mcp-weather-connector

An MCP server that gives an AI agent access to weather forecasts through Open-Meteo — a public API with no key or registration required.

Deliberately small: the repository is about how the tool layer for an agent is structured, not about weather. Weather here is a relatable pretext for showing the trust boundary between the model and the outside world.

Why

Model Context Protocol is a way to give a language model the right to take actions in external services. Once that right exists, the question arises: what happens if the model passes garbage to a tool right as the third-party service returns a 503?

Three decisions are made explicit in this code:

Model input is untrusted. The model can pass an empty string, a ten-kilobyte wall of text, or control characters. All of this is filtered out in server.py before the network call — not in a third-party API.

Errors are distinguished by type. Retrying a 4xx is pointless — it is a request error, and a retry will not fix it. A 5xx is more likely a one-off failure — one retry with a pause. The logic lives in _get_json and is verified by tests that assert the call count.

What goes out is text, not raw JSON. The agent has nothing to second-guess and no way to err when relaying the answer. A 200 response without the required field counts as a failure, not as a reason to return an empty value.

As a separate note: a day count outside the supported range is clamped to the boundary rather than treated as an error. If the model asks for a 30-day forecast, it is more useful to return seven days than to break the conversation with an error message.

Related MCP server: open-meteo-mcp-server

Tools

locate exists for the single scenario where the agent is not sure about the spelling: better to check with the user than to silently return the weather for the wrong city.

Installation

pip install -e .

Connecting to Claude Desktop

In claude_desktop_config.json:

{
  "mcpServers": {
    "weather": {
      "command": "mcp-weather"
    }
  }
}

After the app restarts, the tools will appear in the list of available tools.

Development

pip install -e ".[dev]"
pytest
ruff check .

The tests cover what actually breaks: empty and oversized input, a non-numeric number of days, a 4xx with no retry, a 5xx with one retry, success on the second attempt, a 200 response without a payload, and an unknown weather code. HTTP is mocked with respx — no network is needed in the tests.

Structure

src/mcp_weather/
    client.py    — работа с Open-Meteo: типы, повторы, разбор ответа
    server.py    — MCP-слой: валидация входа и форматирование вывода
tests/
    test_tools.py

The network code is deliberately separated from the MCP layer: the client is tested on its own, and the agent's tools remain a thin wrapper over straightforward functions.

License

MIT