Hey, thanks for the detailed report. You laid everything out really well, down to cache locations and workarounds.
This is a known limitation. Agents Window (Glass) doesn’t correctly resolve the MCP config for remote workspaces yet (WSL and SSH). Editor Window delegates resolution to the remote extension host correctly, but in some cases Agents Window pulls state from the local Windows host. Because of that, after switching, the per-workspace MCP binding goes stale. That matches your symptom. Settings shows Connected, but tool calls fail silently.
Your main workaround, a project-level .cursor/mcp.json in the repo, is the right mitigation for now. It forces workspace-scoped registration on the WSL host where the binary actually lives, and both windows handle that config more consistently than the user-level one.
Related threads on the same topic (you can subscribe for updates):
- User MCP Servers in Agents Window are inconsistent with Editor Window when using Remote SSH
- Agents Window uses 'Windows' .cursor settings, where it should use WSL:ubuntu-24.04
- WSL workspace context lost when switching to Agents Window (still present in 3.1)
I’ve reported the issue internally. I can’t share an exact ETA for a fix yet. I also passed along your feature requests, a warning on switch, a one-click state reset, and an option to pin MCP to the remote host. Let me know if the project-level config stops helping anywhere.