mcp-streamable-http-quickstart

MCP Quickstart Weather App Extensions

After reading the Official MCP quickstart examples on MCP server and client, do you wonder

• How to upgrade the simple stdio-based example to HTTP server/client towards real-world uses?
  • The latest MCP document (June 2025) lists SSE as the default HTTP transport protocol
  • The latest MCP specification (March 2025) further upgrades SSE to Streamable HTTP protocol
• How to replace the Anthropic API with OpenAI API widely used in open source inference servers like vllm?

Goal of This Repository

1. Patch the official MCP quickstart weather app to use:
   • SSE or Streamable HTTP as the transport protocol between client and server
   • OpenAI API for LLM calls
2. Explain each modification for readers to understand these extensions

How to Run

1. Install uv
2. Choose the protocol in your mind, either sse or streamable-http
3. Open two terminals on one host (hardcoded localhost HTTP server in this example)
4. Term 1: run server
   • Go to the server directory weather-server-python
   • Start the server uv run server PROTOCOL_OF_YOUR_CHOICE
5. Term 2: run client
   • Go to the client directory mcp-client-python
   • Setup environment variables for OpenAI endpoint and API
     • export OPENAI_BASE_URL=http://xxx/v1
     • export OPENAI_API_KEY=yyy
   • Start the client uv run client PROTOCOL_OF_YOUR_CHOICE

Explanation of Modifications

Use SSE/Streamable-HTTP Instead of Stdio for Transport Protocol

• Server: use mcp.run('sys.argv[1]') instead of mcp.run('stdio') given sys.argv[1] is either sse or streamable-http
  • SSE protocol: server main endpoint is http://localhost:8000/sse
  • Streamable HTTP protocol: server only endpoint is http://localhost:8000/mcp
• Client: load rs (readstream), ws (writestream) from sse_client or streamablehttp_client intead of stdio_client in the original MCP quickstart example
  • sse_client awaited return
  • streamablehttp_client awaited return

Swap Anthropic API to OpenAI API for LLM call

• Replace the LLM call function
  • self.anthropic.messages.create() -> self.client.chat.completions.create()
  • Dynamic model id for vllm
  • The tools argument uses a little different formatting
• Replace the LLM response object handling
  • response -> response.choices[0].message