WeatherFlow Tempest MCP Server

A Model Context Protocol (MCP) server that gives AI assistants read-only access to your own WeatherFlow Tempest weather station(s): current observations, forecasts, and station metadata.

It is not a general weather service. It does not cover:

• locations away from your station, or global/regional weather (use a public weather API)
• air quality, pollen, or smoke
• severe-weather alerts, radar, or watches/warnings
• historical archives beyond what the live Tempest API returns

🌤️ Features

• Current conditions from your station: temperature, humidity, pressure, wind, rain, solar radiation, UV, and lightning
• Forecasts: hourly (up to 48 h) and daily (up to 10 days), plus a current snapshot
• Station inventory: your stations, their locations, devices, and sensor capabilities
• Caching in memory and on disk, with a refresh option for the latest reading
• Structured results and errors for agents: typed output schemas, explicit units, RFC3339 timestamps, and error responses that include a ready-to-run corrected call

🚀 Quick Start

Prerequisites

• Python 3.13 or higher
• WeatherFlow API token (get one at tempestwx.com/settings/tokens)

Installation

While each client has its own way of specifying, you'll generally use the following values:

Development Version

If you'd like to use the latest and greatest, the server can be pulled straight from GitHub. Just add an additional --from argument:

Install as a Desktop Extension (.mcpb)

For one-click installation in apps that support MCP Bundles (e.g. Claude for macOS/Windows):

1. Download the .mcpb from the latest release.
2. Open the file with your MCPB-capable client to launch the install dialog.
3. Paste your WeatherFlow API token when prompted (the cache settings are optional).
4. Make sure the extension is enabled (in Claude Desktop: Settings → Extensions). It is not turned on automatically until the required API token has been provided.

The bundle uses the MCPB uv runtime (manifest_version 0.4), so the host must ship a recent enough uv/MCPB runtime; it resolves dependencies on first launch — no separate Python install is required.

Troubleshooting: if the server fails to start because its install directory is read-only (uv cannot create a .venv next to the bundle), set the UV_PROJECT_ENVIRONMENT environment variable for the server to a writable path before launching.

📋 Configuration

Environment Variables

Caching & data freshness

The server caches responses in two layers:

• In-memory (WEATHERFLOW_CACHE_TTL / WEATHERFLOW_CACHE_SIZE): all four data tools (tempest_get_capabilities is static and uncached).
• On disk (WEATHERFLOW_DISK_CACHE_TTL, default 24h): tempest_get_stations and tempest_get_station_details only. Stored under platformdirs.user_cache_dir("mcp-server-tempest") in a per-token (hash-keyed) subdirectory.

To bypass the cache for current data, pass refresh=true to tempest_get_observation or tempest_get_forecast. Station data has no refresh argument: to clear it, restart the server (in-memory) or delete the cache directory (disk). Every data result carries retrieved_at, the time the server fetched it from WeatherFlow.

Transport

stdio (the default for uvx mcp-server-tempest and the configuration above).

🛠️ Usage

Tools

The same capability summary is available as the MCP resource tempest://capabilities. It is the authoritative, machine-readable contract; the table above is a quick guide.

Reading results

• Units. Observation values are always metric/SI and are described by the result's units object. Forecast values are described by the forecast's own units. The observation's station_units is the station owner's display preference, not the unit of the values: convert to it when presenting.
• Times. Raw fields such as timestamp, time, and lightning_strike_last_epoch are Unix seconds. Observations also carry observed_at and lightning_strike_last_at, and every data result carries retrieved_at, all RFC3339 in UTC. Hourly forecast entries include local_day and local_hour in the station's own timezone.
• Errors come back as a tool error whose structured content has a symbolic code (branch on it, not on message), temporary (plus retry_after_ms when true), and, when the server can build one, repair: a corrected call to make as-is. The full error contract is the error_channel field of the capability summary. Version 0.11.0 changed the error format (next became repair); see the CHANGELOG.

🌟 Example questions

Once the server is configured, ask your assistant in plain language. Some examples, with the tool it will typically use:

Building your own client instead? Any MCP client works; the tools return structured results described by each tool's output schema. See, for example, the FastMCP client docs.

🤝 Contributing

Contributions are welcome. AGENTS.md is the guide for humans and AI agents alike: environment setup with uv, how to run the tests and linters, branch and commit conventions (conventional commits; signed commits on main), and how pull requests are reviewed and merged. Release notes live in CHANGELOG.md.

Please report security issues privately, as described in SECURITY.md.

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

🙏 Acknowledgments

• WeatherFlow for providing the Tempest weather station and API
• Model Context Protocol for the MCP specification
• FastMCP for the MCP server framework

📞 Support

• Issues: GitHub Issues
• Documentation: WeatherFlow API Docs
• Community: WeatherFlow Community