salesforcecli/mcp — reverse-engineered prompt

Build me a Salesforce MCP server that I can plug into Cursor, Claude Code, or VS Code so an AI assistant can safely work with my Salesforce orgs. I want it to be easy to start through npx and configurable through a simple mcp.json style file where I can choose which orgs are available, which toolsets are turned on, and any specific tools I want, like running Apex tests.

The main goal is to let the AI read and manage Salesforce resources in a controlled way, especially org info, metadata, data, and users. Please make sure it only exposes the orgs and tools I explicitly allow, with sensible security minded defaults. It should also support common flags like debug logging, turning telemetry off, optional dynamic tool loading, and allowing non GA tools when I choose that.

Please include clear setup instructions and working config examples for Cursor, Claude Code, and VS Code with Copilot. If you need details, look up the current Salesforce DX MCP docs online and match the expected behavior closely.