Herdr: tmux, But Built for AI Agents

Herdr is a free, open-source terminal multiplexer built for running AI coding agents — tmux rebuilt around the problem of watching several agents at once. It runs inside the terminal you already use, keeps every agent in its own pane, and marks each pane working, blocked, or idle, so you can see at a glance which agent finished, which is still working, and which one is stuck waiting for an answer. It is written in Rust, released under the Apache-2.0 license, and its latest stable version is v0.9.5 (14 October 2026).

It went viral after NetworkChuck's walkthrough "you need to use Herdr RIGHT NOW!!" (published 6 November 2026), in which his Claude Code launched Codex, Gemini, and Grok in separate panes and the agents talked back to each other. The video is a good demo — but the framing around it needs correcting before it hardens into accepted fact, and that is what this guide does first.

What the viral video gets wrong

NetworkChuck's walkthrough is a real demo of a real tool, and the enthusiasm is understandable. But five claims in the framing around it need correcting before you install anything.

How it works: panes, tabs, workspaces

Herdr takes tmux's model and adds one idea: the multiplexer should know what is running in each pane. The hierarchy is three levels deep. A pane is one terminal process — one agent, one shell, one command. A tab groups panes. A workspace groups tabs and represents one project, so each active project gets its own workspace and its agent state stays readable in the sidebar.

The client-server split is what makes detaching safe. The Herdr server manages sessions, workspaces, and panes in the background; the client is a thin process in your terminal window that captures input and renders output. Close the client, lose your SSH connection, or walk away from your desk — the agents keep running. Run herdr again and you reattach to the same session. The default prefix key is Ctrl+b, the same as tmux, so existing muscle memory carries over; Ctrl+b then ? opens the full keybinding reference, and Ctrl+b then Shift+n creates a new workspace. Mouse control is first-class too: click to focus, drag split borders to resize, right-click for context menus, drag-select text to copy.

The feature that matters: agent state

tmux treats every process the same — a Claude Code instance is just a text stream. Herdr identifies the agent processes running in its panes and shows their state in a sidebar: working (processing a task), idle (ready for a prompt), and blocked (stopped and waiting for input, for example at a first-run permission prompt). Sending a prompt to an agent updates its status immediately, confirming the task was received.

That sounds small until you run four or five agents at once. Without it, checking on each agent means switching into every pane in turn. With it, one glance at the sidebar tells you which one needs you — and Herdr can also emit a sound or a system notification when an agent in a background pane changes state, so a long-running task can ping you instead of you polling it. The practical rule from the comparison guides is simple: one or two agents, use whatever you already have; three or more, the blocked agent is the bottleneck, and that is exactly what Herdr is built to surface.

Agents talking to each other

The second half of the viral demo — Claude Code launching Codex, Gemini, and Grok, and hearing back — runs on Herdr's agent-facing API. The Herdr CLI is available while the TUI runs, and external processes, including agents themselves, can execute herdr commands to create panes, run processes, and read pane output. The documented operations include herdr pane split-right, herdr agent read, and herdr pane close.

A concrete example from the project's own documentation: give one agent a security audit. It uses the CLI to create two panes, runs a secrets scanner in one and a dependency checker in the other, waits for both to complete, reads their output, writes a summary, and closes the temporary panes. One high-level prompt becomes a multi-step workflow that orchestrates itself inside the terminal — no framework required. The same mechanism is how a lead agent hands a review or a test run to a helper agent and keeps working while the helper runs.

The newer layer is herdr-core, a small server on your machine, with the hc CLI. An agent can start a helper in the background — hc run --bg codex "review my diff. Done: hc send t1" — steer it while it works with hc send t3 "take another look", and the helper's report shows up in the lead agent's input when it finishes, so nothing is copied between windows by hand. The agents report back; you only see the result. This is the part the video demonstrates at 12:39, and it is real — but it is also the newest and least battle-tested part of the project, so expect its surface to keep changing.

Remote machines over SSH

Herdr keeps local work and saved SSH machines together in one window, with a combined agent list and independent reconnects. The --remote flag starts the server on a remote machine and the client locally, forwarding communication through the SSH tunnel; the server runs on the remote machine with access to remote files and compute, while the client runs locally and reads your local ~/.config/herdr/config.toml, so all your keybindings, themes, and settings apply to the remote session. Tailscale works too, and the project's homepage states the connection goes straight from your phone or computer to your machines — no account, and your terminals never pass through Herdr's servers.

The newer Herdr App extends this to your phone: see who needs you, open the real terminal, and answer by typing, pasting a screenshot, or talking. Agents keep running on the remote machine while your laptop sleeps or the app is closed. That is the same client-server idea as the TUI, stretched across machines instead of just across detach and reattach.

Setup, step by step

What you need first: macOS or Linux (Windows is a beta preview), a terminal you already use, and — for the agents themselves — the CLIs you want to run. Herdr installs none of them.

Step 1 — Install (Mac or Linux)

The project's install script:

curl -fsSL https://herdr.dev/install.sh | sh

Read the script before you pipe it to a shell. Fetch it first with curl -fsSL https://herdr.dev/install.sh, skim what it does, and only then run it. The same caution applies to every curl-to-shell installer, not just this one. Alternatives: brew install herdr on macOS, mise use -g herdr, or build from source with cargo build --release after cloning the repository.

Step 2 — Install (Windows, beta preview)

powershell -ExecutionPolicy Bypass -c "irm https://herdr.dev/install.ps1 | iex"

Windows support is a beta preview with an endpoint-protected installer. Expect rough edges and missing pieces compared with macOS and Linux, and check the project's Windows support docs before relying on it for daily work.

Step 3 — Start it where the work lives

cd ~/your-project
herdr

Herdr launches or attaches to your default background session. When a session has no workspaces, it opens one automatically — give each active project its own workspace so agent state stays readable in the sidebar.

Step 4 — Run an agent in a pane

Split a pane (mouse right-click, or the keyboard split), then start your agent CLI inside it — claude, codex, opencode, grok, or whichever you use. Herdr detects the agent process and shows it in the agents list with its live status. Repeat for each agent; the sidebar is the whole point, so keep it visible.

Step 5 — Detach and come back

Ctrl+b q        # detach — agents keep running
herdr           # reattach to the same session

Configuration lives at ~/.config/herdr/config.toml and reloads in a running session with prefix + Shift+r. Built-in themes include Catppuccin, Gruvbox, Nord, and Dracula, and the settings menu is at prefix + s. Plugins extend panes and workflows through a marketplace.

Herdr vs tmux vs Zellij vs cmux

The comparison below uses figures verified on 9–15 November 2026. Star counts move daily — Herdr alone went from about 49,500 on 9 November to 50,964 on 15 November — so treat them as a snapshot, not a ranking.

	Herdr	tmux	Zellij	cmux
Latest release	v0.9.5 (14 Oct 2026)	3.8 (9 Sep 2026)	v0.46.0 (11 Nov 2026)	v0.65.2.2 (27 Oct 2026)
GitHub stars	50,964 (15 Nov 2026)	50,307 (15 Nov 2026)	~36,000 (9 Nov 2026)	28,793 (15 Nov 2026)
License	Apache-2.0	ISC	MIT	GPL-3.0-or-later*
Platforms	Any terminal; macOS, Linux, Windows (beta)	Unix-like systems	Linux, macOS	macOS app
Knows agent state	Yes: working / blocked / idle per pane	No	No (plugins can add it)	Notification rings when an agent needs attention
Persistence	Background server; restores layout, can resume supported agent sessions after a restart	Detach / attach	Session resurrection	App sessions
Remote machines	Saved SSH machines in one window, combined agent list	Run on the server, attach over SSH	Built-in web client	cmux ssh workspaces
Agent-facing API	CLI + socket API (+ hc in herdr-core)	Mature command interface	Plugin system (WebAssembly)	Scriptable browser API, cmux claude-teams
Pick it if	You run 3+ CLI agents and want to see which one is stuck	It must already be on every server, scripts must work for years	You want one friendly multiplexer for agents and everything else	You are on a Mac and prefer a polished app to a TUI

* cmux's GitHub API reports no single license assertion for the repository; the GPL-3.0-or-later label (server component excluded) comes from the 9 November 2026 comparison. Zellij's star figure is also from that comparison — the repository's metadata endpoint did not return on 15 November 2026, though its latest release, v0.46.0 (11 November 2026), was confirmed directly.

Honest limitations

• It is four weeks past v0.9.5 and eight months old. The project was created 27 March 2026 and is shipping fast — v0.9.3 on 29 September, v0.9.4 on 13 October, v0.9.5 on 14 October 2026. Expect the CLI surface, the report of agent states, and the newer herdr-core layer to keep changing. The 494 open issues on 15 November 2026 are a normal count for a project this size, not a warning sign on their own.
• Windows is beta. It exists and installs, but it is not the peer of macOS and Linux yet.
• Agent state detection is heuristic. Herdr identifies agent processes and their state from what it can observe in the terminal. A blocked agent that is actually waiting on a slow tool call, or an agent that prints a prompt-like line while still working, can be misread. The status is a signal, not a guarantee.
• Restart does not preserve processes. Layout restores, supported agent sessions can resume, but the original processes do not survive a server or machine restart. Plan around that before leaving a long job running over a reboot.
• The graphical app is newer than the TUI. Herdr's homepage now leads with the Herdr App (macOS, Windows, Linux, and phone) with a gallery of layouts and mods, while the terminal UI from the video remains the multiplexer. If you follow the video, you want the TUI — check which download you are getting.
• It does not isolate agents' work. Herdr owns terminals, not git state. If two agents edit the same files, they will step on each other. Pair it with git worktrees when agents work on the same repository in parallel.

Security and privacy

Herdr runs locally and its homepage states the terms plainly: no account, and your terminals never pass through Herdr's servers — your phone and computers talk straight to your machines over SSH (Tailscale works too). The server listens locally, and the web app runs on 127.0.0.1 in the TUI's case. Configuration is plain files under ~/.config/herdr/ (and ~/.config/herdr-app/ for the graphical app), which an AGENTS.md in the app config explains to your agent.

Three things to keep in mind anyway. First, inspect the install script before piping it to a shell — the one-line install is convenient precisely because it runs code from the internet as you. Second, the agents are the risk, not the multiplexer: Herdr gives agents a socket and CLI they can use to spawn panes and prompt each other, so only enable agent-driven orchestration for agents you already trust with shell access, and review what they run. Third, remote machines mean SSH credentials: saved machines and their keys live on your machine, so the usual SSH hygiene — keys, not passwords, and no shared keys across untrusted boxes — applies.

Real use cases

1. Run a small team of agents on one feature. Claude Code writes the feature, Codex reviews the diff, Gemini writes the tests — each in its own pane, each visible in the sidebar. When one goes red (blocked), you answer it and the rest keep working. This is the NetworkChuck demo, and it is the workflow Herdr is built for.

2. Overnight or long-running jobs without babysitting. Start agents in panes, detach with Ctrl+b q, and let the sound or system notification tell you when one finishes or gets blocked. The server keeps everything running while you are away; reattach with herdr.

3. One window for local and remote work. Keep your laptop's agents and your server's agents in the same sidebar. A long build or a big test suite runs on the remote machine while you work locally, and the combined agent list shows both.

4. Agent-orchestrated chores. A lead agent that can spawn helpers, wait for them, and read their reports turns one prompt into a pipeline: audit in two panes, review in a third, summary in the first. Useful for security audits, dependency checks, and test runs you would otherwise script by hand.

5. Phone check-ins. With the Herdr App, check which agent needs you from your phone, open the real terminal, and answer — type, paste a screenshot, or talk. The agents keep running on your machines while your laptop sleeps.

Verdict

Herdr earns the attention. The core idea is sound — when several agents run at once, knowing which one is stuck is the bottleneck, and a multiplexer that surfaces working, blocked, and idle per pane solves exactly that. The detach-and-reattach model is proven tmux territory, the CLI and socket API make genuine multi-agent pipelines possible without a framework, and the SSH story is the strongest of the four multiplexers compared here. The star count tells the same story: about 19,000 in late July 2026, about 49,500 on 9 November, 50,964 on 15 November — developers are voting with their terminals.

The caveats are real but narrow: Windows is beta, the project is eight months old and shipping fast, restart does not preserve processes, and the graphical app now leads the homepage while the video shows the TUI. None of that stops you from running the free TUI today. Install it, run two agents you already use, and watch the sidebar for a day. If the red dot saves you one hunt through four panes, it has already paid for itself — and it cost nothing.

Frequently Asked Questions

Is Herdr free?

Yes. Herdr is free and open source under the Apache-2.0 license, with no paid tier listed in its repository. The model subscriptions or API billing for Claude Code, Codex, Gemini, or Grok are separate — Herdr installs none of them and charges nothing for itself.

Is Herdr a replacement for tmux?

It is a tmux-style multiplexer rebuilt for AI agents, not a drop-in replacement. It keeps tmux's model — panes, tabs, detach and reattach, a Ctrl+b prefix — and adds per-pane agent state (working, blocked, idle), notifications, a socket and CLI API for agents, and SSH machines in one window. If you need tmux's decades of scripts and universal availability on servers, tmux is still the safer pick.

How do I install Herdr on Mac, Linux, and Windows?

Mac or Linux: curl -fsSL https://herdr.dev/install.sh | sh (read the script first), or brew install herdr, or mise use -g herdr. Windows: powershell -ExecutionPolicy Bypass -c "irm https://herdr.dev/install.ps1 | iex" — a beta preview. Then run herdr from any project directory.

Can agents really talk to each other in Herdr?

Yes, through the CLI and socket API — not by magic. One agent can spawn a pane, send another agent a prompt, and wait until it reports done; the newer herdr-core hc commands (hc run --bg, hc send) do the same with reports delivered back into the lead agent's input. You set the workflow up; the agents execute it.

What happens to my agents if I close Herdr or restart my machine?

Closing the client is safe: the background server keeps agents running, and herdr reattaches to the same session. Restarting the server or the machine is different — the saved layout restores and supported agent sessions can resume, but the original processes do not survive.

Does Herdr work on Windows?

There is a Windows beta preview with a PowerShell installer, and the graphical Herdr App lists Windows support. The terminal multiplexer experience is fullest on macOS and Linux; treat Windows as beta and check the project's Windows support docs before daily use.

Sources

• GitHub — herdrdev/herdr (README, license, install, features): https://github.com/herdrdev/herdr — checked 15 November 2026; 50,964 stars, 4,126 forks, Apache-2.0, created 27 March 2026, v0.9.5 (14 October 2026)
• Herdr docs — install and quick start: https://herdr.dev/ — checked 15 November 2026 (homepage describes the Herdr App, herdr-core, and the hc CLI)
• Herdr vs tmux vs Zellij vs cmux (comparison, verified 9 November 2026): https://andrew.ooo/answers/herdr-vs-tmux-vs-zellij-vs-cmux-ai-coding-agents-2026/
• Better Stack — Herdr: Terminal Multiplexer with Built-in AI Agent State Awareness: https://betterstack.com/community/guides/ai/herdr-ai-agent/ — published 8 June 2026; used for the interface, CLI commands, SSH --remote flag, and notifications
• NetworkChuck — "you need to use Herdr RIGHT NOW!!" (YouTube, 6 November 2026): https://www.youtube.com/watch?v=9_FzA-J9S2E — the walkthrough that sparked this article; includes an HPE Networking sponsorship segment

Related reading: gputaskman: the Free GPU Task Manager from Microsoft's Scott Hanselman · uv: the Python Package Manager That Replaced pip for Most Developers · Colibri: Run 744B AI Models on a Potato PC