hckrnws

[delayed]

I'm so married to my existing tmux workflows and layout that I'm not sure whether I'd ever feel open to trying out something like this. At the same time, orchestrating multiple agents with native tmux and git worktree does feel cumbersome.

I am in the same boat. It's the unix philosophy. tmux does its job well enough and it is already scriptable. I don't think I was sticking to the same layout for a long time because different projects tend to require different layouts. But that's the fun part of streamlining down your automation. I haven't felt the need to explore other options yet because I haven't felt the limitation of tmux yet.

I don't consider tmux to be very unix-y. It does two different jobs -- connection persistence and window management. Most people these days only use tmux for the former now that tabbed terminal programs are commonplace, but the complexity of the latter still infects tmux.

I built a tool that lets you keep your existing tmux workflow and adds a simple dashboard for monitoring and discovering agents inside tmux: https://github.com/sethdeckard/atria

if you want you can file an issue with your current workflow? happy to see if we can do a PR to support this natively in tmux-ide

What tools do you use on tmux to achieve your workflow ? Do you do just prompting or do you code too ?

tmux by itself lets you create any number of sessions, windows and panes. You can arrange them for anything you want to do.

Having a pane dedicated to some LLM prompt split side by side with your code editor doesn't require additional tools, it's just a tmux hotkey to split a pane.

There's also plugins like tmux resurrect that lets you save and restore everything, including across reboots. I've been using this set up for like 6-7 years, here's a video from ~5 years ago but it still applies today https://www.youtube.com/watch?v=sMbuGf2g7gc&t=315s. I like this approach because you can use tmux normally, there's no layout config file you need to define.

It lets me switch between projects in 2 seconds and everything I need is immediately available.

Oh I mixed up tmux and termux.. woops.

I'm also trying to build something similar for agent orchestration where one terminal is controlling multiple terminals. I tried using tmux but it's very good at sending the initial text to the tmux sessions, but I've not been able to get an agent to have a proper back and forth controlling multiple tmux sessions. I know we can use send-keys, but reading the session or knowing when that session is complete is kind of up in the air. And then if the main orchestrator terminal has checked all the sessions to see if they're actually working and doing things, the main session kind of stop so I've kind of been thinking about a cron that periodically checks in and nudges it to check the sessions again. Are they still working? Do they need more guidance? Essentially having one terminal control others, but having that back and forth with the terminals has been pretty challenging to achieve. Have you gotten anywhere with this?

It sounds like maybe you haven't seen agent teams mode, which this project is using.

https://code.claude.com/docs/en/agent-teams

Oh will definitely look into this! I ditched CC for Pi a few months ago so maybe there’s an equivalent feature.

yup core to this project!

I recently built something just like this: https://github.com/sethdeckard/atria

It supports tmux, but you can use it without via embedded terminal. It also has native integration with a few select terminals that expose the right kind of APIs.

Installs as single binary (written in Go) with no external dependencies.

If this supported Gemini and Codex I would find it useful. I never run more than one Claude, it's always a mix.

Both are supported! You can customize the startup command in ide.yml. I will update the docs to highlight that :)

Looks like a great implementation. I want to question the basic user story, which seems to be: "I am a software developer who wants to improve productivity by running multiple simultaneous agents that are roughly isomorphic to a human software developer team."

I am burning a lot of tokens every day at work and on personal projects. It's helpful. I generally work in tmux with github copilot in one pane, and a few other terminal panes showing tests and current diff.

I find it really important to avoid the temptation to multi-task by running multiple agents. For quite varied tasks, productivity gains from multi-tasking have proven to be illusory. Why would it be different with writing software?

https://en.wikipedia.org/wiki/Human_multitasking

https://github.com/ehrlich-b/wingthing Here’s my take on this idea, also FOSS. Does tmux style sessions so you can come and go. It also exposes a web terminal so you can get remote access - but you can also run fully locally for less latency.

Interesting, but I wonder if this shifts too much complexity onto the user.

tmux is powerful, but not exactly approachable, and "multi-agent orchestration" on top of it feels like something that could get hard to reason about quickly. Curious how you think about UX here.

Good points and indeed thinking about this quite a bit. Currently leaning towards a CLI first approach so that Claude/Cursor/[insert coding agent] can configure and control the ide. Feels a bit meta, but also makes it extremely user-friendly.

Tmux is pretty easy to pick up and build muscle memory by learning a few keyboard shortcuts from a basic youtube video and it's handy when you don't want to switch screens between multiple terminals just for one thing.

The ability to split and divide the screen pretty simply with a few keys is handy for anyone who spends enough time in a shell - the abilty to save that layout for the shell items you're using to load up easily again the next time is valuable too.

Multi agent orchestration likely just means keeping track of a few different windows all on one screen.

https://cmux.com/

I love cmux! ironically you can use tmux-ide within cmux. The idea is to make it an agent development environment that's great when ran on a remote machine :).

I built a very similar one that I use every day, smux: https://github.com/gergomiklos/smux. Took only 1 hour with claude.

Terminal-first makes a lot of sense for anything that runs on remote servers. I've been on helix+tmux for about a year and the main friction is onboarding teammates who are VSCode-native. Nice to see projects pushing in this direction. Does it handle multi-pane debugging or is that still a manual tmux split?

I like a good CLI, I like a good TUI, but why are we eschewing high-DPI graphical displays, icons, true-color, images, and everything else in favour of a 1980s terminal? Is it just to be retro, or do you guys really think you're seeing some advantage restricting yourself like this? Do you all REALLY hate moving your hand over to your mouse?

Congrats on getting this out. What was the most surprising part of the build?

most surprising was that something this lightweight made such a big impact on my productivity. its really nice to have persistent Claude teams on my remote machines that I can always access no matter what.

Can somebody develop a mobile app that natively supports tmux

I assume that you've tried Termux and somehow that doesn't meet your needs? (Also, you didn't specify whether you are on Android/iOS)

both ish and termux are great options on mobile/iPad!

For all the hype of AI agents, you never see people taking on real challenging projects like this. Just low hanging fruit.

solid work. would love to see a comparison with the latest claude/gpt models

So basically tmuxinator?

Thank you for sharing your work. Really cool.

From my perspective, this is cool, but since tmux is kind of permanent, you open your layout, set 1,2,3 screens for agents, you might add gemini and opencode. then open vite for server and one for shell for example. Then you can just close it and reopen whenever you want to work on it.

And that is it. If I am missing something, processes taking memory or such, I have a machine with memory (I know, flexing how expensive things are), please explain.

[dead]

[dead]

the token costs are real. we switched to smaller models for 80% of tasks and barely noticed

[flagged]

[flagged]