Then there are the cases where you want the LLM to actually interact with the page, using the current web page state and your credentials.

For example, one might want to tell it to uncheck all the “opt in” checkboxes in the page… And express this task in plain English language.

Many useful interactive agent tasks could be achieved with this. The chatbot would be merely the first step.

Maybe consider static ip assignment in your DHCP server (e.g. internet router) if at all possible… Then you can add a name to it to /etc/hosts.

Alternatively you could use Avahi to provide mdns names to your local network

I guess it’s worth checking if those names point to the expected binaries, but I also think it would be highly unlikely they would be anything else than just /usr/bin/ssh and /usr/bin/ssh-agent.

As mentioned, -v (or -vv) helps to analyze the situation.

My theory is that you already have something providing ssh agent service, but that process is somehow stuck, and when ssh tries to connect it, it doesn’t respond to the connect, or it accepts the connection but doesn’t actually interact with ssh. Quite possibly ssh doesn’t have a timeout for interacting with ssh-agent.

Using eval $(ssh-agent -s) starts a new ssh agent and replaces the environment variables in question with the new ones, therefore avoiding the use of the stuck process.

If this is the actual problem here, then before running the eval, echo $SSH_AUTH_SOCK would show the path of the existing ssh agent socket. If this is the case, then you can use lsof $SSH_AUTH_SOCK to see what that process is. Quite possibly it’s provided by gnome-keyring-daemon if you’re running Gnome. As to why that process would not be working I don’t have ideas.

Another way to analyze the problem is strace -o logfile -f ssh .. and then check out what is at the end of the logfile. If the theory applies, then it would likely be a connect call for the ssh-agent.