SSH-Agent Forwarding and GNU Screen

GNU Screen and OpenSSH are two incredibly useful tools. In this tutorial, I hope to help users to work around a small, but annoying, problem that can arise when using SSH-Agent forwarding and Screen. This tutorial assumes you have knowledge of SSH, SSH-Agent, SSH-Agent-forwarding and GNU Screen. I also assume you’re using a shell that can handle Bourne shell variable syntax.

The Problem

SSH-Agent forwarding breaks when screen is re-attached on a different host.


The problem involves three hosts*, A, B, and S. All of these hosts have Agent-Forwarding enabled in their SSH configuration.

I log into host A and run ssh-agent, authenticating to my key. I connect from machine A to machine S using my agent/key to authenticate. On machine S, I start screen. I open several screen windows, all running shells. These shells inherit the ssh-agent variables from the parent shell that spawned the screen. (This is where the problem starts.)

I notice that it’s time for work, so I drive to work and log into machine B, starting ssh-agent and authenticate to my key. I connect from machine B to machine S using my agent/key to authenticate. I resume the running screen that I started when logged in from machine A. Naturally, all of the ssh environment variables are still set the way they were prior to the screen de/re-attach. So now, despite having ssh-agent forwarding turned on and having an active agent, I can’t connect to remote hosts via host S without some kind of magic.

*This problem could also involve two connections from the same host (which will be on different client ports,) to a second host. I think the above example is clearer.

SSH Environment Variables

When you connect, using ssh-agent for authentication, to a host that has ssh-agent forwarding enabled, the following shell variables will be set:

SSH_CLIENT= 4671 22

The SSH_CLIENT variable contains three space-separated pieces of information. The IP address of the client making the connection, the client’s port number, and the server’s port number.

The SSH_TTY variable is set to the path to the tty device used for the shell opened when the SSH connection was created.

The SSH_AUTH_SOCK variable identifies the path of a unix-domain socket used to communicate with the agent. In other words, this variable sets up a "path" back to the host that is actually running the agent, so that future ssh connections can talk to/use that agent.

The SSH_CONNECTION variable contains four space-separated pieces of information. The IP address of the client, the port number of the client, the IP address of the server and the port number of the
server. This variable is necessary to prevent agent hijacking.

So, now that I’ve reiterated the ssh(1) man page, we can see why, when we connect from A to S, and then from B to S, ssh-agent forwarding fails. In the initial connection, S defines the SSH environment variables to point to host A. When screen starts its shells, they inherit those variables. When host B connects and grabs that instance of screen, the sub-shells have no idea that anything has changed, so the variables now point to the "wrong" machine.

The Solution

The environment variables listed above are set correctly when host B connects to host S, so we need a way to get them propagated to the shells running inside screen. I wrote a small script to cull them out and stick them in a file. The script also re-formats them so that they will be easily integrated into the shell’s environment. I called this script "grabssh."


for x in ${SSHV} ; do
    (eval echo $x=\$$x) | sed  's/=/="/
                                s/^/export /'
done 1>$HOME/bin/fixssh

The output of this script looks something like this:

export SSH_CLIENT=" 4685 22"
export SSH_TTY="/dev/pts/22"
export SSH_AUTH_SOCK="/tmp/ssh-jtF17289/agent.17289"
export SSH_CONNECTION=" 4685 22"

Rather than having to type in "grabssh" before resuming any screen, I created an alias:

alias Attach='grabssh ; screen -d -R'

So now every time I resume my screen, the correct ssh-agent variables will be written to ${HOME}/bin/fixssh. I wrote another alias to source that file:

alias fixssh='source $HOME/bin/fixssh'

Now, when I’m in a screen-spawned shell and I want to take advantage of ssh-agent forwarding, I run "fixssh" and everything just works. You might even go so far as to create an alias for ssh like this:

alias ssh='fixssh; ssh'

However, that alias might someday drive you insane if you’re troubleshooting some ssh problem and you forget that it is there.

If you’re an EMACS user, you’ll likely find Mike Prentice’s addendum to this article helpful. Thanks, Mike!

2 thoughts on “SSH-Agent Forwarding and GNU Screen”

Leave a Reply