weather-packing-bot

A scheduled bot that posts a Slack message with a multi-day weather forecast and a packing recommendation for a recurring three-city travel route. It is also a small reference implementation of an MCP (Model Context Protocol) server and client pair: the forecast data is served by a standalone MCP server and consumed by an MCP client, rather than being fetched directly by the bot script.

Overview

The bot runs on a weekly schedule (Fridays, 15:00 CEST) and posts a Slack message covering the weather from that day through the following Monday evening. The forecast and packing recommendation are itinerary-aware: they account for which of three cities the traveler is expected to be in on each day of that window, rather than showing all locations' weather indiscriminately.

GitHub Actions (scheduled, Fridays)
        │
        ▼
bot/packing_advisor.py            MCP client
        │  spawns as subprocess, communicates over stdio
        ▼
mcp_server/weather_server.py      MCP server, queries Open-Meteo
        │
        ▼
   Slack incoming webhook

Why an MCP server and client

The server exposes weather tools without any knowledge of who calls them or why - it could be this bot, an IDE assistant, or any other MCP-capable client. The client does not import the server's Python functions directly; it discovers available tools at runtime (session.list_tools()) and invokes them by name over the protocol (session.call_tool(...)). This separation is what distinguishes the design from simply calling a function in the same process.

Components

mcp_server/weather_server.py

An MCP server (using the MCPServer class from the mcp package) that wraps the Open-Meteo API and exposes four tools:

Cities are defined in the COORDINATES dict and can be extended freely.

bot/packing_advisor.py

An MCP client that:

1. Computes the forecast window - the current date through the following Monday, inclusive - via forecast_window().

2. Spawns weather_server.py as a subprocess, performs the MCP handshake, and calls get_forecast_all and render_temperature_chart for that window.

3. Matches each date in the window against WEEKLY_ITINERARY, a dict mapping weekday to the city (or cities) the traveler is expected to be in that day, so that only relevant locations inform the output.

4. Passes the matched days to recommend_packing(), a rule-based function using temperature and precipitation thresholds (no external AI call).

5. Formats a Slack Block Kit message - day-by-day breakdown, packing list, and chart - and posts it via an incoming webhook.

Travel itinerary

WEEKLY_ITINERARY in bot/packing_advisor.py encodes a recurring weekly route:

WEEKLY_ITINERARY = {
    4: [("Nijmegen", "arrival")],       # Friday
    5: [("Nijmegen", "")],              # Saturday
    6: [("Nijmegen", "")],              # Sunday
    0: [("Den Bosch", "day"), ("Kerpen", "evening, back home")],  # Monday
}

Monday has two entries because it is a transition day: the itinerary places the traveler in Den Bosch during the day and back in Kerpen by evening. This dict is the single place to edit if the route changes.

The chart

render_temperature_chart produces a plain-text, monospace chart rather than an image. This is a deliberate constraint: a Slack incoming webhook can only post text or Block Kit JSON - it cannot upload a binary file. Posting an actual image would require a Slack bot token (an installed Slack App with the files:write scope) and a call to Slack's files.upload Web API instead of the webhook. That is a viable extension but requires managing an additional credential; the current implementation avoids that requirement entirely.

Setup

Prerequisites

• A Slack app with an incoming webhook URL.

• A GitHub repository to host the code and run the scheduled workflow.

Local run

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env   # fill in SLACK_WEBHOOK_URL
export $(grep -v '^#' .env | xargs)
python bot/packing_advisor.py

SIMULATE_DATE (see .env.example) can be set to an ISO date to exercise the itinerary/forecast-window logic without waiting for an actual Friday.

The MCP server can also be inspected on its own:

pip install "mcp[cli]"
mcp dev mcp_server/weather_server.py

This opens the MCP Inspector in a browser, useful for viewing tool schemas and calling tools such as get_forecast or render_temperature_chart directly.

GitHub repository configuration

• Add SLACK_WEBHOOK_URL as a repository secret (Settings → Secrets and variables → Actions → Secrets).

• No other repository configuration is required; the itinerary is defined in code rather than as a repository variable.

Schedule

.github/workflows/weekly-forecast.yml runs every Friday at 13:00 UTC (15:00 CEST) via cron: "0 13 * * 5". Cron schedules in GitHub Actions do not observe daylight saving time, so during the CET (winter) period this fires at 14:00 local time instead of 15:00; the cron expression can be adjusted seasonally if that offset matters. The workflow also supports manual triggering (workflow_dispatch) with an optional simulate_date input, useful for testing before relying on the schedule.

Known limitations and possible extensions

• The Slack message includes a text-based chart rather than an image, for the reason described above. A PNG chart via matplotlib, combined with a Slack bot token and files.upload, is a possible follow-up.

• recommend_packing() is rule-based. Replacing it with a call to an LLM (e.g. the Anthropic API) would allow more natural-language output; the MCP tool results are plain dicts, so they can be passed directly into such a call.

• The itinerary is a fixed weekly pattern. A calendar-integration source (e.g. reading actual travel dates from a calendar) would generalize it beyond a repeating weekly route.

Notes on the source material

The original draft server script had several issues that were corrected during development of this implementation:

• An unused, non-existent httpx2 import.

• get_params() built a parameters dict but did not return it.

• The Open-Meteo client was constructed only inside if __name__ == "__main__", making it unavailable to the tool functions that referenced it at module scope.

• process_response attempted to assign into a list using a dict_keys object as an index, which is not valid.

• Tool functions returned pandas DataFrame objects and numpy scalar types, which are not JSON-serializable; MCP tool results must be serializable, since they are transmitted as JSON.