arcadeai/arcade-mcp ? reverse-engineered prompt

Build me an open source Python framework for making MCP servers and custom tools that can plug into AI agents. I want it to be easy to start a new server, add tools with simple decorators, and run it either locally over stdio or as an HTTP service for clients like Claude, Cursor, and VS Code.

It should support the main MCP features, like tools, resources, prompts, logging, progress updates, and sampling, plus a clean way to test tool quality with real model calls. I also want built in support for authenticated tools, so a tool can ask for OAuth access to services like GitHub, Google, Slack, or similar, and the secret tokens should stay hidden from the client and the model.

Please include a few working examples, a CLI for creating, configuring, and deploying servers, and make sure the whole thing feels simple enough for someone to use without reading a ton of docs first.