displaced it there. Those tools add local monitoring tmux was never designed to provide. The tool itself runs as a background process that survives the disconnect and adds little memory load on a small server. Version 3.6b shipped on May 20, 2026. The program has been part of the OpenBSD base system since 2009.

How it works

• Runs inside your terminal: tmux uses a client-server model, so its sessions live in a background process rather than in any single window.
• Sessions, windows, and panes: a session is the whole workspace, windows act like tabs, and panes are splits inside one window.
• Client-server architecture: the tmux server holds sessions in the background; closing the terminal or losing an SSH link only detaches the client and leaves the processes running.
• The prefix key: every command starts with a prefix, C-b by default, followed by a key. C-b % splits a pane left and right, C-b " splits it top and bottom, and C-b d detaches the session.
• Attach and reattach: tmux new -s NAME starts a named session, tmux attach -t NAME reconnects to it, and tmux ls lists what is running.
• Wide reach: it installs through apt, dnf, pacman, port, or brew and runs on Linux, macOS, BSD, and Solaris.

How to use it well

1. Name every session (tmux new -s agents) so tmux ls and your reattach commands stay readable.
2. Turn on mouse mode (set -g mouse on in ~/.tmux.conf) so you can scroll and click between panes.
3. Log in straight to a live session with ssh server -t 'tmux attach || tmux new'.
4. Add tmux-resurrect and tmux-continuum through the Tmux Plugin Manager to restore layouts after a reboot, which a plain session does not survive.
5. Script repeatable layouts with tmuxinator or a shell script instead of splitting panes by hand each time.

Two mistakes recur:

• Starting a long agent job outside tmux, so it dies the moment the connection drops.
• Leaving agents running unattended with permission checks disabled, which removes the last guardrail on what they can change.

# Start or reattach to a persistent session over SSH
ssh server -t 'tmux attach || tmux new -s agents'

# Inside tmux:
#   C-b %   split panes left/right
#   C-b "   split panes top/bottom
#   C-b d   detach (the agents keep running)
# Later, from any machine:
tmux attach -t agents

How it compares

What it costs

• tmux is free and open source under the ISC license, with no paid tier, no account, and no telemetry.
• Its main alternatives are also free to use, so the real cost is the time spent learning each one.

The verdict

For a developer whose agents run on a rented server, tmux is still the tool the newer options measure themselves against, because a session started inside it keeps long-running work alive after the connection drops or the laptop closes. The Mac applications are a real gain for anyone monitoring several agents on one screen, and cmux or Termdock may suit that local job better. In practice the tools split by location: the desktop apps watch agents on one machine, and tmux runs the server those agents sit on.

Pros:

• Sessions survive SSH drops and reconnect from any machine
• Runs on nearly every Unix system
• Near-zero memory overhead on small servers
• Fully scriptable, with a large plugin ecosystem

Cons:

• The prefix-key system takes time to learn
• No native Windows build
• No agent-specific notifications or visual monitoring
• Sessions do not survive a server reboot without extra plugins

Professional Tutorial: Build a Pi-to-Pi Agent Communication Bus

The useful version of multi-agent coding does not start with ten agents. It starts with two Pi Coding Agent sessions, using the terminal coding harness to hold different context: one session sits near

The Implicator

Repo Radar: 5 GitHub Projects Worth Your Week

Repo Radar's fourth issue leaves the coding agents alone and looks at the tooling around them. The five projects below were all pushed within the last 48 hours. Memory between sessions, code indexing,

The Implicator

15 CLI Tools That Make AI-Assisted Coding in the Terminal Actually Work

The terminal is no longer a fallback interface for developers who refuse to move on. It is the main stage. AI coding agents run there, generate files there, commit code there. The graphical IDE still

The Implicator

Marcus Schuler

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: [email protected]