AI Workbench cannot connect to DGX Spark

AI & Data Science NVIDIA AI Workbench
1

If you are reporting a bug or error, consider submitting a Support Bundle to aiworkbench-ea@nvidia.com. This will help us solve your issue more quickly.

Please describe your issue or request: (tick the boxes after creating this topic):

Please tick the appropriate box to help us categorize your post
Bug or Error
Feature Request
Documentation Issue
Other

Setup:

  • Windows 11, AI Workbench Desktop App
  • DGX Spark with AI Workbench CLI installed
  • Connected via Tailscale
  • Docker Desktop running with WSL 2 integration

The Problem:

I’m trying to add my DGX Spark as a remote location in AI Workbench but keep getting this error:

{
  "error": "Failed to get context SSH fingerprints.",
  "detail": "cannot connect to ssh-agent: problem connecting to OpenSSH Authentication Agent, ensure the service is running"
}

Testing my ssh setup everything seems fine outside the workbench app.

  • SSH works fine manually - I can ssh into the Spark from both PowerShell and WSL without a password
  • The SSH agent IS running in WSL - ssh-add -l shows my key loaded
  • I made it persistent in both .bashrc and .profile
  • Even non-interactive shells work: wsl -u workbench -d NVIDIA-Workbench -- ssh-add -l returns my key
  • The AI Workbench service is running (curl localhost:50001/v1/version works)
  • Docker is working

What I’ve tried:

  • Both Windows paths (C:\Users\...) and WSL paths for the SSH key
  • Using the public key .pub file and the private key file
  • Starting the Windows OpenSSH Authentication Agent service
  • Fresh SSH keys created specifically for AI Workbench
  • Using Tailscale hostname, IP address, everything
  • The CLI command nvwb-cli create context runs without error but doesn’t actually create anything

The frustrating part:

NVIDIA Sync connects to the same Spark just fine using its own SSH keys. But AI Workbench’s GUI can’t find the SSH agent even though I’ve verified it’s running and accessible in the exact type of WSL session the GUI uses.

It feels like there’s some environment variable not getting passed or the GUI is looking in the wrong place for the agent.

Has anyone else encountered this issue before? Am I missing something in setting up the app?

Thanks for any help!

2

An update to this, I was able to get the workbench gui to communicate with my dgx spark by manually creating the ssh context within WSL. That solved that issue and I can now use workbench with my spark. I do still have an issue of having nvidia sync being unable to work with workbench for now but I will look into it as well. Admittedly it is disappointing the software suite doesn’t work together seemlessly.

3

I believe the issue is that you are running the SSH agent in WSL. This isn’t supported by Workbench. On Windows installs it uses the Windows OpenSSH Authentication Agent service, a port of the OpenSSH SSH agent. From Windows you can use ssh-add or ssh-add.exe to communicate with the Windows SSH agent service, after it has been started.

Just like you describe, if you add the protected private key to the Windows SSH agent and select that key’s public key when creating the Workbench context, Workbench should be able to verify that the private key is held by the SSH agent.

4

I attempted to setup the SSH agent on windows at first but it wasn’t being recognized. Setting up the SSH agent in WSL was my workaround solution afterwards.