micytao/vllm-playground — reverse-engineered prompt

Build me a modern local web dashboard for running and chatting with vLLM models. I want to start and stop a vLLM server from the browser, connect to remote vLLM servers with a URL and API key, and switch between several instances in tabs so I can compare local, container, and remote backends.

The app should have a clean chat interface with streaming replies, markdown, image upload for vision models, tool calling, structured output options, and a place to connect MCP servers with approval before tools run. Add pages for managing instances, viewing logs and basic observability, running simple benchmarks, counting tokens, and loading saved recipes or configs.

Please include support for GPU, CPU, macOS Apple Silicon, custom virtual environments, containers, and deployment notes for OpenShift or Kubernetes. Also include a Claude Code page that lets open source models served by vLLM act as the backend. Look up current vLLM docs online if needed.