HA docker/ssh/modal etc feedback thread

[hollow shore]

heya <@&1478898669712310353>! i'm working on a refactor to better unify how hermes agent works with different execution environments. right now we support local, ssh, docker, modal, daytona, singularity. if you have any issues, suggestions, general feedback (+ve / -ve) for any of them, shoot em over here

first order of business is to make better interfaces so that we have less repetitive code across the environments

[limpid barn]

"paste install curl -sSL install-hermes.sh"
"hermes --code HAKE-ME23"
"... you are connected to claude-code on machine X, on codex on machine Y, and models W, Y on gateway BANANAPOKE.localhost"

headscale would be nice 😄

[hollow shore]

should be even more easy to add post this 😄

HA docker/ssh/modal etc feedback thread

[glacial sky]

i am tired of having to ssh to use tui. when will we get an official full-featured tui-like webui? that we could still access via tailscale for example. but something that just looks and works better than discord/telegram on phone. with every tui feature

[weary kayak]

A refactor makes sense given that I would like to see Podman added. It has enough similarities with Docker that some of the helper functions should probably be put in a utility module.

[mint acorn]

you know what I've been thinking would be killer, is if it could manage multiple terminal environments and run things in parallel across them, rather than being constrained to one. Both multiple of the same kind + mix/match. So maybe it has a local or a docker backend as our "home", but it can spin up and manage multiple sandboxes on Daytona/Modal. Or any other combination. I think with the right interfaces it wouldn't be too hard to do

That would be a big step towards turning it into an orchestrator of agents imo, especially w subagents already here

[placid sparrow]

For Hermes of course lol

I've been thinking about a tool routing layer where in a thin client setup (local CLI, remote agent), tool calls get dispatched either client-side or server-side based on config. WebSocket relay in the API server... Is that inline? 😯👀

oops that sent backwards

So my server can host hermes while I use cli locally on windows - tool access on windows,etc

[hollow shore]

do u mean u want the TUI running locally and the agent wherever u put it?

that's alr being cooked up by @winter whale / @fathom sundial :)))
they're working on decoupling the agent and the UI.

so the UI can be webUI/or own vibecoded UI/TUI

[glacial sky]

yeahhh

great epic

i love how the tui works, but its only usable via pc, thats the thing

i wish we had something ideal for phone

[hollow shore]

i wanna make it really easy to add in more environments for sure!

[hollow shore]

can u explain better?
😅 dont think i understand

[placid sparrow]

Sorry XD

[hollow shore]

is this same as the decoupling?

[placid sparrow]

I believe we were on the same scope/idea

[mint acorn]

what would you wanna see on mobile? I've been working on a mobile UI for myself + gonna share with everyone here once I feel like it's good

can dm or @ me in #hermes-agent

[acoustic zephyr]

One thing I do a lot of handoff from the TUI to Telegram, but they don’t have knowledge of each other’s conversation. It’s a pretty decent friction point for my usage. I’ve started doing more tmux usage so I can continue a TUI session on my phone

[placid sparrow]

I host hermes on my ubuntu server, along with the rest of the home-lab, but thought it would be nice if herme's could exacute tool calls etc on the local machine rather than only server side. Memory, skills, model management, config, etc - the rest of the system remaining server side centrally. Just to add - but I think yeah same ideas basically 😛

[fallen idol]

+1 for this

[mint acorn]

I think if you name your sessions with /title you can then try to /resume across devices? I haven't tried it but I don't see why it wouldn't work

[glacial sky]

yeah i agree on this one, continuing sessions on different endpoints is a mess in hermes, a giant friction point imo

[glacial sky]

yeah you can, but sometimes it just doesnt pick it up so well

[mint acorn]

I see

[glacial sky]

ive tried that and it did work tho

[stiff comet]

I think if we get a good websocket/json-rpc interface that's well defined to all the Hermes functions

the web interfaces will build themselves.

[unique raft]

I would be happy if I could run an Hermes agent on a machine in an isolated network, and whenever the agent needs to perform an action, it could connect via a secure channel to a machine with powerful CPU and memory resources and execute tasks within a containerized environment.

[acoustic zephyr]

Yeah I’ve been waiting for that one instead of building my own

[stiff comet]

Seems like an obvious solution. Afterwards feeding the spec to hermes and asking him tio build a web ui

[acoustic zephyr]

You can already do that with the SSH back end, right?

[stiff comet]

is like 10 mins fo work

[acoustic zephyr]

The PR for it has been open for a while

https://github.com/NousResearch/hermes-agent/issues/360
which is blocking for https://github.com/NousResearch/hermes-agent/issues/501

[stiff comet]

I have seen that looks pretty complete

[hollow shore]

yea

[stiff comet]

that would also streamline the gateway and cli better too

[sleek hull]

@solar spearbin Please consider implementing the Agent Protocol to enable the "blue band" notification UI in cmux, similar to Claude Code. This visual signal is essential for knowing when the agent needs approval while running in the background. Native macOS notification support would give Hermes a significant edge over OpenCode/KiloCode for power users.

Also please note that I am building: https://discord.com/channels/1053877538025386074/1486768386926055635

This is uniquely challenging given Apple containers(MacOS 26 beta) do not support inter-container networking. I am implementating container-compose for declarative vSock integrations which facilitates air-gapped hardware isolated containers which I believe is an ideal deployment model for long running Hermes and Honcho Hub/DB containers which are taking in high privacy information and conducting experimental code execution. e.g. hyperagents.

[glossy wraith]

i'll be happy with anything that makes it much easier to work with my hermes remotely, since it runs in a VM not local 🤣 makes CLI and workspace folders etc kind of difficult but i get by - not quite docker/etc but would be nice if i had a way to just hermes on my pc and have the remote instance know what's up, i doubt overlay file systems are good for that i'm not an expert in any of this specifically

[hollow shore]

does something like the ssh backend work for u where ur VM can be wherever/whatever?

https://hermes-agent.nousresearch.com/docs/user-guide/configuration#ssh-backend

[hollow shore]

what happens/doesnt happen properly? i'll try it rn as well

[glacial sky]

for me they often behaved like we started a new convo, but with the memory of the last session. not a continuation of the session - might be me tho

[hollow shore]

ur saying u would like to have ur remote hermes be able to access ur local dirs?

or u want ur hermes to run locally but work in a VM.

to be fair we have also started architecting a hermes agent that runs on the cloud as well as ur desktop seamlessly.

[hollow shore]

gotcha. thanks

[glacial sky]

hermes is running on a vps and i havent even thought of him running locally. what i mean is that I, myself want to be able to access the local directories kinda like an FTP tab in an official hermes phone UI app

so i would never have to actually ssh into him from my pc again lol

[hollow shore]

fair, i face that too. rn i just ask it to send me the docs i wanna read over telegram/discord.

but this is something that will be architected and fixed on a wider product level!

[acoustic zephyr]

I just use Transmit on Mac to SFTP to the box my Hermes is running on. It can sync dirs iirc.

[sleek hull]

I use Hermes agent on Nvidia Jetson Orins for controlling robots and doing research.

Agent Protocol Support: Emit "attention" signals (JSON-RPC or ANSI) to trigger the cmux blue-band notification on macOS via tmux 3.6a allow-passthrough. This is a massive UX win for backgrounded remote tasks.

Persistent SSH Sessions: Allow the ssh-backend to attach to existing multiplexer sessions (e.g., tmux 3.6a attach) instead of spawning new SSH processes per tool-call. This would drastically cut latency over tunnels.

[stiff comet]

I sitll think this is the answer for literally 95% of all remote access solutions. #1491227373692129530 message

[hollow shore]

https://github.com/NousResearch/hermes-agent/pull/4561

btw, draft draft draft PR here.

@sleek hull re: speedups it has a lot of speedups in execution in SSH mode.

gonna try and optimise other modes too

[languid cedar]

I'm not sure if this would have anything to do with the refactor your working on, but I found that my LLM got pretty confused when I was using docker terminal environment. It thought that terminal=local, and execute_code=docker, so it was trying to touch a file with python and then confirm it was visible locally with terminal commands, which of course worked fine, but it was really doing everything in the sandbox

[hollow shore]

this has a lot to do w/ the refactor 😄

[hollow shore]

can u explain this better? u had the docker backend?

[broken tiger]

These are excellent suggestions.

[hollow shore]

but it was using the local terminal?

[languid cedar]

yeah, so hermes is running in a VM (but not in docker itself), and I had configured terminal.backend=docker. I was asking it to write some python code to create/update a JSON file of websites I wanted it to save. but that json file was ending up in the sandbox, until I figured out how to map the right volumes.

[broken tiger]

Having a kubectl-like tool that allows an orchestrator to gain functionality over a swarm of agents for issuing frequently ran scripts over datasets, or deterministic tasks, instead of dynamically inferring them.

There are pros and cons to this. Sort of kneecaps trajectory building. But practically very useful!

[hollow shore]

i think orchestrators are a fake thing

everyone can just vibe code their own based on their own setup

[hollow shore]

hate to be that guy but have u looked at the mount setup here?

https://hermes-agent.nousresearch.com/docs/user-guide/configuration#docker-volume-mounts

[languid cedar]

I wasn't sure if those mounts would be relevant, for whatever reason the agent thought a persistent file should live in .hermes/data, so I setup a mount for that. But it seems like some of the other tools weren't working in docker, so I've since gone back to just the default terminal=local.

[broken tiger]

Yeah, they sort of are. But you’re mainly packing on weight from claw users running various models in various configurations. “Vibe code it” leads to various results. A million of Theseus ships pull into harbor for repairs on some routinely helpful utilities.

[sleek hull]

@Myrddin Exactly. I’m currently running a K3s cluster on each of the Jetsons, using Hermes Agent as a sidecar pod to drive a DAG of ROS nodes and infra.

The goal is to use Hermes to iterate on experiment parameters while Honcho captures the memory of which combinations yield new capabilities. It creates a closed-loop where Hermes builds the trajectory and Honcho synthesizes the long-term "lessons learned" from the hardware states. The edge robots integrate with Qwen 3.5 VLM to advance visual causal reasoning.

[hollow shore]

dude i lowkey feel like ur using a clanker to send messages

[broken tiger]

Yeah, we’re cooking similar stews, my friend

[earnest sorrel]

Provide a shell tool that ssh's into your server before executing the command

[sleek hull]

sorry trying to be concise.

will stay human only.

[hollow shore]

im using my hands to type 😄
least ya could do is do the same

[sleek hull]

Totally legit

[hollow shore]

fair! something that should get fixed in the refactor

[broken tiger]

Well anyways, standardize a bridge layer for multi tenant/multi platform

If that’s not an interesting consideration, I guess I just don’t know what you’re asking

[mellow silo]

I second podman. Lacks root by default vs docker.

[hollow shore]

https://github.com/NousResearch/hermes-agent/pull/6085

some fixes merged onto here 😄

[hollow shore]

https://github.com/NousResearch/hermes-agent/pull/6283

and another PR opened

[oblique wing]

why would u ever want to use ai agents to deal with repeating deterministic tasks?
also there already is delegate_task, it just has execute_code hardcoded to disabled atm

[broken tiger]

When you have values that can be known that the agent might need as context for something else, but they can run a script instead of making an approximation, or writing new novel scripts every time it needs to know something?

[oblique wing]

for a) the main ai agent can simply execute or queue the scripts to run in the background and get the results when they are done. there is nothing you gain from invoking several subagents
b) is not really a repeating deterministic task

[broken tiger]

Oh, ok

[sleek hull]

Thanks for getting this merged into main. Looking forward to using it in the next release.

[limpid barn]

@hollow shore et all, you people aware of https://github.com/strongdm/leash ?