Multi-channel messaging (Telegram, Discord, WhatsApp etc)

[n0madic]

Thanks for the interesting project.
I'm interested in remote communication with the bot via messengers, primarily Telegram.

[yiwang]

what should LocalGPT really internally support?

[jcorbin]

Okay so I'll start out by sharing more about how I actually run localgpt, and other agentic harnesses like pi, claude, etc:

• my main Linux system is an amd strix halo rig able to run GLM-4.7-flash locally, so I lean into the "actually all local" aspect as hard as possible
• part of that is that I've hardened my system, e.g. so that any user's home is private, no shared world read into /home/other users
• then I just add additional low-privilege users to my system (no sudo access, no additional groups for access to hardware, etc)
• and grant my high-privilege normal user password-less sudo access and the ability to write into the low-side's home using classic unix group permissions
• for me this gives me plenty of containment, without feeling like I need to use a full VM, container, or even the landlock/seccomp support in localgpt ; also fwiw, when I tried to use the the sandboxing stuff, it crashed with a rust error of somde sort, so I just turnt it off and moved on... but separate issue

So basically my model is "low-priv alternate user runs each agent, high-priv 'Me' user has full access to just asymmetrically access them, but not the other way around".

How I then use localgpt:

1. I've written a small systemd user unit that spawns the localgpt server, which my high side user owns the build for, and installs Today's Binary into /usr/local/bin, but this way the AI user also cannot modify its harness binary, nor even has easy access to its source code
2. for interactive sessions beyond Telegram, I just sudo -u agent tmux attach and run localgpt chat inside its tmux; I could probably get away without tmux, and just sudo -H -u agent localgpt chat
3. it's also possible to try out the gui or the new 3d generator stuff by opening up and passing my high-user's wayland socket along, thus granting the low user ability to open windows in my normal graphical session ( biggest risk to doing this is clipboard access, but generally wayland's security model is Fine Actually for thing liek this )

So for me, I'd ideally like to see the localgpt server support its own unix domain socket for opening up local chat sessions much like the cli does today, but keeping all the tool calls and such on the other side like the current telegram sessions do. Nothing says "actually local" in current year than "TCP Socket? Nah. Unix (file) Socket." ;-)

Another use of this, is that it would allow to wholly separate things like telegram/discord/whatsapp network integration into wholly external processes, which could also be ran as a different user, such that tokens/credentials are never injected at any level into the AI user's namespace; this would be arguably better than the current popular trend of doing network layer proxy injection of secrets.

[yiwang]

@claude please analyze this thread and give advice based on best practice for this type of service