Herdr is an open source terminal multiplexer built for AI coding agents, and this Herdr tutorial distills the ten best videos about it into one guide to running Claude Code, Codex and other agents side by side. The videos come from ten sessions by nine creators, about three hours in total, published between September 7 and September 28, 2026. I read every transcript end to end, checked the commands against the official Herdr documentation, and rebuilt the material in the order you would actually do the work. The video above, Dave Ebbelaar’s full setup guide, is the one to watch if you only watch one. Notes on all ten follow the tutorial.

TLDW

Herdr does three things that a plain terminal does not: it keeps every agent running in a background server after you close the window, it shows whether each agent is working, blocked or idle, and it exposes a command line interface that the agents themselves can use to open panes and prompt each other. Installing it is one command, and you can drive it entirely with the mouse, so the learning curve is the mental model (session, space, tab, pane) and not the keyboard. The setup that the more experienced creators converge on is Herdr on a rented server with one orchestrating agent handing work to others in separate git worktrees. The caveats are real: it is version 0.9, backed by a US$6 million seed round, and an always-on API that lets agents spawn terminals deserves the same care as any agent with shell access.

The 10 Videos at a Glance

#	Video	Channel	Length	Best for
1	How to set up Herdr for multi-agent coding (full guide)	Dave Ebbelaar	28 min	The complete first setup, if you come from Cursor or a desktop app and not from the terminal
2	Herdr: My Command Center for Claude Code, Codex, and VPS	Promovaweb (Portuguese)	29 min	Seeing a month of real daily use from a server, including from a phone
3	Herdr Is AMAZING: Tutorial and Setup (VPS)	ByteGrad	19 min	A fast overview, and letting an agent do the server setup for you
4	Run a Team of Coding Agents in the Cloud with Herdr	Dave Ebbelaar	26 min	The careful server build: SSH keys, a non-root user, firewall, logins
5	Herdr: 1 Million Installs in 6 Months	AwesomeFOSS	8 min	Deciding whether to bother, with the numbers and the caveats
6	Herdr: Managing Agents Like A Power User	NeuralNine	20 min	Keyboard shortcuts, the config file and the command line
7	Herdr: Run Parallel AI Agents with Tidewave	Dashbit	9 min	Git worktrees, so parallel agents do not collide
8	How I Orchestrate AI Coding Agents with Herdr + AI DevKit	Codeaholicguy	9 min	One manager agent spawning six workers
9	Herdr Shortcuts: Understand Ctrl+B	yousef	10 min	Absolute beginners on Windows who have never seen a prefix key
10	Herdr’s Agent-to-Agent is WILD	Fru Dev	16 min	Agent to agent messaging, and uses outside of coding

Thoughts

Ten creators who did not coordinate describe the same product in almost the same three sentences, which is unusual and worth noticing. Persistence, status, and control by the agents themselves. The first two are tmux with better manners. The third is the actual product. AwesomeFOSS calls it “quietly the biggest idea” in the tool, and the videos that are most excited (Dave Ebbelaar watching Codex open a tab and start Claude in it, NeuralNine asking one agent to summarize two others, Codeaholicguy spawning six workers, Fru Dev routing messages between tabs) are all excited about that same feature. If you install Herdr and only use it as a nicer set of terminal tabs, you have skipped the part everyone is talking about.

The adoption story is the mouse. Dave Ebbelaar says several times that he is “not a terminal wizard” and had never used tmux, Neovim, lazygit or zoxide before this month, after ten years in tech. yousef is on Windows with the plain command prompt and admits he used Herdr for a week without knowing the shortcuts. NeuralNine, at the other end, rebinds the prefix to Ctrl+A and moves with Vim keys. All three are comfortable in the same tool. A terminal program that right-clicks is what lets people leave the Codex and Claude desktop apps, and the reason they want to leave is specific: the moment you use more than one vendor’s agent, no single vendor’s app can hold your work.

Three of the ten videos are really about renting a server, and the Portuguese video makes the case best: Luiz at Promovaweb runs the cheapest Hetzner machine, had not restarted it in almost a month, and reconnects from Windows, Mac, Linux, a tablet and a phone to the same screen. Laptops sleep, and Herdr on a laptop only survives a closed terminal, not a closed lid. What differs is the care taken. Dave builds a non-root user, key login and an IP firewall. ByteGrad lets a browser-driving agent do the setup and offers to copy agent credentials onto the box, then says plainly that it is an experiment and not for anything important. Follow Dave’s version.

What nobody covers is more telling than what they do. Not one video mentions what five simultaneous top-tier agents cost, although ByteGrad lets slip that fast mode is “draining your budget”. Only AwesomeFOSS reports that the original processes do not survive a full machine restart (the layout comes back, and an agent resumes only if it supports resuming), a detail that the “close your laptop, nothing is lost” framing in the other videos glosses over. Only Dashbit points out that worktrees move the conflict to the pull request and do not remove it. And only NeuralNine shows the approval question on screen: his orchestrating agent stopped and asked him before pressing Enter on another agent’s permission prompt, because Herdr’s own skill tells it to. That one behaviour is the difference between delegation and an unattended loop of agents approving each other.

The last two videos point somewhere the first eight do not. Fru Dev’s demonstration has nothing to do with code: a “chief” agent as the only place you type, and one long-lived agent per folder of your life (insurance, taxes, health) that keeps its own context and history. Codeaholicguy makes the same point from the engineering side, that a spawned Herdr agent is a real process you can click into and steer, unlike a subagent whose work you never see. And Dashbit closes with the most useful sentence in the whole set, that worktree, environment file, dependencies and port are “reusable principles” and Herdr is one popular place to apply them. That is the right way to hold this tool. It does not wrap or replace your agents, so leaving costs nothing, and the habits you build in it will outlast it.

The Tutorial: Herdr Step by Step

The steps are in the order you would do them, not the order of the videos. Each ends with the moments in the videos where you can watch it happen.

Step 1: Understand what you are installing

Herdr is a single binary for macOS, Linux and Windows, released under the Apache 2.0 licence. It is a server that runs in the background and a client that draws in whatever terminal you already use. Your agents (Claude Code, Codex, Grok, opencode, Pi, Cursor and others, 22 detected out of the box according to the Herdr site) run inside it exactly as they do today, with the same logins and subscriptions. Herdr is not a model and does not wrap them. It owns the terminals. If you know tmux, it is tmux that understands agents. If you do not, think of the project list in the Codex or Claude desktop app, except that any agent from any vendor can sit in it.

Shown in: AwesomeFOSS at 0:00, ByteGrad at 2:16, Dave Ebbelaar at 1:30.

Step 2: Pick a terminal that passes the mouse through

Herdr is “mouse first”, so the terminal matters. Dave Ebbelaar found that iTerm blocked some mouse actions and shortcuts and moved to Ghostty (macOS and Linux). yousef had trouble with PowerShell on Windows and got it working in the plain command prompt. On a phone or tablet, Luiz uses Termius. If a click or a shortcut does nothing, suspect the terminal’s own key bindings before you suspect Herdr.

Shown in: Dave Ebbelaar at 3:47, yousef at 6:50, Promovaweb at 23:39.

Step 3: Install and start it

macOS and Linux:

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

Windows (PowerShell):

irm https://herdr.dev/install.ps1 | iex

Homebrew, Nix and manual installs are listed in the docs. Running the installer again updates an existing install. Then type herdr. On a fresh Linux server the installer may tell you the binary is not on your PATH yet and print the line to add; run that line first.

herdr

Shown in: NeuralNine at 1:31, Dave Ebbelaar at 0:45, ByteGrad at 9:52.

Step 4: Learn the four levels

Everything in Herdr is one of four things, from the outside in:

• Session. A whole Herdr workspace on the server. You get one called default. Extra named sessions are for grouping, for example one per client.
• Space (also called a workspace). One project, which in practice means one folder or git repository. The left sidebar lists them with their branch.
• Tab. A screen inside a space. One for agents, one for tests, one for the dev server.
• Pane. One terminal inside a tab. Split a tab right or down to get more.

Below the spaces, the sidebar has an Agents list. A pane appears there only once you start an agent in it, so a plain shell or a running dev server stays out of the way. To make a space, click New and cd into the project. To start an agent, type its normal command (claude, codex) in a pane. Right-click any pane, tab or space for split, rename and close. Rename things as you go: a tab called “deploy” and a pane called “server 01” are what make a screen of ten terminals readable.

Shown in: Dave Ebbelaar at 5:19, Promovaweb at 7:42, NeuralNine at 11:25.

Step 5: Prove persistence to yourself

Do this once, because it is the reason the tool exists. Start an agent, give it a task that takes a minute, then quit the terminal application completely. Open a new terminal and type herdr. Everything is where you left it and the agent kept working. To leave on purpose without quitting anything, detach with the prefix key and then q (by default Ctrl+B, then q).

Know the limits. On a laptop this survives a closed terminal, not a sleeping or powered-off machine. After a full restart Herdr restores the layout, but the original processes are gone, and an agent picks up again only if it supports resuming a session. Installing the integrations in the next step is what makes that resume work.

Shown in: Dave Ebbelaar at 2:16, NeuralNine at 3:48, AwesomeFOSS at 6:51.

Step 6: Install the integrations and read the status lights

Herdr detects agents on its own, but the integrations make status reporting and session resume reliable. Either open Menu, Settings, Integrations and tick each agent, or run one command per agent:

herdr integration install claude
herdr integration install codex

Each agent in the sidebar is then in one of a few states: working, blocked (it has asked a question or needs a permission), idle, and done. In the themes shown in the videos, working is yellow, blocked is red and done is green, and a sound plays when an agent needs you. Clicking an agent in the list jumps to the right space, tab and pane. This is the feature that replaces walking from window to window to find out who is waiting.

Shown in: Dave Ebbelaar at 8:26, NeuralNine at 4:35, ByteGrad at 3:03, Dave Ebbelaar (cloud) at 21:52.

Step 7: Learn six shortcuts, then stop

Shortcuts are two steps: press the prefix, let go, press a key. The default prefix is Ctrl+B, the same as tmux. Press the prefix and then ? to see every binding. The full table is in the cheat sheet below. The six worth learning first are split down, split right, new tab, next tab, close pane and detach.

Everything is set in one file, ~/.config/herdr/config.toml on macOS and Linux and %APPDATA%\herdr\config.toml on Windows. A minimal example that changes the prefix:

[keys]
prefix = "ctrl+a"
new_tab = "prefix+c"
next_tab = "prefix+n"
previous_tab = "prefix+p"
split_horizontal = "prefix+minus"

After editing, reload without restarting anything:

herdr server reload-config

Two approaches to the config appear in the videos. NeuralNine keeps the prefix style and adds direct bindings (Ctrl+Alt plus arrows to move between panes, a custom key to move a pane into its own tab). Dave Ebbelaar removes the prefix entirely so Herdr behaves like a normal app: Cmd+T for a tab, Cmd+W to close, Cmd+K for a searchable jump to any space, and a key that jumps to whichever agent just finished. He did not write the file. He pointed Codex at the config and the docs and described what he wanted, which is the fastest route for anyone.

Shown in: NeuralNine at 6:54, NeuralNine at 9:10, Dave Ebbelaar at 14:33, yousef at 5:18.

Step 8: Let one agent drive the others

Anything you can click, the command line can do: herdr session list, herdr workspace list, herdr pane rename, and so on (herdr --help lists it all). Herdr ships a skill file that teaches an agent those commands. Install the skill globally for every agent you use (the docs have the one-line installer), and then you can type a request like this into any agent:

Use the herdr skill to open a new pane to the right in this tab,
start Claude Code there, and have it review this codebase.
Tell me when it is done and summarize what it found.

The first agent reads the skill, splits the pane, launches the second agent, types the prompt into it, waits for its status to change, and reads the result back. You watch it happen in real panes and can click into any of them to steer. This works across vendors, so a Codex orchestrator can hand implementation to Claude and review to Grok. Expect the orchestrator to ask you before it approves another agent’s permission prompt. That is in the skill on purpose. Fru Dev shows that simple messaging between two named tabs works even without the skill installed.

Shown in: Dave Ebbelaar at 17:33, NeuralNine at 14:30, Codeaholicguy at 3:03, Fru Dev at 0:54.

Step 9: Give parallel agents their own worktrees

Two agents editing the same checkout will overwrite each other. Dashbit’s four steps solve it, and they apply to any tool, not only Herdr:

• Create a git worktree. Right-click the space, choose New worktree, name the branch. It appears under the main checkout in the sidebar.
• Copy the environment files. Files like .env.local are not in git, so a new worktree does not have them. Claude Code and Codex read a .worktreeinclude file that lists what to copy. Herdr does not read it by default, so Dashbit uses a community plugin that does.
• Install dependencies in the worktree (npm install, mix setup).
• Run the dev server on its own port so ten servers do not collide.

PORT=3001 npm run dev

Repeat for each agent with the next port. Locally the agents cannot touch each other’s files. The conflicts arrive later, when the branches are merged. Delete a worktree from the same right-click menu when the branch is done.

Shown in: Dashbit at 1:32, Dashbit at 2:19, Dashbit at 3:53, ByteGrad at 17:29.

Step 10: Move it to a server so the lid can close

This is optional and it is where the tool pays off. The outline, following Dave Ebbelaar’s careful version:

• Rent a small Linux server. He suggests at least 8 GB of RAM. Both he and ByteGrad use Hostinger, which has a server template with Herdr preinstalled. Luiz runs the cheapest Hetzner plan and mentions the free Oracle tier.
• Create an SSH key on your own machine with ssh-keygen and add the public key to the server when you create it. Keep the root password in a password manager.
• Add a shortcut to ~/.ssh/config so you connect by name instead of IP address.
• Create a non-root user for daily work, give it your key, and install Herdr for that user. whoami tells you which account you are in.
• Add a firewall rule in the host’s panel that allows port 22 only from your IP address and drops everything else. This needs a fixed address, which usually means a VPN with a dedicated IP.
• Log in to everything on the server. gh auth login for GitHub, claude auth login for Claude Code, and the device login for Codex. A server has no browser, so each prints a URL and a code that you open on your own machine.
• Make a folder for repositories, clone your projects into it, install the integrations and the skill as in steps 6 and 8, and start herdr.

Host agent-cloud
    HostName YOUR.SERVER.IP
    User developer

After that, ssh agent-cloud and then herdr puts you back in the same screen from any device. Two refinements: Cursor and VS Code can connect to the same server over SSH, which gives you a file tree and a diff view beside the Herdr session, and Herdr can attach to a remote machine directly from your local client, keeping your local key bindings (herdr --remote user@host, or save the machine with herdr machine add).

Shown in: Dave Ebbelaar (cloud) at 4:51, Dave Ebbelaar (cloud) at 11:00, Dave Ebbelaar (cloud) at 13:17, Dave Ebbelaar (cloud) at 16:22, NeuralNine at 16:49, Promovaweb at 1:34.

Step 11: Fill the gaps an IDE used to fill

Leaving an editor means losing the file tree and the git panel. The creators replace them with terminal tools kept in their own tabs:

• Neovim (nvim .) as a file browser in the first tab.
• lazygit for branches, diffs, commits and pushes. Luiz adds lazydocker for containers and btop for watching the machine.
• zoxide to jump to a project by name (z cookbook, or zi for a searchable list) instead of pasting paths.
• A floating notifier. ByteGrad uses a community heads-up display that pops up over a full-screen application when an agent finishes.
• Plugins. The marketplace had more than 1,300 as of the September 26 AwesomeFOSS video. Luiz uses two, one for an editor-like layout and one for naming windows.
• Tidewave for browser work: click an element or run an accessibility scan on the running app, copy the generated prompt, paste it into the agent in that worktree.
• AI DevKit if you want a ready-made agent management skill with Herdr as its runtime.

Shown in: Dave Ebbelaar at 21:20, Dave Ebbelaar at 24:21, Promovaweb at 18:17, ByteGrad at 17:29, Dashbit at 4:38.

Step 12: Settle into a daily pattern

The patterns the creators have landed on after one to four weeks are worth copying outright. Keep one long-lived tab per concern (deploy, copy, specs, implementation) so each conversation keeps its history and you never mix topics. Use one agent to build, a second to check the first, a third to test and deploy. Keep a single “chief” agent as the place you type and let it route work to the others. And when you come back from a break, look at the Agents list before anything else: red first, then green, and leave yellow alone.

Shown in: Promovaweb at 22:06, Promovaweb at 26:47, Fru Dev at 4:46, Dave Ebbelaar at 16:03.

Cheat Sheet

Default key bindings. “Prefix” is Ctrl+B unless you changed it.

Keys	What it does
Prefix, then ?	Show every key binding
Prefix, then –	Split the pane (new pane below)
Prefix, then v	Split the pane (new pane to the right)
Prefix, then h / j / k / l	Move focus left / down / up / right
Prefix, then x	Close the pane
Prefix, then c	New tab
Prefix, then n / p	Next / previous tab
Prefix, then Shift+X	Close the tab
Prefix, then Shift+N	New space
Prefix, then w	Move through the list of spaces
Prefix, then Shift+W	Rename the space
Prefix, then Shift+D	Delete the space and everything in it
Prefix, then s	Settings (themes, sounds, integrations)
Prefix, then q	Detach. Everything keeps running

Commands.

Command	What it does
herdr	Start, or reattach to the default session
herdr --help	Every command, and the path of your config file
herdr session list	List sessions on this machine
herdr workspace list	List spaces and their IDs
herdr workspace close <id>	Delete a space
herdr integration install <agent>	Install the integration for one agent
herdr server reload-config	Apply changes to config.toml
herdr server stop	Stop the background server. Spaces are kept
herdr --remote user@host	Attach to Herdr on another machine with your local settings
herdr machine add <name>	Save a remote machine so its agents show beside local ones

The Videos, One by One

1. How to set up Herdr for multi-agent coding (full guide)

Dave Ebbelaar, 28 minutes. Watch on YouTube. This is the video embedded at the top of the post.

A week of daily use written up as a guide by someone who had never used tmux and had been living in the Codex and Claude desktop apps. It covers install, spaces, integrations, sessions, a prefix-free keyboard setup, agent delegation, and the terminal tools that replaced his editor. The most complete single video in the set.

Key takeaways

• The reason to switch is mixing vendors. A single-vendor desktop app is excellent until you want Claude, Codex and Grok reviewing each other.
• Install the integrations first, then the Herdr skill globally for every agent, or delegation will not work.
• He never read his own config file. He pointed Codex at it and at the docs and described the result he wanted.
• His shortcuts drop the prefix: Cmd+T, Cmd+W, Cmd+K to search and jump, and a key that jumps to the agent that just finished.
• Neovim, lazygit and zoxide are what made Herdr a complete replacement for an editor, and all three were new to him.

Chapters

• 0:45 Install, first run, and the empty screen that confuses everyone.
• 5:19 Spaces as projects, panes, tabs, and the Agents list.
• 8:26 Settings, integrations, and named sessions.
• 12:16 The config file, colours, and a keyboard setup without a prefix.
• 17:33 Agent delegation: Codex opens a tab, starts Claude and prompts it.
• 21:20 Neovim, lazygit and zoxide to replace the editor.
• 25:53 Starting a Herdr agent from a dictation app through a custom MCP server.

“To me this is what AGI feels like. Just like you give it a task and it just like it navigates. It knows what to do and it just runs.”

Dave Ebbelaar, watching Codex start a Claude session in a new tab and type the prompt into it

2. Herdr: My Command Center for Claude Code, Codex, and VPS

Promovaweb (Luiz), in Portuguese, 29 minutes. Watch on YouTube.

A long-time tmux user shows the setup he has run for about a month: Herdr on a small server in Germany, reached from every device he owns. Less a tutorial than a tour of a working day, and the best evidence in the set that the server approach holds up. The transcript used here is an English translation.

Key takeaways

• His server had not been restarted in almost a month, and a class he recorded on Friday was on screen, unchanged, on Monday.
• He prefers Herdr to the Orca desktop app because Orca needs a good machine that stays on, and its mobile app is a different program. On a phone he opens the real Herdr in Termius.
• Herdr is useful without any agents: one tab of htop panes watching several servers over SSH, each pane renamed for its machine.
• He keeps one permanent tab per concern (site, copy, development, deploy) so each agent conversation keeps its own history.
• The only shortcut he says you need is Ctrl+B then q. And you cannot run Herdr inside Herdr.

Chapters

• 0:47 What tmux did for years, and why closing a laptop kills an agent.
• 1:34 Running on a server and connecting from Windows, Mac, Linux, tablet and phone.
• 7:42 Tabs, spaces, the Agents list and what the status colours mean.
• 9:14 Herdr compared with the Orca desktop app.
• 13:01 A monitoring tab: split panes, htop on several machines, renamed panes.
• 18:17 lazygit, lazydocker and btop in their own tabs.
• 23:39 The same session opened on a phone.
• 26:47 Three Codex agents: one builds, one checks, one tests and deploys.

“I leave, I go to sleep. The moment I close my laptop, it’s dead. So Herdr has the function of keeping that running for you.”

Luiz of Promovaweb, translated from Portuguese, on why a normal terminal fails for long agent runs

3. Herdr Is AMAZING: Tutorial and Setup (VPS)

ByteGrad, 19 minutes. Watch on YouTube.

A quick product overview followed by a server install from a plain Ubuntu image, with a twist: he hands the setup to the Codex desktop app, which drives the host’s web terminal in its built-in browser.

Key takeaways

• His working layout is two different agents side by side (Codex, and opencode with a DeepSeek model) with Neovim along the bottom.
• An agent on your own computer can do the whole server setup through a web console: install the agents, create the project folder, build a 2 by 2 layout, rename the tabs.
• He describes this as experimental and says not to do it with important projects or sensitive data.
• opencode crashed when its pane was too narrow. Layout can break an agent.
• Separate spaces do not isolate agents working on the same repository. You still need git worktrees.

Chapters

• 0:00 A working layout: two agents, an editor pane, a checks tab.
• 2:16 How Herdr describes itself, and the three statuses.
• 3:03 A floating notifier that shows finished agents over a full-screen app.
• 9:07 SSH in, create a user, install Herdr, fix the PATH.
• 11:24 Handing the rest of the setup to an agent through the web terminal.
• 15:11 Asking for a 2 by 2 and a 3 by 1 layout instead of building them by hand.
• 16:43 Reconnecting from a laptop, then worktrees and automations.

“Herdr does not wrap them or replace them. It just owns the terminals.”

ByteGrad, reading and endorsing the line from the Herdr site that explains why your agent setup does not change

4. Run a Team of Coding Agents in the Cloud with Herdr

Dave Ebbelaar, 26 minutes. Watch on YouTube.

The follow-up to video 1, two weeks later: the same setup moved onto a rented server, done properly. This is the one to follow line by line for step 10, because it is the only video that treats the server as something to secure.

Key takeaways

• Use a server with at least 8 GB of RAM. Hosting prices have doubled or tripled over the past year because of a RAM shortage, by his account.
• Two accounts: root for system changes, a “developer” user for daily work and for the agents.
• Key login plus a firewall rule that allows port 22 only from a fixed IP. A firewall configuration can be reused across servers.
• Every login on a server is a device login: the tool prints a URL and a code, and you finish in the browser on your own machine. Codex may need device login enabled in your account’s security settings first.
• Cursor or VS Code over SSH gives you a file tree and source control on the server, with Herdr running in the editor’s terminal panel.

Chapters

• 0:47 The architecture: a server holds the work, devices tap into it.
• 4:51 Root password and SSH key.
• 8:39 Connecting from your own terminal, and an SSH config shortcut.
• 11:00 A non-root developer user.
• 13:17 Firewall: one port, one IP address.
• 16:22 Installing and logging in to GitHub, Claude Code and Codex.
• 21:06 A repositories folder, integrations and the skill.
• 23:27 Opening the same server in Cursor over SSH.

“We want to essentially disconnect them from the physical devices that we have around them and rather have a centralized place where all the work lives.”

Dave Ebbelaar, on why he expects agents to move off laptops and onto servers

5. Herdr: 1 Million Installs in 6 Months

AwesomeFOSS, 8 minutes. Watch on YouTube.

No setup at all. A fast, numbers-led briefing on what Herdr is, how quickly it grew, and four things to know before depending on it. Watch this first if you are still deciding. All figures are as of its September 26 publication.

Key takeaways

• The repository appeared on March 27 with version 0.1.0 the same day. Six months later: about 41,000 GitHub stars, over a million release downloads, 88 contributors, 89 releases, version 0.9.1.
• It had passed Zellij in stars and was within about 9,000 of tmux, which has been around since 2007.
• The company has raised US$6 million. The code is Apache 2.0 and can be forked, but something adjacent will be monetized eventually.
• After a full machine restart the layout is restored but the original processes are not. Resuming depends on the agent.
• If you run one agent once a week, the tmux you already know is fine. At three agents a day it becomes the missing layer.

Chapters

• 0:45 Downloads, stars, plugins, contributors, release pace.
• 1:32 The US$6 million seed round.
• 3:03 The feature list, ending on agent-native control.
• 4:35 The star chart against tmux and Zellij.
• 5:20 A before and after of daily life with several agents.
• 6:51 Four honest caveats.
• 7:36 The verdict.

“The multiplexer wars are back, and this time the contestants are judged on how well they babysit robots.”

AwesomeFOSS, on why a six-month-old tool is growing this fast in the most conservative category in computing

6. Herdr: Managing Agents Like A Power User

NeuralNine, 20 minutes. Watch on YouTube.

The keyboard and command line video, from a Linux and Vim user. It is the clearest walk through the prefix keys and the config file, and it has the best demonstration of one agent controlling others, including the moment it stops to ask permission.

Key takeaways

• He names the status colours as he watches them change: yellow while working, red when approval is needed, green when done, with a sound on red.
• Config changes do nothing until you run herdr server reload-config.
• Bindings can replace the defaults or extend them. He keeps prefix plus h, j, k, l and adds Ctrl+Alt plus arrows.
• Stopping the server does not delete spaces. You have to close them, with a key or from the command line.
• He thought the sidebar had a bug showing the wrong project, then corrected himself: it shows the space name, not the current folder.

Chapters

• 1:31 Install and first look.
• 3:48 Persistence, shown by closing the terminal mid-task.
• 4:35 Two agents, two projects, and the status colours.
• 6:54 The prefix key and every basic shortcut.
• 9:10 config.toml: custom bindings and settings.
• 12:56 Reloading config and the command line tool.
• 14:30 One Claude instance opens panes, starts two agents and prompts them.
• 16:49 Summarizing other agents, and connecting to a remote server.

“So I can have one agent managing the other agents, which makes it easier for me to just communicate with this one agent.”

NeuralNine, after his orchestrating agent summarized the work of two others

7. Herdr: Run Parallel AI Agents with Tidewave

Dashbit, 9 minutes. Watch on YouTube.

The tightest video in the set. Four steps for running agents in parallel without collisions, shown first on a Next.js site and then on an Elixir and Phoenix project, followed by Tidewave for sending browser context to an agent.

Key takeaways

• The four steps: new worktree, copy environment files, install dependencies, run the dev server on its own port.
• Claude Code and Codex honour a .worktreeinclude file for copying environment files. Herdr needs a community plugin to do the same.
• Worktrees prevent local collisions only. Merge conflicts still appear when the pull requests are opened.
• Tidewave’s toolbar turns a click on a page element, or an accessibility scan, into a prompt you paste into the agent for that worktree.
• The same worktrees and running servers work with Tidewave’s graphical app, which has a code review tab for the diff.

Chapters

• 0:46 The interface, and using the mouse.
• 1:32 The four steps, and creating a worktree from the right-click menu.
• 2:19 Environment files and the worktree include plugin.
• 3:53 Two dev servers on two ports from two branches.
• 4:38 Tidewave: browser context for the agent.
• 6:55 Two agents, two models, two tasks from the toolbar.
• 7:41 The same setup from the Tidewave app.

“These are reusable principles, and Herdr was just one very popular example.”

Dashbit, closing the video by pointing out that the four steps work in any agent tool

8. How I Orchestrate AI Coding Agents with Herdr + AI DevKit

Codeaholicguy, 9 minutes. Watch on YouTube.

A short, unpolished demo from the author of AI DevKit, which used tmux as its runtime for spawning agents and now supports Herdr as well. One Codex session scans a command for performance problems, then spawns a separate agent for each improvement.

Key takeaways

• AI DevKit installs with npm and has a setup command that checks your machine and wires it to the agents you already have.
• Its built-in agent management skill first lists running agents, reuses one if it fits, and otherwise spawns a new one of the type you asked for.
• You can name the agent type per task: a Pi agent to brainstorm one item, Codex workers for the rest.
• The manager watches its workers and tries to unblock them without you.
• Six Codex workers were running from one instruction, each visible in the sidebar.

Chapters

• 0:00 Herdr as a second runtime beside tmux.
• 0:47 Install and setup in about thirty seconds.
• 1:32 A Codex session proposes a list of performance improvements.
• 3:03 Spawning a Pi agent for the first item.
• 5:21 Steering a worker directly by clicking into it.
• 7:37 Six Codex workers started and instructed.

“You can see a new agent is spawn and I never do anything to spawn it. The whole spawning process come from the agent.”

Codeaholicguy, pointing at the sidebar as the manager agent creates its first worker

9. Herdr Shortcuts: Understand Ctrl+B

yousef, 10 minutes. Watch on YouTube.

An unedited clip from a live stream by a Windows user learning in public. It covers one idea, the prefix key, slowly and with mistakes left in. Skip it if you have used tmux. Watch it if “press Ctrl+B, let go, then press a key” is new to you.

Key takeaways

• “Prefix” in the help screen just means Ctrl+B. It is written that way because you can change it.
• Pressing the prefix shows a bar at the bottom of the screen, and ? from there opens the full list of bindings. That list is how he learned the tool.
• He used Herdr for more than a week by clicking, without knowing the shortcuts.
• PowerShell gave him trouble on the day. The plain command prompt worked.
• Typing an agent’s command in a pane only works if that agent is installed. Otherwise you get an ordinary shell error.

Chapters

• 0:46 What Herdr is, and two finished projects showing green.
• 1:31 A first split and a first close with the prefix.
• 3:47 Doing the same things with the mouse and the menu.
• 5:18 What “prefix” means.
• 6:50 Shells on Windows, and the PowerShell problem.
• 7:37 The key binding list, a new space, and a split to the right.

“You don’t necessarily need to know these shortcuts. I’ve been using Herdr without knowing all these shortcuts because, and this is part of why people really like it, it has the option to just click.”

yousef, explaining why a shortcut video is still useful to people who will never use the shortcuts

10. Herdr’s Agent-to-Agent is WILD

Fru Dev, 16 minutes. Watch on YouTube.

Two Claude sessions in two tabs pass messages to each other with no skill installed, and the rest of the video is about what to build on that: a single “chief” agent in front of one long-lived agent per area of your life. The enthusiasm is turned up high, and the second half promotes his own app, but the idea is the most original in the set.

Key takeaways

• A plain instruction (“using Herdr, find agent B and send this exact message”) was enough. The agent listed the panes, found the target, sent the text and verified it landed.
• Each space points at a folder, and the folder is the context: every insurance document in one, every tax document in another.
• The history stays in the receiving agent’s tab, so you can drop into it later and carry on that conversation directly.
• For coding, the same shape is a driver agent with separate front end and back end agents in their own spaces.
• He plans a small decision model in front of the chief to choose which agents a request should go to, and whether it deserves several models’ opinions. This part is a plan, not a demo.

Chapters

• 0:54 Agent A finds agent B and sends it a message.
• 4:00 Agent B writes its paragraph, unprompted by a human.
• 4:46 The chief of staff idea.
• 7:03 An example: a new law, researched and passed to the insurance agent.
• 8:33 The same pattern for a codebase.
• 11:36 A decision model to route requests.
• 13:53 Messaging across vendors, and a council of models.

“This now becomes your chief of staff, your main entry point.”

Fru Dev, on talking to one agent in the morning and letting it brief the others

Start with the eight-minute AwesomeFOSS briefing if you are undecided, Dave Ebbelaar’s full guide to get set up, Dashbit’s worktree video before you run agents in parallel, and the cloud setup once you want the work to continue with the laptop closed.

Related Reading

• Herdr the official site, with the install commands and the list of supported agents.
• Herdr documentation configuration, agents, connecting machines and the socket API, and the source for every command in this post.
• Herdr on GitHub the Apache 2.0 source, releases and issue tracker.
• tmux (Wikipedia) the terminal multiplexer Herdr borrows its prefix key and detach model from.
• git worktree the git feature that lets several agents work on one repository at once.