Martin Zajíc

building self-hosted home for my codex, claude code, pi and muse sessions across pc, laptops and vps. dev at @keboola, PHP, Linux.

[Martin Zajíc @zajca]

i run codex, claude code, hermes, pi and muse sessions on my pc, old laptops and vps. one manager agent starts them and watches which one is working or waiting, so i only check that one. it's open source and self-hosted: https://t.co/fYPNjWaIw7

[Martin Zajíc @zajca]

@bentlegen @exedev I went other way around build https://t.co/fYPNjWaaGz to run agents across machines and watch their state, next are plugins for runtime like kvm/microvm/...

[Martin Zajíc @zajca]

@aarondelasy had same problem, that's why sessions on my vps run under https://t.co/fYPNjWaaGz daemon and not in my ssh. network drops, laptop sleeps, sessions keep going and I just attach again. no tmux needed. do you start them over plain ssh?

[Martin Zajíc @zajca]

similar here but I didn't want to build own dashboard. I run claude, codex and pi sessions on my pc, old laptops and vps through https://t.co/W5YcLPluMG and one manager claude session starts them and watches which one is working or waiting. it reports me few times a day so I'm not the one tracking all sessions. how do your duos decide when they really need you?

[Martin Zajíc @zajca]

This bit complicated best to ask agent on code, since Pohunek supports multiple, so each is different. OSC 7501 is main source, second is screen reading and fallback is if bytes flowing from PTY. This is combined and debounced. Pohunek also records subagents and background tasks, but they currently have no effect on session state.

[Martin Zajíc @zajca]

i run codex, claude code, hermes, pi and muse sessions on my pc, old laptops and vps. one manager agent starts them and watches which one is working or waiting, so i only check that one. it's open source and self-hosted: https://t.co/fYPNjWaIw7

[Martin Zajíc @zajca]

did similar but without hooks. my manager is normal claude session which starts claude/codex/pi sessions on my machines through https://t.co/W5YcLPkWX8 and sees which one is working or waiting for input. Next is MCP to register in Grok bot as I don't trust Elon that much to connection my VPN to Grok.

[Martin Zajíc @zajca]

@prathamgrv I don't have 256GB mac and I can't train any models. But having 20 and more parallel agents is no problem. I mostly use manager which is dividing agents work across all my devices. (finally good excuse not to sell old laptops)

[Martin Zajíc @zajca]

I think it's still more complicated than. But I was on vacation last week, I did prepare work for agents on 3 different projects. Manager agents were reporting 3 times a day what was done and what is the plan. It worked really well I was checking very occasionally. I actually had to slow done agent on one project because it was overwhelming rest of the team.

[Martin Zajíc @zajca]

@Tony_Stef_ had same problem so i dont use routines for this. my manager is normal claude session on my machine and it starts full claude/codex sessions by itself through https://t.co/fYPNjWaIw7.

[Martin Zajíc @zajca]

@kaelig Same for me, manager can do lot of things. For example manager for pohunek project is running for weeks independent, reaching me only via matrix chat and preping artifacts for me to decide. My biggest work is to plan features with agent and tell what is in next milestone.

[Martin Zajíc @zajca]

@k4k0ch Pohunek works cross devices, so prompt is only way now. But agent is aware of host so can use file. I was thinking about some way to share files and configs, but trying to not over engineer things. Same with some agents communication.

[Martin Zajíc @zajca]

@SSShken @AhmedReza Thing is you don't want to watch agents, but rather warch one manager session and this session will watch agents over shoulder. This is why I have build https://t.co/fYPNjWaaGz multiplatform and headless, I'm running my agents across 4 devices including mac and watch over manager

[Martin Zajíc @zajca]

i have claude as manager agent and it gives work to codex for review and second opinion, sometimes i swap codex for pi with openrouter models. all sessions run on my pc, laptop and vps with https://t.co/W5YcLPkWX8 so manager can start them anywhere. limits i still watch manually tbh. And didn't use cursor for more than year

[Martin Zajíc @zajca]

@k4k0ch yeah that's more planning problem than merge problem. I have one manager agent for this, it plans first what is shared, lets one session build the log, and only then starts the 3 features in parallel on top of it

[Martin Zajíc @zajca]

I'm using codex for reviews and second opinion on problems. But in my workflows I can swap codex for Claude or Pi with models from openrouter. Recently I was doing lot of coding with meta models in Pi. I just instruct Claude which I'm using as manager to give some tasks different model and evaluate their work. It works surprisingly good.

[Martin Zajíc @zajca]

@GripFangWolf I was on tmux and ssh too, now my machines at home run https://t.co/fYPNjWaaGz and sessions keep running there, from laptop I see all of them. maybe worth trying when your limit resets?

[Martin Zajíc @zajca]

@hqmank same for me, I stopped switching between sessions. I run claude code and codex on my own machines with https://t.co/fYPNjWaaGz and mostly talk only to one manager agent, it starts the others and I just check which session is waiting for me