meteoswiss-llm-tools

🌦️ MeteoSwiss LLM Tools

Swiss weather data for AI assistants — powered by MeteoSwiss Open Government Data (OGD), the same data behind the MeteoSwiss app and website. Free, no API key required.

meteoswiss-mcp.ars.is — try the hosted service instantly, no setup needed.

This repo is also a working answer to a design question: how should you give AI agents access to a public dataset? It implements the same MeteoSwiss data access twice — as an agent skill (markdown instructions plus bash scripts, no server) and as an MCP server (structured tools, fuzzy matching, caching, hosted). The two approaches are compared honestly in the skill vs. MCP case study.

A third piece, meteoswiss-forecast-evals, demonstrates eval-driven interface design: a promptfoo suite measuring how well 13 LLMs read the forecast JSON, which settled a real design decision — local-time timestamps beat UTC, with hour-level lookups scoring ~100% vs. ~0%.

What the tools provide:

• Multi-day forecasts for ~6000 Swiss locations (postal codes, stations, place names)
• Real-time measurements from ~300 stations (~160 full weather + ~140 precipitation-only), updated every 10 minutes
• Station discovery by name, canton, or GPS coordinates
• Pollen monitoring from ~15 stations across Switzerland
• Climate series from the National Basic Climatic Network (NBCN), going back decades
• MeteoSwiss website search and content retrieval

What this repo demonstrates

• An agent skill — teach an agent to fetch open data directly with curl/awk/jq: ~630 lines of markdown and bash, zero infrastructure. → packages/meteoswiss-skills
• An MCP server — the same data as structured, validated tools with fuzzy station matching, geocoding, TTL-tiered caching, a real test suite, Docker, and a hosted instance. → packages/meteoswiss-mcp
• Eval-driven interface design — treat tool output as an interface for a language model, and measure its legibility before shipping. → packages/meteoswiss-forecast-evals

Read the comparison: Skill vs. MCP Server: Two Ways to Give AI Agents the Same Data.

Choose your approach

Both approaches answer the same weather questions. Which to install depends on your agent:

	MCP Server	Agent Skill
What it is	Standalone server exposing 7 structured tools via MCP	Markdown instructions + 5 bash scripts the agent runs directly
Works with	Claude Desktop, Claude.ai, Cursor, any MCP client	Claude Code, Cursor, any agent with shell access
Coverage	Forecasts, current weather, stations, pollen, climate series, website search	Forecasts, current weather, stations, pollen
Extras	Fuzzy matching, geocoding, caching, DE/FR/IT prompts, structured JSON	No server, no Node.js — just curl, awk, jq
Size	~6.6k lines TypeScript, tested in CI	~630 lines markdown + bash
Install	One-liner (hosted), npm, or Docker	Plugin marketplace, Skills CLI, or symlink

Full comparison — parity matrix, engineering trade-offs, context cost, when to choose which: docs/skill-vs-mcp.md.

MCP server — quickstart

Use the hosted instance (no installation):

# Claude Code
claude mcp add meteoswiss https://meteoswiss-mcp.ars.is/mcp

For Cursor, install from the Cursor Directory or add manually via Settings → MCP.

Or self-host with Docker:

docker run -p 3000:3000 ghcr.io/eins78/meteoswiss-mcp:latest

See the meteoswiss-mcp README for Claude Desktop setup, environment variables, and full documentation.

Agent skill — quickstart