ChatGPT Remote plugin: allow configuring a custom self-hosted Remote MCP backend
Summary
Please allow the ChatGPT Remote Desktop Commander plugin/app to use a custom self-hosted Remote MCP backend instead of being tied only to the hosted Desktop Commander backend.
This would make the existing Remote Device OSS self-hosting capability usable from ordinary ChatGPT chats without requiring a separate Custom GPT.
Related to #664 (OpenAI / ChatGPT compatibility), but this request is specifically about backend selection/configuration in the ChatGPT plugin/app.
Existing capability
The Remote Device OSS already supports overriding the hosted backend with:
MCP_SERVER_URL=<custom backend>That path is already useful for self-hosting the Remote MCP/control plane while leaving the Desktop Commander tool engine unchanged on each endpoint.
I have validated a private self-hosted backend implementation against the existing Remote Device flow, including:
/api/mcp-infobootstrap;- device authorization flow (
/device/start,/device/poll); - persisted Supabase session;
- device presence;
- Realtime remote-call dispatch;
- real Desktop Commander tool execution on the selected device;
- explicit device routing with no silent fallback.
The Desktop Commander tool engine itself was not reimplemented; the existing OSS Remote Device continues to invoke the local Desktop Commander MCP.
Current limitation
The hosted ChatGPT Remote Desktop Commander plugin is available directly from normal ChatGPT chats, but its backend is not user-configurable.
A self-hosted backend can therefore work with Remote Device OSS, but cannot currently replace the hosted backend while preserving the normal ChatGPT plugin experience.
Using a separate Custom GPT + Action can prove the backend, but it does not solve the main UX requirement: using Desktop Commander from ordinary ChatGPT chats through the existing plugin/app.
Requested change
Please add a backend configuration mode to the ChatGPT Remote Desktop Commander plugin/app, for example:
Backend: Hosted / Custom
Remote MCP server URL: https://example.invalid/...
Authentication: <supported credential/configuration>The hosted service should remain the default. Custom mode should only change where the plugin sends Remote MCP/control-plane traffic; it should not redefine Desktop Commander tools.
Ideally the configuration should be stored per user/plugin connection and be removable/resettable back to the hosted default.
Security expectations
- Do not embed or publish a user's private backend URL or credentials.
- Do not silently fall back from a configured custom backend to the hosted backend.
- Keep device selection explicit.
- Preserve the existing Remote Device authentication/session model, or document any additional credential contract required by the plugin.
Acceptance criteria
- In ChatGPT, configure Remote Desktop Commander to use a custom Remote MCP backend.
- Connect a Remote Device OSS instance using the same backend.
- From a normal ChatGPT chat, list the devices exposed by that backend.
- Execute a normal Desktop Commander tool such as
start_processon one explicitly selected device. - Confirm that the request is handled by the custom backend, not the hosted Desktop Commander backend.
- Switching back to
Hostedrestores the current behavior.
This would let users keep the official ChatGPT integration while self-hosting only the Remote MCP/control-plane component, which the existing OSS agent architecture already largely supports.
Source: wonderwhy-er/DesktopCommanderMCP