Self-hosted Cursor workers on multiple VMs — good setup or too much orchestration overhead?

I’m experimenting with a small multi-machine Cursor setup and would be interested in feedback from others using something similar.

The setup is simple:

  • one main PC as the “control” machine
  • one Windows VM for a large Delphi development environment
  • one Windows VM for PHP/web development
  • one self-hosted Cursor Worker on each VM

The idea is that the agents work directly where the real toolchains, libraries, project paths and local services already exist, instead of rebuilding everything on the main PC.

For focused tasks this works surprisingly well: code analysis, edits, builds, tests and even some deployment work can happen directly on the appropriate VM.

Where it gets difficult is orchestration across machines: Git/Gitea credentials, SSH keys, user contexts, permissions, deciding which machine should execute which step, and the amount of explicit control needed for more complex workflows.

My current impression is that self-hosted workers are very useful as specialized remote development environments, but using them as a fully automated multi-agent development pipeline quickly creates a lot of administrative overhead.

I’d be interested to know:

  • Does anyone use a similar “central PC + multiple specialized workers” setup?
  • Is there a cleaner way to coordinate several workers?
  • How do you handle Git credentials, SSH keys and permissions?
  • Do you let workers commit/push/deploy themselves, or keep those steps on the main machine?
  • Any architectural or security concerns with this model?

I’m not really looking for maximum automation — more for a practical setup where AI agents can work in the actual development environments without creating a second infrastructure just to manage the agents.

we ended up somewhere close to this. the workers are great when the agent just needs to sit where the toolchain already is, it’s the glue between machines that eats the time. what cut most of the credential mess for us was keeping merge and deploy on the main box only. each VM gets a scoped token that can push to its own branch and nothing else, so if a worker does something weird it can’t ship it