Remote Desktop MCP (DV)

简体中文 · Setup · Security · Review

Experimental v0.2: upstream transitive dependency advisories remain. See the dependency review before deployment. Connect an AI assistant to your development machine through Desktop Commander and the official OpenAI tunnel-client, using a persistent stdio process. This project packages the integration used by DV; it is not a replacement implementation of Desktop Commander or a desktop screen-control application.

Features

• Files, terminal sessions, search and document tools provided by Desktop Commander.

• A small stdio adapter that filters info/debug MCP log notifications and non-protocol stdout while preserving responses, errors, progress and higher-level notifications.

• Upstream process failure propagates to the supervisor; no shell pipeline or jq dependency.

• Private TOML configuration, loopback health endpoint and generated tunnel YAML with an environment reference for credentials.

• A read-only smoke check that initializes MCP, discovers tools and calls list_sessions without displaying local data.

• Linux/WSL2 persistent tunnel deployment; Windows local stdio support. CI checks both operating systems.

Related MCP server: GPT Web Codex

One-command setup (Linux / WSL2)

curl -fsSL https://raw.githubusercontent.com/ThinkerQAQ/remote-desktop-mcp/v0.2.0/install.sh | bash

Already downloaded/cloned the repository? Run bash setup.sh. Windows PowerShell: ./setup.ps1 (requires initialized WSL2 Ubuntu; copies source to the Linux filesystem).

The wizard installs missing Node.js/tunnel-client from official distributions with SHA-256 verification, runs locked npm installation and a real read-only MCP smoke test, asks for your own Tunnel ID and hidden runtime API key, saves private configuration, starts a persistent process and checks tunnel readiness. Re-running reuses saved inputs; port conflicts select a free loopback port. No global Node or existing DV tunnel is replaced.

You must create your own tunnel/runtime key and register the custom MCP in ChatGPT. The wizard prints the official account links and next steps. Registration and remote tools/call remain pending until you complete them; local/tunnel success is not presented as full remote acceptance. No Edge debugging setup is needed for this file/terminal integration.

See the full wizard guide for repeat runs, headless deployment, recovery, credentials and limitations. Source download requires curl/tar; Ubuntu/WSL can install remaining base utilities via apt (sudo may prompt). Other distributions must provide those utilities. Dependencies are installed under the checkout; upstream package install hooks may run. Upstream dependency advisories remain as disclosed above.

For development: npm ci, npm run check, npm run smoke. For local MCP clients, execute node /absolute/path/scripts/proxy.mjs directly. The official tunnel-client is installed separately by setup; no copy is bundled in Git.

Scope and limits

DV and Remote Desktop Commander overlap in file and command capabilities. DV uses your own tunnel-client deployment; Remote Desktop Commander provides its own remote service and device selection. This project adds reproducible deployment and transport checks, not a new tool engine or an autonomous agent.

This adapter is not a sandbox or an authorization layer. It exposes upstream tools, including shell execution. Desktop Commander settings are not a substitute for OS isolation. Run under an account with only the access you intend to delegate. Requests and results sent through the tunnel can reach your AI service; “local execution” does not mean “local-only data.” Upstream telemetry/settings remain upstream behavior; this adapter does not disable them.

Persistent processes reduce reconnects; they cannot remove platform authorization requirements. Terminal execution does not activate a desktop window, but commands that launch GUI programs can affect focus. Native Windows tunnel service management and automatic updates are outside v0.2.

Upstream and license

Integration code: Apache-2.0. Desktop Commander: MIT, pinned to 0.3.0. OpenAI tunnel-client: Apache-2.0, separately installed; tested with 0.0.16. See third-party notices. This is an independent project, not an official OpenAI or Desktop Commander product.