nexrad-mcp

An MCP server that gives an AI client (Claude Desktop, Claude Code, any MCP client) direct access to raw NEXRAD radar data — the same data classes a human interrogates in RadarScope:

• Level II dual-pol moments (reflectivity, velocity, CC, ZDR, spectrum width, differential phase, opt-in KDP retrieval), queried at any point, at any tilt, for the latest volume or any archived timestamp;

• Level 3 derived products (digital VIL, enhanced echo tops, 1-hour and storm-total rainfall, hydrometeor classification, storm cell tracks, mesocyclone detections);

• Active NWS warnings for a point;

• NWS hourly forecasts and thunderstorm outlook for a point;

• Lightning detection via GOES GLM satellite data.

Unlike weather MCPs that return pre-rendered radar images, this decodes the actual data with Py-ART and MetPy, so you can ask questions like "what's the CC over my house right now, and is that a debris signature or just clutter?" — the same interrogation you'd do by hand in RadarScope.

For AI agents: self-install

If you are an AI agent with shell access and want to use this server, install and register it in one command (Claude Code):

claude mcp add nexrad -- uvx --from git+https://github.com/gm2211/nexrad-mcp nexrad-mcp

Verify the installation by piping an initialize + tools/list JSON-RPC handshake into the server and checking that 13 tools are listed:

printf '%s\n' \
  '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"probe","version":"0"}}}' \
  '{"jsonrpc":"2.0","method":"notifications/initialized"}' \
  '{"jsonrpc":"2.0","id":2,"method":"tools/list"}' \
  | uvx --from git+https://github.com/gm2211/nexrad-mcp nexrad-mcp

Expect a tools/list response containing 13 tools. Note that macOS has no timeout command; if you need a time limit, prefix the pipeline with perl -e 'alarm 120; exec @ARGV' instead.

The first run downloads heavy dependencies (Py-ART, scipy, and friends), so allow a few minutes before deciding it has hung. Subsequent runs start in seconds from uv's cache.

If a pre-0.2.0 version of this server was installed before, run uv cache clean nexrad-mcp once — otherwise uv stays pinned to the old commit.

Agents without shell access (for example, claude.ai web chat) cannot self-install this way. In that case the server must be hosted remotely (nexrad-mcp --transport http) and added as a custom connector by the user, or wired into Claude Desktop's claude_desktop_config.json by the user.

Sandboxed agents with a shell but no control over their MCP client config (for example, Cowork sessions) don't need MCP registration at all: call the library directly instead, e.g. uvx --from git+https://github.com/gm2211/nexrad-mcp python -c "from nexrad_mcp.forecast import get_hourly_forecast; print(get_hourly_forecast(38.9, -77.0, 6))" — every MCP tool is a plain function in nexrad_mcp.radar, .level3, .forecast, .lightning, and .warnings.

Related MCP server: Weather MCP Server

Installation

Requires uv. The first run is slow (a minute or two): uvx resolves and installs Py-ART/MetPy, which pull in scipy and friends. Subsequent runs start in seconds from uv's cache.

Claude Code

claude mcp add nexrad -- uvx --from git+https://github.com/gm2211/nexrad-mcp nexrad-mcp

Claude Desktop — add to claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "nexrad": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/gm2211/nexrad-mcp", "nexrad-mcp"]
    }
  }
}

Any other MCP client — configure the server command as:

uvx --from git+https://github.com/gm2211/nexrad-mcp nexrad-mcp

(stdio transport; no API keys or credentials — all data sources are public.)

Transport flags — by default the server speaks stdio. To host it over streamable HTTP instead (e.g. as a remote connector), use:

nexrad-mcp --transport http --host 127.0.0.1 --port 8748

• --transport: stdio (default) or http

• --host: bind address for http (default 127.0.0.1)

• --port: bind port for http (default 8748)

For development, clone and uv sync, then run uv run nexrad-mcp.

Tools

site is a 4-letter radar ID (e.g. KLWX = Sterling VA) — use find_nearest_radar if you don't know it.

RadarScope data-parity coverage

Beyond RadarScope parity — forecasting (RadarScope doesn't do this at all):

Data sources

• Level II: NOAA Open Data volumes on AWS S3, discovered via the nexradaws index. Decoded with Py-ART.

• Level 3: the public unidata-nexrad-level3 S3 bucket (anonymous access). Decoded with MetPy; scaling cross-checked against Py-ART.

• Warnings, hourly forecast, thunderstorm outlook: api.weather.gov (requires only a User-Agent header).

• Lightning: GOES GLM (Geostationary Lightning Mapper) Level 2 data on the public noaa-goes19 (GOES-East) and noaa-goes18 (GOES-West) S3 buckets, anonymous access. Decoded with netCDF4.

Try it without MCP

uv run python -c "from nexrad_mcp import radar as R; \
import json; print(json.dumps(R.query_point('KLWX', 38.905, -78.235), indent=2))"

How the polling works

Each call lists the current UTC day's volumes for the site via the nexradaws index (the main archive bucket blocks anonymous listing, so we go through the index rather than raw S3), grabs the newest *_V06 key, downloads it, and decodes it. A completed volume in VCP 212 lands every ~4–6 minutes, so re-querying more often than ~30–60s just returns the same file. Volume loads are LRU-cached so repeated queries on one volume don't re-decode. Level 3 products update on a similar cadence and are fetched by day-prefixed key listing (the bucket is a flat namespace, SSS_PPP_YYYY_MM_DD_HH_MM_SS).

Caveats

• Not for life-safety. This is an analysis aid. For warnings, always use NWS / official sources (get_active_warnings surfaces exactly those). The tools return numbers; interpreting a debris signature vs. clutter still needs judgment (which is exactly why it pairs well with an LLM that can weigh reflectivity + velocity + CC together).

• Beam height rises with range: at 100+ km the lowest tilt is well above ground, so a clean surface reading is best within ~60–80 km of the radar. get_vertical_profile reports the actual beam height of every sample.

• Radar-estimated rainfall (DAA/DTA) is an estimate; gauges beat radar.

License

MIT. NEXRAD data is public domain (NOAA); Level 3 mirror courtesy of Unidata.