Practical setup Client Setup
← Back to project top Copy, paste, try

Connect desktop-touch-mcp to the client you already use

This page keeps the setup practical. Pick your client, copy one snippet, and you should be able to get moving without reverse-engineering transport names or config shapes.

Current scope: all examples on this page assume a local Windows 11 machine. This project is not yet a multi-OS setup story.

Start here

Choose the transport first.

Before transport, one important scope note: the runtime and client setup on this page are documented as Windows 11 only for now.

Stdio is the simplest path when your client can launch local MCP servers itself.

npx -y @harusame64/desktop-touch-mcp

HTTP is the right path when your client wants a server URL.

npx -y @harusame64/desktop-touch-mcp --http

The default local MCP endpoint is:

http://127.0.0.1:23847/mcp

Optional health check:

http://127.0.0.1:23847/health

Custom port example:

npx -y @harusame64/desktop-touch-mcp --http --port 8080
Rule of thumb

If the client can launch, use stdio. If it asks for a URL, use HTTP.

Use stdio

Claude Code, Codex, Gemini CLI, and many local clients are comfortable launching a local command.

Use HTTP

VS Code, URL-based clients, and ChatGPT Developer mode fit better with a local endpoint.

Keep both

It is reasonable for public docs to show stdio first and HTTP second for the same project.

Tested shape

These setup patterns are based on current official client docs.

This project has also been manually checked against Claude, GitHub Copilot CLI, VS Code / Copilot Chat, OpenAI Codex, and Gemini CLI.

Claude Code

CLI registration or shared project JSON.

Fastest path:

claude mcp add --transport stdio desktop-touch -- npx -y @harusame64/desktop-touch-mcp

HTTP path:

claude mcp add --transport http desktop-touch http://127.0.0.1:23847/mcp

Shared project .mcp.json for stdio:

{
  "mcpServers": {
    "desktop-touch": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@harusame64/desktop-touch-mcp"]
    }
  }
}

HTTP version:

{
  "mcpServers": {
    "desktop-touch": {
      "type": "http",
      "url": "http://127.0.0.1:23847/mcp"
    }
  }
}
GitHub Copilot CLI

Interactive form or ~/.copilot/mcp-config.json.

Interactive path:

/mcp add

Then enter:

  • Server Name: desktop-touch
  • Server Type: Local or STDIO
  • Command: npx -y @harusame64/desktop-touch-mcp

For HTTP, choose HTTP and use http://127.0.0.1:23847/mcp.

JSON config for stdio:

{
  "mcpServers": {
    "desktop-touch": {
      "type": "local",
      "command": "npx",
      "args": ["-y", "@harusame64/desktop-touch-mcp"],
      "env": {},
      "tools": ["*"]
    }
  }
}

HTTP version:

{
  "mcpServers": {
    "desktop-touch": {
      "type": "http",
      "url": "http://127.0.0.1:23847/mcp",
      "tools": ["*"]
    }
  }
}
VS Code / Copilot Chat

Use .vscode/mcp.json.

Stdio:

{
  "servers": {
    "desktop-touch": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@harusame64/desktop-touch-mcp"]
    }
  }
}

HTTP:

{
  "servers": {
    "desktop-touch": {
      "type": "http",
      "url": "http://127.0.0.1:23847/mcp"
    }
  }
}

Once it is registered, open Copilot Chat, switch to Agent mode, and enable the server from the tools picker.

OpenAI Codex

Use ~/.codex/config.toml.

Stdio:

[mcp_servers.desktop_touch]
command = "npx"
args = ["-y", "@harusame64/desktop-touch-mcp"]

HTTP:

[mcp_servers.desktop_touch]
url = "http://127.0.0.1:23847/mcp"

Check what Codex sees:

codex mcp list
ChatGPT Developer mode

Treat ChatGPT as an HTTP-first client.

The current public OpenAI docs describe ChatGPT Developer mode with remote MCP over streaming HTTP or SSE. For desktop-touch-mcp, that means the local HTTP mode is the right public setup story.

npx -y @harusame64/desktop-touch-mcp --http
http://127.0.0.1:23847/mcp
  1. Open Settings -> Apps -> Advanced settings -> Developer mode
  2. Enable Developer mode
  3. Create an app for your MCP server
  4. Add the local MCP endpoint
Gemini CLI

Use settings.json.

Gemini CLI supports both user and workspace config:

  • ~/.gemini/settings.json
  • .gemini/settings.json

Stdio:

{
  "mcpServers": {
    "desktop-touch": {
      "command": "npx",
      "args": ["-y", "@harusame64/desktop-touch-mcp"]
    }
  }
}

HTTP:

{
  "mcpServers": {
    "desktop-touch": {
      "httpUrl": "http://127.0.0.1:23847/mcp"
    }
  }
}

After registration, /mcp is the quickest way to check status.

Tiny advice

Make the first run obvious.

Show stdio first

It is the least surprising local path for most technical readers.

Show HTTP second

It removes ambiguity for URL-based clients and web clients.

Keep one snippet per client

Readers should never have to translate one client’s config shape into another by hand.

Go to the repository Back to project top