I've tried Cmux, but it didn't do it for me, since the agent statuses were displayed on the workspaces and having multiple agents in the same workspace would sometime produce confusing results.
I've been using Herdr since the start of the week and so far it's been the best in terms of visibility of what my agents are doing and which of them need attention. The only wart I've noticed so far is that the performance is not always great — sometimes I see the text appearing with a noticeable delay as I type it.
Then I just work through the alerts and handle them as they arrive.
I guess there's an edge case where you don't know if an agent is running in a tab at all (while herdr will show "Agent running") but in practice I'm the one starting agents, and they run until they alert, so I'm not hitting a case where I think an agent is running but it's not.
The hitching is compounded for me because I tend to run the agents in squads: the agent I talk to operates a strict no-coding 'producer' mode, it tasks a sub-agent to do the research or coding, then the results go via a file to a critic or review agent; keeps the producer context very minimal and lean. Not convinced it's as necessary as just starting new contexts frequently with Fable etc.
My general rule is that I won't commit code a human hasn't seen/reviewed to production codebases, and I know I won't maintain that rule if I have to read all the slop that gets generated first time round without an AI reviewer pass.
So far my producer skill has survived 4.6 thru fable in succeeding to treat the review/critic output skeptically, as a likely yes-man or team player.
The key is to remember that, as of Fable, the size of the training corpus segment representing people responding to AI-generated content is still relatively tiny. Telling Sonnet 4.6 "this is code an agent produced" has a near decorative effect with no apparent significance, Sonnet 4.8 shows some misgivings, and when I experimented with Fable it seemed to do well at anticipating the kind of slop 4.6 would throw you.
Interestingly, to me, telling Fable that code a previous Fable agent wrote was AI generated seemed to raise some kind of "I'm being benchmarked" flag; expanding the reasoning finds it being evasive and mistrusting; look past the null derefence because this must be a trick question type thing.
In practical terms it often means that you can give the agent a ticket and it'll work on it for several hours until the PR is ready. Occasionally it'll run into some question that it needs you to clarify something or make some directional choice, but overall it's pretty autonomous.
Because this takes so long, you normally run 3 or 4 of those in parallel.
Then between shepherding those session, you also run several PR review sessions at the same time. Those also run for quite some time when doing deep investigations.
And then you also have long running discovery/design sessions for various other projects or problems you're working in between.
Assuming I'm working on one repo, I'll have different worktrees, each for a related area. For example, one worktree for each of: ui, small bug fixes, feature A, feature B, and so on.
Each worktree will have one active write agent, but I have a special docs/plans folder where I can have additional agents doing research and saving their findings. Agents in "plan mode" are write restricted to just that folder.
So look at a bug fix worktree. I could easily have 5 agents doing RCA into various bugs, each in plan mode. One at a time will get promoted to write mode to fix its specific bug while I continue discussion/RCA with the others. After fixing several bugs I'll spin up a "quality pass" agent that will make sure all tests/lint etc. pass and then give me a list of touched surfaces to manually verify before merging the branch and closing the worktree.
Note, I'm working solo at the moment so there's no PRs needed, but it would look quite similar if I had to make a PR for each bug fix, just probably with more worktrees.
A lot of this is personal taste but the general thing I get most value from is asking an agent to speculatively build every idea, instead of writing down ideas in some backlog for later (it never happens later).
> Popular with engineers from... (bunch of logos)
> Individual engineers, not company endorsements.
You might as well say Herdr runs on 3 billion devices, go all in! 16 billion devices! Every human on earth installed Herdr twice! (source: study amongst users of herdr, n=5)
I've seen VCs very much encourage this (and shadier) behavior. There's a line between hustling / being an entrepreneur and actively trying to mislead people... clearly we don't all see that line being in the same place (:
edit: to be clear, I didn't mean to make such a broad statement about VCs. I'm sure plenty of people do this on their own
There's the more traditional tmux/zellij way. I have a tailscale and mosh and a script where I can say something like "cw api do-this-feature" and it'll go into the folder of my api, create a worktree, and fire up claude code to use it (and another tab with lazygit and delta so I can do side-by-size diffs). And modified tmux's defaults so the session & window names are more targeted. Then can have it in Ghostty locally or Moshi (https://getmoshi.app/) on my phone. Moshi is basically a terminal for iOS but geared toward doing agentic stuff. Has a lot of shortcuts you can add for common commands of tmux, CC, can pinch to change font size, has a built-in diff tool, etc. Also has a small hook you can put in your agent so you get notifications on your phone. And it can directly open the particular tmux/zellij/herdr pane from that (has specific support for those three).
So that approach is nice in that you get full access to the harnesses, with no worries that a wrapper has a bug or doesn't support some new feature. But it still doesn't flow as well for me as a "real" wrapper does.
For the real wrapper approach, I started with Happy (https://happy.engineering/) and currently am using Happier (https://happier.dev/). Unlike Moshi, these are open source and basically control the various harnesses and have their own UI. Also have a relay so tailscale isn't necessarily needed, along with encryption in such a way that the person owning the relay can't decrypt the sessions. Happier even has a desktop app. Right now I'm running the dev version of Happier from source on my own machine w/ tailscale, with the testflight version of the app. There's various rough-around-the-edges aspects (and they're currently in the middle of a re-write), but it is nice to have a real UI with tabs for all your sessions, can click in to expand tool calls, to start a new session just select the harness, folder, type a new worktree name, etc and it starts going.
So I don't know. I like aspects of both of these and kind of depends on what I'm up to and my current mood. Nothing feels totally perfect yet IMO.
For me the killer feature is the git worktree management.
But in essence it is tmux on steroid (Specifically for agents)
Hopefully any of those get some traction
If you're a Mac user, and can forgive the shameless plug, you might find belfry.robgough.net useful. Connects to local and remote machines entirely through tmux, ssh and libghostty - populating the sidebar from tmux session info, and optionally adds a little Claude visibility in there for good measure.
There's even an iOS/iPadOS version – though for the moment you'll need to build that yourself. Source is all on Github.
Can you explain what you struggle to do in tmux with the mouse? I'm using it as we speak and everything is clickable, panes are resizable, right clicking a windows give you a context menu... Not sure what else is missing?
I generally run a session per project. Then within each session, a window for Claude, a window for running whatever dev server/logs, a window for neovim, and finally a plain terminal window for things like kicking off deploys etc.
Once I tried it, I can never go back. It is so simple and worked exactly as I thought how it should work.
I would say it is just a modern version of tmux but really thought about user experience of more novice users. For people who come from the mouse world, moving to this is quite seamless.
Also, coming to the comments, I would think maybe `zellij` would also work well. Or use `zmc` and build something on top. `zmc` is great, as the then the tab/window management can be handled by the desktop or TUI, depend on what someone want to use.
A few small downsides: I can't copy/paste in wezterm using the keyboard/vim keys because it is constantly drawing the screen and unselects my selection. The mouse drag in herdr works very well though. It'd also be nice if you could rebind key mappings in the UI, because I still haven't rebound the keys and am using the mouse.
So I will start a workspace for each different thing I'm working on, label it "Studio Shed Packet", "Teapot game", "Mux experiment". Then in each one I run a Claude Code. Then I can see the status of my Claude Codes just by glancing at the sidebar, rather than having to switch between screens to see what is waiting next.
I've been using it around a week so far.
I do think that there should be a protocol such that using these tools becomes more standardized.
For example I know that cmux is trying to support a tmux based setup so it feels like tmux could be most of that protocol.
There are a lot of tools like this popup and I really think making switching easier is the only way I will try a different tool.
With the long waits for agents to do stuff I really don't see how one can get anything done without multitasking with multiple worktrees in parallel. So I'd want support for listing the worktrees and then have a list of agents within each worktree.
Emdash and Nimbalyst have this kind of UI. Unfortunately both of them want to manage the state of each worktree group themselves; I'm looking for something that just would just call git worktree directly so that I can switch more seamlessly between CLI and IDE/TUI..
https://gist.github.com/iaindooley/cc8a61a1ff0fe23526c850906...
You include a script .worktree in your repo that does any copying or symlinking to setup the target directory.
It also has a headless mode so that it does the worktree operations without the tmux which I use for executing pi -p prompts in worktrees.
How do you manage that? How do you successfully navigate complex merges using ai?
But I do have many years of experience working in a larger team and it's the same problem there (just that people want to merge after some days of working). I'm not sure if AI changes the picture much vs working in a team.
Either way one has to plan ahead a bit and select tasks that are not going to trample on each other. I can typically imagine roughly what the code generated is going to be (at least what files will likely be involved in what way) and when selecting tasks to work on I take into account if it's going to likely cause conflicts.
In my experience if different branches work on different things, Claude have no issue doing "trivial" merges where you just ended up changing different aspects of the same lines. Of course, if two branches rewrite the same pieces of code there's a problem -- so don't do that..
echo “resolve conflicts” | runpi
Where runpi is my pi -p wrapper. I’ve never had a regression from it, but it gives me a report at the end so I can double check the decisions if I need to.
The skill is basically don’t use automatic resolvers, err on the side of including both sides, refer to recent commits, missions and runfiles for context and in your report to me use real branch names not HEAD and incoming because I can never remember what those refer to.
But I think git worktrees are a bit more ergonomic, I don't have to think about local vs upstream there's just one place to push.
I like to organize my projects like this:
myproject/.repo/git # bare repository .. my own convention..
myproject/main # worktrees from ../.repo/git
myproject/feature1
myproject/feature2So, I no longer use worktrees, and just copy and existing/clone a new folder.
I think all my problems go away with jujustu.
I prefer to manage my worktrees manually with a super simple script.
Has anyone got a tool/setup that is tmux-like but the remote terminals/panes are all local/native windows?
I do separation of concerns with the agent orchestrator (Nemesis8): https://github.com/deepbluedynamics/nemesis8. That can be run with or without Hyperia. I do not suggest anyone run agents on their bare metal. Putting them in a container gets a lot of wins, especially around log aggregation. Working now on a Splunk/Loggly-like interface for searching logs, tool runs (useful in tuning a custom local MoE drafter) and full session suspend, stop, detach, and search. It also does single MCP tool installs for all agents. Nemesis also supports dynamic port exposure to the host metal, for testing agent builds inside their containers.
Hyperia has a lot of extra features as well that I have found personally useful:
- Sticky notes (search too) - addressable panes in addressable tabs, tabs in windows, multiple windows - full ACLs across panes, notes, tabs, windows - Poke-a-pane to keep an agent going (any agent, not just CC which has a timer function) - webpanes with markdown extraction, JavaScript injection - directory pickers for people who find cd'ing to things confusing or those weary of typing nearly the same directory path over and over again in new terminals (not perfect, but I'm iterating on it) - a built in agent loop (in the Rust sidecar) that allows using local models for tool calls (needs a trained drafter to make it viable) or using a local model for token maxxing (compresses reads of panes by frontier models) - pane splits down/up/left/right and quick layouts.
As for whether it was "vibe coded" or not, or Herdr for that matter, I don't think that term is useful, other than for quick judgment. No, this is not a one-prompt project. I've spent 100s of hours on it, started out with Hyper, and did a crazy amount of planning on how to architect it. I have done systems architecture for a living before, and have a strong search background. People who hate on AI, and therfore projects done with AI, are threatened. Nothing more. That's why they shortcut with "AI slop" or "Vibecoded. Nope". That's just ignorance speaking from a standpoint of fear.
Slop, whether AI or human, is an effort problem: https://deepbluedynamics.com/blog/ai-slop-effort-problem. Looking at Herdr, it looks solid. Judge the product by it's outcomes, it's use, not whether or not AI wrote it or not. That's the moment we're in though, for now, so downvote or not. I don't care.
I couldn't click fast enough, then discovered that "Hyper Terminal" has nothing to do with "HyperTerminal" and couldn't click "close tab" fast enough.
How does one describe what's happening with stuff like this? Where a business tries to intercept people who are still learning the lay of some land, to get them to pay for something that they just haven't learned yet can be essentially free to them? Is there a word for it?
Doesn't tmux and zellij do all of these things that 'herdr' does?
It sets up a bunch of panes for you and tracks your agents' status and everything is clickable. You could probably do much of the same by setting up tmux appropriately. Or you could use this, which is already set up.
I tried herdr today, and it's not bad. Better than my previous many terminal tabs. Copy-and-paste is wonky (and it's ridiculous that copying is documented in multiple places in the docs, but pasting is never mentioned). And when an agent is waiting for a shell command, it shows as "idle", which IMO is wrong. Still, seems OK so far, I'll stick with it for a while.
I don't think you can avoid those much longer.
- everything is mouse clickable
- tmux style display-popups are used for friendly UI interactions everywhere
- it has a UI for agents running in panes, with a cool status (idle/working) display
- has opinionated defaults like automatic clipboard copy on mouse text select
- makes nested sessions easier & has default affordances for remote SSH attach
- is generally prettier
- uses display-popups for notifications
Otherwise it seems exactly like tmux.
for the lazy.
I work on both MacOS and Linux. Even if I were to figure out a perfect window management approach, I'd have to do it twice, and then figure out a way to "attach" to a layout of windows essentially. And then make it cross-platform... or live with zero organization, or at best some kind of flat structure. That's a immediate non-starter. Even if I focused only on a single OS, I still don't consider this a viable approach. It would be like removing browser tabs, and saying let's let the OS manage all this nasty UI stuff.
So then the next immediate idea is let's just have the Terminal itself manage UI. Ghostty supports tabs, splits etc. Great. But if I need to restart Ghostty for whatever reason, my entire layout vanishes and I have to set it all up again?
For me tmux is not just about the attach/detach part. A session is a personal "layout" of the workspace I've organically landed on for this task. Maybe it's a few windows with splits. Maybe it's just the single window with or without a split. Maybe it's a more browser-like just X amount of windows (tabs). If/when things get overwhelming I branch off into a new session for one (or more) sub-tasks. Little fzf-based utilities allow for nice fuzzying over sessions and what not. Or maybe you just run the different sessions in actual different physical OS windows. You can mix and match however you feel that day.
So if I'm in project x, working on feature y, and I detach and walk over to my other computer I can SSH over and continue as if nothing at all changed and retain the exact layout of what I was using before. Simultaneously in another physical window (or a tab) I can just switch to another local (or not) session and instantly be mentally in a different context, working on a different project or feature. Or I can SSH into a server and resume a debugging session left (intentionally) open for a while (which may consist of a few splits/windows/what not).
Maybe I'm missing something. I'm curious how your day-to-day workflow looks like. I imagine if I was 100% bound to a single machine, I could make it work with a tiling wm. But even then Ghostty crashes or something else happens and poof your entire work layout is gone and needs to be re-created (even if the individual pieces are safe).
(I know Wezterm has remote session multiplexing in beta, but I’d be surprised if you can re-attach from a different machine.)
Lately I’ve been experimenting with embracing the total scriptability of my favorite terminal, Kitty. Its creator is famously outspoken about non-native multiplexers, precisely because of the pitfalls you mentioned and because they hold back the terminal ecosystem.
Couple of recent issues I’ve had: washed-out colors in Claude Code, desktop notifications not passing through, scrollback weirdness, and image protocol mismatches.
I used to treat Zellij as an OS to build my own IDE in. Now it’s Kitty.
I save layouts as a Kitty session. Important limitation: this doesn’t persist running processes. And so enter zmx. Whenever I need a process I can reattach to I wrap it in zmx.
A terminal agent can interact with the kitty cli to create splits and such. I’m still on the fence as to whether that’s safe. Probably not. Also, tmux is likely better represented in the training data. Currently experimenting with running claude inside tmux (multiplexing) inside nono (sandbox).
Runs agents in tmux in sandboxes.
Gondolin is sand boxes done right...
So for a new worktree the agent can create a tab in zellij with multiple panes, properly named etc.
It works really nicely.
Although since I got the 52 inch Dell monitor I tend to just put everything visible at all once most of the time.
Until I try it, though, I don't think I'll understand why it needs a specific feature for running an agent tool in a side panel, which is appartently one of it's main selling point.
You end up having to wrap everything in adapter scripts because it is so verbose as well as so unmemorable, along with that, it has no built in help, not even -h
I never heard of this project but in 5 minutes I got the features that I wanted without having to introduce muscle memory. I just wanted a dead simple way to point and click to select different terminals. The only thing I neeed to remember now is "ctrl-b + v" to create a new pane.
Quite handy to apply since you can instruct models to run devservers in herdr panes or workspaces rather than doing long running shell commands or scripts in the agent session.
(You could probably do the same with tmux and others but I've had better success with monitoring herdr state)
But they seem to be targeting very different scopes use wise.
I think if you use this to do stuff like, edit files, run htop, drop into an interactive shell, compile some code, my guess is you're gonna be fighting uphill and have a bad time.
On the other hand, if you're only using tmux so that you can have a bunch of terminal windows for agents. Then this is probably going to make your workflow a fair bit smoother.
I use it primarily as a replacement for tmux to do terminal stuff, not use agents. It's awesome. It worked out of the box, has actually good, working mouse support, quickly ships fixes, and the defaults are sane with obvious behaviors represented in an intuitive TUI.
I think this is why it's described as "mouse first" terminal.
1- https://github.com/dcolinmorgan/herdr-remote Support Cloudflare, Telegram, .. quite buggy to me
2- https://github.com/0cv/herdr-mobile-relay Forked from the original one but improved and usable on phones, however the scope narrowed down to Cloudflare only
Similarity to tmux is the point
I've been dailying this (inside tmux!) for multi-agent orchestration and for state detection for agentic fleet orchestration, it's excellent. Highly recommend!