v0.2.0 — MCP Apps + Remote Parquet

What's new

MCP Apps Server — interactive Plotly charts rendered directly in Claude Desktop:

• analyze_air_quality — time series for Saharan dust, SO₂, PM2.5 with forecast stitching + UBA station overlay
• show_air_quality_map — spatial distribution on a European map (scattergeo), variable dropdown with lazy loading
• compare_cities — multi-city comparison on shared time axis
• query_measurements — query the remote Parquet archive
• 3 MCP prompts: luftqualitaet, staedtevergleich, saharastaub_lage

Remote Data via DuckDB httpfs — no local CAMS download, no API key needed:

• Daily CI workflow uploads analysis.parquet + forecast.parquet to the data-latest release
• MCP server reads directly from GitHub via DuckDB read_parquet('https://...')
• Coverage: Central Europe (44–56°N, 2–18°E), 14 days analysis + 5 days forecast

MCP Apps protocol (SEP-1865):

• CallToolResult with content (model summary) + structuredContent (UI data)
• App-only tools (visibility: ["app"]) for lazy data loading
• UI resources with CSP metadata, @modelcontextprotocol/ext-apps SDK v1.3.2

CLI — still works as before with local CAMS downloads (~/.cdsapirc required).

Setup (Claude Desktop)

{
  "mcpServers": {
    "dust-analyzer": {
      "command": "uv",
      "args": ["run", "--directory", "/path/to/dust-analyzer", "python", "-m", "dust_analyzer", "--mcp"]
    }
  }
}

No API key needed — data comes from the public GitHub Release.