Agent-Manager: A Tmux TUI for Running Claude Code, Codex and OpenCode

I've been using herdr for a couple of weeks. The only real advantage is that it makes the agent status glanceable in the sidebar, which can be nice if you don't want to cycle through your tmux panes to check if something's blocked or complete. Other than that, it's kind of just tmux (which is nice, I like tmux).

I've tested out just about every tool like this, and ended up coding my own, it is really minimal though.

The primary reason people reach for these tools are two to three reasons:

1. Tmux does not natively show agent statuses of agents / notify you when one needs input. Helpful when you have a huge list of small things to fix: I just spin up N agents in parallel to handle all of them, then I go over and review.

2. Tmux does not handle worktree handling. If you wanna make changes in parallel you cannot have two agents make db migrations at the same time. The way to solve for this is to have them work on two different worktrees with separate environments, ports etc.

3. Tmux tree view is not super beautiful, especially for viewing agents.

I built this as a tmux plugin since I don't really want to opt-out of tmux as a whole, especially since I like my keybindings there.

https://github.com/Ymirke/tmux-agent-switcher

My main reason for herdr over tmux is the priority queue of agents requiring intervention.

I built yoloAI, which is a sandboxer so it uses tmux to keep the agent alive inside. Since I can just reattach to them or use VS code tunnels, I don't really have a need for a multiplexer, though.

I’ve been using herdr lately and before that I was using zellij and before that tmux and before that gnu screen

For me the main advantage of herdr is that when I have the codex integration activated, herdr will keep track of the codex session ids and automatically restore them after restart

With the other multiplexers I had to create a separate tool that would save/restore layout/tabs/panes/session ids. And I had to run it manually. With herdr it just works, almost out of the box

Oh, it’s also very nice that you can create groups of tabs, so it’s easier to organize if you have more than 3-4 terminal sessions open (I usually keep at least 3 tabs, each split into an agent pane and a zsh pane)

And lastly, mouse scroll support is a lot better. I’ve had many issues with scrolling codex’s output. tmux and zellij’s scrolling wouldn’t play nicely with codex’s output, but herdr seems to work very well

- I get OS level notifications when the agent is ready.

- I don't need to use TMUX's wonky keyboard shortcuts

- I moved off of TMUX anyway, when ssh-ing into remote servers became 'not a thing'

- automatic coding session management, like auto resume, layout manamgement, etc. out of the box, without setting up TMUX, which could be a wormhole for some, by the time you fix mouse scrolling, cut/paste, etc...

I'm currently using CMUX, and been pretty happy with it it at this time, but loving the evolution of all these tools

Most of the AI tools people are building have little to zero value.

If you're making valuable tools, anthropic, openai, or one of the other big players in the space will hire you/buy you out and then they will release the tool themselves.

I'm getting really tired of everyone doing their own special rain dances and then acting like they made it rain.

I build one, an inhouse one for myself. I think it's a worth doing and see if it works for you. Especially for someone who works independently most of the time. I get it to follow and support my working style, workflow and idiosyncrasies. I started out because I wanted something that queues my work serially for each project so I can work on main and I can type what I want to do in it instead of writing to an ever-growing TODO list.

I then make it show icons and indicators for blockers, ready sessions that need attention, TODO lists that include notifications and actions (because it's so easy to miss things while working on/with many sessions). Shows costs (loop coordinator and non-LC costs, by day, etc) so I can optimize costs.

It's fun, a learning experience and it's more productive. With excellent SOTA and coding agents now, we should be building tools to improve our work even more than before. Try it.

yes exactly my thought. They're quite the same, a feature more a feature less

I tried herdr recently, and I missed many tmux features I really liked.

tmux has a better `prefix-[` mode where you 1) can do visual selection using line/block mode 2) you can use `o` to switch to other side (just like vim) 3) it pauses/freezes the state of the output so you have time to capture

In herdr, its very basic and incomplete if you're a vim/nvim/tmux power user. And if you're in a TUI that shows some text and then hides it, its hard to do the selection in herdr before it disappears.

To give herdr credit, I did like it's `prefix-e` which opens the history in vim so you can you real selection utils.

As I was looking at herdr, I found there many other solutions that were like it built on the more robust tmux. I am still evaluating, but there are plenty of solutions that have hooks to observe what's going on in agents in other panes/windows like herdr.

I like the Primeagen's `tmux sessionizer` style solutions for nvim to work with git worktrees that each agent may be in. And I like the tmux-agent-sidebar.

I started using herdr last night, and was able to get a workflow I've wanted to set up for a while working, which is to be able to talk to a managing agent in my car and check on / update claude sessions. It uses Telnyx SIP and LiveKit cloud, and I'm using Qwen3.6 35B-A3B quite successfully as the voice layer router. It's still very early but it works quite well.

If you decide to try it out, please make sure to read the security implications. There is no sandboxing added as of yet, and guardrails are pretty minimal.

https://github.com/jonesy827/mate

This is great! It would be really helpful if it had some measure of the number of users (stars?) and a contributor count. It's pretty hard to determine which of these projects are going to stick around. Total length of time under active development and percentage of code committed by agents would also be informative.

OPs tool is not a sandboxer and isn’t like the other projects in your list. OP is an agent multiplexer, of which there also are many similar projects, but not a sandboxer.

Do you know how does Herdr compare to native “claude agents” view?

Thanks for sharing your experience with Herdr! I've been considering taking the plunge for a few weeks - going to do a one day experiment today.

yes but if you don't create a company you can't get funding.

Different scope, as far as I can tell from their site: they describe a multiplexer for all work, spanning remote hosts, sandboxes, CI, production systems and shared terminals. This is a Go binary that drives tmux on the machine you're already on, and every session it creates is an ordinary tmux session you can attach to without it. Much smaller bet, and it stops where tmux stops.

I used exactly this setup before (mosh + tmux + Tailscale + VPS), so maybe I can help. Both tools are terminal UIs, so they fit naturally in what you have: you run them inside tmux on the VPS, and mosh handles the bad connection like it does for everything else. But honestly they don't change much when the connection is bad, because the UI still goes through the connection. This is one of the reasons I built ShellTeam (github.com/sebderhy/shellteam): the sessions stay on the VPS and the UI is a web app, so when I lose internet on a train I just reload the page and everything is still there. I'm obviously biased, but for intermittent connections I think a web UI works better than a TUI over mosh.

It's useful for running multiple long running tasks alongside your "primary" work. As an example I currently have 3 tabs open running a PR review, 1 tab running "autonomous development" for a small time feature and then my "primary" window for working on a feature.

These types of responses are always so low value, they should be auto filtered out. How is Claude Code different than Codex, how is x different from y, etc. Read the repo and figure it our for yourself, good grief.

agent-deck was the first one I actually tried to adopt, landed a couple of commits in it. wanted it to be the one.

what I kept missing was answering an agent without attaching to it. so space on a blocked one, type, enter, and it lands in that pane. on a project row it spawns a new agent already working on what you typed. the other half is review, ctrl+r opens what it changed as whole files with the diff tinted, c comments a line, C sends the notes into the pane.

deck does plenty I don't, worktrees, cost, forking. different bets.

Normal GUI runs local. I have mosh running from Blink on my iPhone to my server and I can continue months long session of AI

herdr's a whole runtime, own multiplexer, plugins, socket api. mine sits on tmux and doesn't try to replace it, sessions stay normal tmux sessions.

overlap on the status sidebar is real. two things I lean on that I couldn't find a key for over there: space on a blocked agent types the answer into its pane without attaching, and ctrl+r opens its repo as full-file diffs where a comment on a line goes back to that agent as a prompt. herdr's socket api can push text into a pane, so the first one is scriptable at least. correct me if I missed something.

mouse and windows, herdr's ahead.

To be clear on the GUI point, since my last reply here answered a different question: this is a terminal tool and it won't convert you. The reason I stay in one is that the agents already live in terminals, so the manager is a view over what's running rather than a second place things live.

No need to comment.

100% agree. AI isn't hacking, it's a completely different crowd. So much front page space is wasted talking about the latest version, some funding controversy, or these useless add-on tools.

Yes, I've found that T3 Code fills this gap nicely.

You're not, and this won't convert you. The reason I stay in a terminal is that the agents already live in one, so a TUI is a view over what's running rather than a second place things live. If your agent already runs in a GUI, that argument doesn't apply to you.

I had the same experience, but someone calmed me down by saying "we're all working on solving this problem."

I’ve been working on something like this for a few years. Literally.

And I’ve noticed that my hobby project went from being a niche toy to something everyone was independently building almost overnight. My guess is that AI now makes it trivial for people to get started.

In any case, I since decided it would be better to focus on building my tool for myself rather than releasing it, like I was originally planning to do.

I'm seeing a lot of this across all sorts of apps right now. We're at a sort of cambrian explosion point for software.

I too built my own equivalent of this (belfry.robgough.net). I take it as a good thing, we're all using these tools and coming to similar conclusions.

I suspect over time it'll settle down, once we've gotten comfortable with the new ways of working and found the new normal.