swiss_meteo_mcp

This is an experimental project to learn about MCP + GitHub Copilot, using the latter to create the MCP server code and then access it through the Chat funciton of copilot. I also managed to use the MCP server through Amazon Q chat (see further below).

Journey

• I prompted the goal as stated above for an initial pass.

• After uvicorn was used wrongly in the .vscode/mcp.json config, which is for http endpoints, not for stdio ones, debugging back and forth with copilot fixed the issue and I had a running MCP server scaffolded (that were unrelated to swiss meteo):

• Availabilities:

  • prompt to summarise text, which could either be detailed or brief (/mcp.swiss-meteo-mcp.summarize-notes ).

  • tool to add a note (#add-note).

  • resource to list notes, not accessible explicitly (just implicitly) through copilot chat.

• One can see below that it works as expected, with autocompletion, after the server was started:

• I wanted to add a tool to get the current weather from swiss meteo (hence the name of the package), but there is no official API I found out with copilot chat.

• I then asked for a few alternatives and decided to go with wttr.io and let copilot implement it.

• I also factored out some functionality so I could let copilot write a simple unit test that worked.

• So my own new tool is #get-temperature

Lessons Learned

in general

• Building an MPC server seems a standard task. GitHub Copilot managed kind of well, but not flawless.

• Integration of mcp servers is working well with copilot it seems.

• Invocation of tools and resources is possible implicitly or explicitly. When used, it is communicated transparently in the copilot chat.

• MCP servers can run locally (with like uv run or container isolation) or remotely.

• MCP prompts are really just parametrised snippets.

• resources are read-only info.

• tools are supposed to have side effects.

GitHub Copilot

• can also write git commit messages (not so impressive)

• used 20% of free code suggestions in one day (and might be used for retraining).

Amazon Q

• Can discover MCP server in .vscode/mcp.json in own project, start it when a relevant question is asked in the Q chat and use the tool to answer the question, but using Python source package not MCP tool over stdio.

• to use the mcp server outside the source repo, we had to put the mcp.json file into the amazon q global directory at ~/.aws/amazonq/mcp/mcp.json and rename the servers: key to mcpServers: for Q to recognise it:

{
  "mcpServers": {
    "swiss-meteo-mcp3": {
      "type": "stdio",
      # not portable, most mcp servers use uvx run [pypiname.module@latest] to install on fly and run
      "command": "/Users/lorenz/git/swiss_meteo_mcp/.venv/bin/python", 
      "args": [
        "-m",
        "swiss_meteo_mcp.server"
      ]
    }
  }
}

Other

• drag and drop screenshot into VS Code editor works well