Smarter Weather MCP Server

Complete weather overview for a location: current conditions, daily forecast (day/night periods, SPC threats, severity, CAPE, UV), active alerts, and convective outlooks in one call. Data is pre-aggregated across NBM, HRRR, GFS, RTMA, and SPC and unit-converted server-side. This is the primary weather tool; reach for lower-level tools only when you need raw observations or a specific dataset. Accepts a place name directly. Examples: {"location": "Denver"} or {"location": "Portland, OR", "days": 5} or {"lat": 41.4, "lon": -92.9}.

Expert forecaster text products. type=afd: Area Forecast Discussion. type=hwo: Hazardous Weather Outlook. type=now: WFO short-term NOW. type=fwf/hls/esf: local fire weather / hurricane local statement / hydrologic discussion. type=mcd: SPC Mesoscale Discussion. type=mpd: WPC Mesoscale Precipitation Discussion (flash flood). type=swo/fwd/ero: national outlook discussions. type=tcd/tcp/tcm/twd/two: NHC tropical text (type=two is the text TWO, not GIS nhc_two). type=pmd: WPC/CPC desk discussion (pass awips_id for a specific desk, e.g. PMDSPD). type=pwo: SPC public weather outlook. National types (swo/fwd/ero/tcd/tcp/tcm/twd/two/pmd/pwo) need no location; `day` selects the outlook day for swo and fwd. summary_only=true returns the pipeline LLM summary without the full body. Examples: {"location": "Des Moines", "type": "afd"} or {"type": "swo", "day": 2, "summary_only": true}.

Probabilistic forecast guidance from NBM for one aspect of the weather: percentile ranges (p10-p90), exceedance probabilities, and ensemble spread. Use this for any question about odds, ranges, potential or confidence ("how much could we get", "worst case for the wind", "how sure is this") -- a deterministic forecast value cannot answer one. Reading the percentiles: p50 is the most likely outcome, p90 is the reasonable worst case when the risk is the high end (snow totals, wind, rainfall), and p10 is the reasonable worst case when the risk is the low end (cold, minimum visibility, ceiling). A single percentile is not the forecast -- report the likely value with the tail that matters, and label which is which. Aspects: precip (PoP, QPF + percentiles), snow (accumulation percentiles, >1/2/4in probabilities, snow level), ice (freezing rain, accretion), temperature (temp/dewpoint + stddev), wind (speed/gust percentiles), severe (hail/tornado/damaging-wind probabilities), aviation (LIFR/IFR/MVFR visibility + ceiling probabilities), confidence (ensemble stddev; low spread = settled forecast, high spread = details still in play). Examples: {"location": "Denver", "aspect": "snow", "hours": 72} or {"lat": 32.9, "lon": -97.0, "aspect": "severe"}.

How accurate our forecasts have actually been near a location, measured against observed analysis truth. Returns bias (positive = the model runs high), mean absolute error, RMSE, and a skill score against local climatology, per model, weather variable, and forecast lead time; continuous and vector entries also carry persistenceSkillScore, skill against the analysis at forecast issue time (null means not enough persist pairs, not zero skill -- do not compare it to skillScore as if they shared a denominator), and analysisDisagreementMae, the analyses' own disagreement at that lead -- a floor on how good the forecast can look, not a skill score and not an excuse (null means the sibling row is missing or below minimumSamples); for probability forecasts, the Brier score and a reliability breakdown. Use this to qualify a forecast rather than assert it -- "NBM has been running 1.8F warm at 3-day leads near you, so treat that 72 as around 70" -- and to answer "how much should I trust this forecast", "is the model biased here", or "how accurate were you last month". Evidence is reported at three scopes side by side: the exact point (strongest, slowest to accumulate), the ~50km neighborhood, and the ~300km region. Prefer the most specific scope that has samples. Metrics below minimumSamples observations are withheld and listed under insufficientHistory with their count -- say that history is still accumulating rather than treating thin numbers as evidence. Coverage is a rolling recent window over verified US variables, not all of history. Entries are per model and their samples are not matched, so never conclude that one model beats another by comparing their numbers here. Each entry states the truth field it was measured against -- one designated analysis per variable -- so never compare numbers carrying different truth values either. Each entry also states the regime it was measured under: ALL for every observation regardless of weather, or a conditioned tier such as SEA:DJF (winter), SCN1:WINDY / SCN1:WET / SCN1:QUIET (what the forecast was showing), or JC1:NW (a circulation pattern). Pass the regime parameter to ask for a conditioned track record. It falls back, so asking for SCN1:WINDY and getting back regime ALL is a successful answer, not a missing one -- always read the regime field and qualify the claim with it, because "NBM runs warm here when it shows windy" and "NBM runs warm here" are different statements. Regimes overlap by construction across families, so entries under different regimes are alternative answers to one question and must never be compared or added; within SCN1: the labels are mutually exclusive. Entries with a categorical block answer a yes/no question instead of an error magnitude -- did it rain, at the thresholdMm stated on the entry -- with pod (of the times it happened, how often we called it), far (of the times we called it, how often it did not happen), and frequencyBias (above 1 = we call it too often). Use these for "will it actually rain" questions, where a small average error means nothing if the rain lands in the wrong hour. A null rate means the sample cannot answer it -- the event has not happened, or been forecast, enough times to divide by -- and must be reported as unknown, never as zero. The counts beside it are still evidence, and for a rare event they are often the whole answer: "it has only rained twice here in the record" is a useful thing to say.

Day-of-year climate normals (NCEI 1991-2020 30-year averages) for a US location, from the nearest station with a record. Returns normal high, normal low, and normal mean for each date in the window, plus the station and how far away it is. Use this whenever a question needs a baseline rather than a forecast: "is this warm for October?", "what is a typical high here in January?", "how does this week compare to normal?". Pair it with get_forecast to say how far above or below normal the coming days run. Covers dates by day of year, so it answers for any date, past or future -- these are long-period averages, not a forecast and not observed history for a specific year.

NWS daily climate data: type=reports returns CLI daily climate reports (observed high/low/precip vs normals per station); type=records returns RER record event reports (record highs/lows/rainfall actually set). Filter by wfo (3-letter office, e.g. DMX), station, date (YYYY-MM-DD), start/end range, or hours lookback. Examples: {"type": "records", "hours": 48} or {"type": "reports", "wfo": "DMX", "date": "2026-07-04"}.

Nearest RAOB (radiosonde) vertical soundings to a point. Each sounding carries: profile (pressure-indexed thermodynamics: pressure_hpa, height_m, temperature_c, dewpoint_c, wind arrays), wind_profile (height-indexed winds for hodographs/shear), and derived indices (sbcape/mucape/mlcape + cin, lifted_index, k_index, total_totals, pwat_mm, freezing_level_m, lcl/lfc/el, bulk_shear_0_6km_kt). Soundings launch at 00Z/12Z so data can be hours old. Example: {"location": "Norman, OK"}.

Render the nearest RAOB (radiosonde) sounding as a Skew-T log-P + hodograph chart image for visual analysis: temperature/dewpoint traces, wind barbs, height-banded hodograph, and a derived-indices table (CAPE/CIN, lifted index, PWAT, shear, LCL). Soundings launch at 00Z/12Z so data can be hours old. Use get_sounding for the raw profile numbers. Example: {"location": "Norman, OK"}.

Radar-identified storm cells near a location, merging NEXRAD Level III algorithm output from the nearest radar site: storm tracks (cell position, movement, forecast positions), hail index (probability of hail/severe hail + max expected size), mesocyclone detections (rotation), and TVS (tornado vortex signatures). Use during active convection to see what the radar algorithms flag. An empty result means no detected cells -- common outside active storms. Example: {"location": "Norman, OK"}.

Recent NWS Local Storm Reports (LSRs) -- verified reports of tornadoes, hail, damaging winds, flooding near a location. Use to confirm severe weather occurrence or assess reported damage. valid_time is event occurrence (UTC); cite in local time. Example: {"location": "Wichita", "hours": 12, "type": "H"}.

AirNow air quality at a location (CONUS): current overall AQI plus per-pollutant detail (PM2.5, ozone, PM10 concentrations) and the AirNow AQI forecast. AQI scale: 0-50 good, 51-100 moderate, 101-150 unhealthy for sensitive groups, 151-200 unhealthy, 201-300 very unhealthy, 301+ hazardous. pollutants=["aqi"] (default) is the cheap headline call; add pollutant keys or include_forecast=true when the user digs in. Example: {"location": "Boise", "pollutants": ["aqi", "pm25"], "include_forecast": true}.

NWS watches, warnings, advisories. Point (city/ZIP/lat+lon): containing polygons. BBox or US state/DC/CONUS (codes, full names, US/national): intersecting polygons. A city miss is not a statewide all-clear — query the state or a bbox; never say regional inventory is impossible. NY/WA and "New York State"/"Washington State" are states; "New York"/"Washington" stay cities. Omit at for now; at (ISO-8601 UTC) is the snapshot then. Empty = all-clear or purged (~24h). alert_id = detail+geometry, ignores at. Ex: {"location":"WI","events":["Tornado Warning"]}.

Find the optimal time window for an activity based on weather criteria. Scans the forecast and returns daylight-aware periods matching all conditions. Criteria are expressed in the selected `units` system (default imperial: °F, mph, miles, feet). Example: {"location": "Boulder, CO", "criteria": {"min_temperature": 55, "max_wind_speed": 15, "max_precipitation_probability": 20}, "hours": 72, "activity_duration_hours": 3}.

Compare forecast variables across multiple locations side-by-side in one batched call. Returns a distilled per-location series matrix for direct comparison -- prefer this over N sequential forecast calls. Locations accept place names directly. Example: {"locations": [{"location": "Denver"}, {"location": "Boulder, CO"}], "variables": ["temperature_2m", "precipitation_probability"], "hours": 48}.

Current weather right now at a location from two independent sources in one call: the RTMA gridded analysis (exact-point values, updated sub-hourly) and the nearest METAR station observation (ground truth with raw METAR, flight category). Use the analysis for point-accurate values and the station for verification. For a forecast, use get_forecast. Example: {"location": "Pella, IA"}.

Raw time series from a specific dataset for specific variables at a point. Power-user access to any gridded product (NBM, HRRR, GFS, RTMA, MRMS, air quality, ...). Time modes: hours (next N hours, default 24), time_start+time_end (explicit ISO-8601 window), or latest=true (single most-recent value). reference_time pins a specific model run, and each returned series reports the run that served it (reference_time, or reference_times when a series mixes runs) — check it before comparing two runs, since a run older than about 48 hours may no longer be available. For blended forecasts use get_forecast instead. Examples: {"location": "Denver", "dataset_id": "hrrr_surface", "variables": ["temperature_2m"], "hours": 18} or {"lat": 41.4, "lon": -92.9, "dataset_id": "rtma_conus", "variables": ["temperature_2m"], "latest": true}.

Discover the datasets (model grids, analyses, observations) available at a location, with per-dataset freshness (data age, latest model run). Datasets vary by domain (CONUS/Alaska/Hawaii). Use this to find dataset_id values for query_dataset and describe_dataset, or to assess whether data is current before making decisions. Example: {"location": "Anchorage"}.

Variables available in a dataset, with standard names, units, descriptions, and the time range of available data. Use before query_dataset to discover valid variable names. Example: {"dataset_id": "nbm_conus"}.

Growing Degree Units (GDU / GDD) for a US location (CONUS, Alaska, Hawaii), computed from daily max/min temperatures. Pass a crop id (e.g. "corn", "soybean", "wheat") to use calibrated base/upper thresholds, or crop="custom" with base_temp_c (and optional upper_temp_c / method). Without season_start you get per-day GDU across the forecast horizon; WITH season_start (YYYY-MM-DD) you get the cumulative season-to-date total (observed history + today + forecast) plus a per-day cumulative series -- the number a grower tracks against crop milestones. Answers "how many growing degree days has my corn accumulated since May 1?" and "what's the GDU forecast this week?".

Blended hourly forecast: temperature, feels-like, humidity, wind, precipitation probability/amount, conditions, and icon per hour. Snapped to the current hour so hourly[0] is "now". Timestamps are UTC ISO 8601; convert to the local timezone before presenting. Ask for the days you need up front -- one call with days: 4 beats four calls. ALWAYS check hourly_coverage before answering about a specific hour: it reports first_time and last_time (the window the rows actually span), sample_interval_hours (past the first day rows are every 2-3h, not every hour), and truncated: true when upstream returned less than you asked for. If the hour the user cares about is after last_time, say the forecast does not reach that far yet rather than answering from the nearest row you do have. For one stretch of time ask for that stretch with hours_from/hours_to: it comes back hour by hour even where the full range would be sampled. Accepts a place name or coordinates. Examples: {"location": "Portland, OR", "days": 2} or {"lat": 41.88, "lon": -87.63, "hours_from": 36, "hours_to": 48}.

Real-time lightning near a location: GLM satellite flash count (30km/10min) and MRMS ground-truth lightning density + 30-minute probability. The summary field is ready-to-use. A zero flash count means no lightning inside that window -- report it as a quiet observation scoped to the window in `scope`, never as a data gap. Only call when storms may be active or the user asks about lightning. Example: {"location": "Tampa"}.

Resolve a place query to candidate locations with coordinates. Accepts city names ("Denver"), city+state ("Portland, OR" via query), ZIP codes ("50219"), or partial input with fuzzy=true for autosuggest-style matching ("bost" -> Boston). Returns ranked candidates with lat/lon. Most weather tools accept a `location` string directly and geocode internally -- use this tool only to disambiguate ("which Springfield?") or to present location choices to the user. Example: {"query": "Springfield"} returns all major Springfields ranked by place importance.

Render a weather map image for visual analysis. Simple form: pass `product` (a viz-catalog product_id like "mrms_qpe_01h_pass2_conus", "goes_truecolor_conus", "spc_day1_categorical", "hrrr_precip_hybrid_derived_conus" (future radar), "hrrr_subhourly_conus" (15-min Future Radar), "mrms_radar_nowcast_conus", "rtma_conus", "nbm_daily_temps", or "nexrad_l3:{SITE}:{PRODUCT}" for single-site radar, e.g. "nexrad_l3:TLX:N0B") plus a location and zoom (5=regional, 8=metro, 10=city). Composed form: pass `scene` -- a declarative scene document layering basemap + multiple weather products + active alerts + storm features + inline GeoJSON in one image (layers draw bottom-to-top, under basemap labels). Example scene: {"scene":"1.0","view":{"center":{"lat":43.8,"lon":-91.2},"zoom":8},"layers":[{"type":"weather","product":"goes_truecolor_conus"},{"type":"weather","product":"nexrad_l3:ARX:N0B"},{"type":"alerts","filter":{"events":["Tornado Warning"]},"onError":"skip"}]}. Alert filters (all optional, AND-combined): `ids` (specific alerts), `events`, `severities`, `minSeverity` (Extreme>Severe>Moderate>Minor>Unknown). Single-site radar keys: the address is `nexrad_l3:{SITE}:{KEY}` where KEY is `N{tilt}{measurement}` and tilt 0 is the 0.5 degree sweep -- N0B reflectivity (dBZ, where and how heavy), N0G base velocity (knots toward/away from the radar), N0S storm-relative velocity (storm motion removed, so a couplet is rotation rather than translation -- prefer it for rotation questions), N0C correlation coefficient (0-1, debris and hail), N0X differential reflectivity (dB). Legacy codes (N0V, N0R, N0Q) are accepted as aliases. Not every site produces every key; when a render reports which keys a site has, retry with one of those. Optional `time` (unix seconds): closest frame. Forecast (HRRR/nowcast/NBM) honors future times; analysis (MRMS/NEXRAD/RTMA/GOES) clamps to latest past. Pass `time` for future-radar asks — do not claim that capability is missing. Product ids must be real viz-catalog entries -- shorthand like "radar" or "reflectivity" is not one. Omit `product` for the default hybrid precip still. For Alaska and Hawaii prefer a local site or `mrms_precip_hybrid_derived_alaska` over CONUS mosaics, which do not cover them. Returns the rendered image plus per-layer resolved valid times.

METAR surface observations from weather stations: temperature, wind, visibility, ceiling, flight category, raw METAR. Nearest mode (default) returns the closest N stations to a location; station mode returns history for a specific ICAO identifier. Examples: {"location": "Denver", "n": 3} or {"station": "KJFK", "hours": 6}.

Hazard outlooks affecting a location. hazard=severe returns SPC convective outlooks (Day 1-8 categorical risk + tornado/wind/hail probabilities); hazard=fire returns SPC fire weather outlooks; hazard=rain returns WPC Excessive Rainfall Outlook polygons (days 1-3); hazard=heat returns the NWS HeatRisk index at the point (0 none .. 4 extreme, days 1-3). include_narrative=true adds the forecaster discussion for severe (SWO), fire (FWD), or rain (QPF/QPFERD; one PIL for all days). An empty result means no outlook covers the point -- not a failure. Examples: {"location": "Moore, OK", "hazard": "severe", "include_narrative": true} or {"location": "Phoenix", "hazard": "heat"}.

Aggregate a weather variable over one or more time periods. Returns server-computed totals, maxima, minima, or averages per period. Period start/end times should use the user's local timezone boundaries (not UTC midnight). Response includes the converted value and unit per period. Ideal for questions like "total rainfall today and tomorrow" or "peak wind speed this weekend". Accepts a place name directly. Example: {"location": "Portland, OR", "variable": "precipitation", "aggregation": "sum", "periods": [{"start": "2026-07-08T07:00:00Z", "end": "2026-07-09T07:00:00Z", "label": "Today"}]}.

Current data-freshness status of the weather platform: overall state, per-source states (ok / degraded / outage / no_signal), open incidents with cause attribution (provider outage vs internal processing delay), and active provider advisories. Use this when a user asks whether data is current, when other tools return surprisingly stale data, or before presenting time-critical weather. If a source is degraded or in outage, tell the user their data may be stale rather than presenting it as live. No inputs. Refreshed about every 5 minutes.

National population-exposure headline for a risk-zone outlook product: how many people are inside risk bands at or above min_level. Powers headlines like "~57M people under major heat risk tomorrow". hazard=heat covers NWS HeatRisk days 1-3 (levels: 1 minor, 2 moderate, 3 major, 4 extreme). Pass product_id directly for other risk-zone products. Example: {"hazard": "heat", "min_level": 3}.

Resolve coordinates to a human-readable place (city, state, county, timezone). Use when you have lat/lon but need a display name or the local timezone. Example: {"lat": 39.74, "lon": -104.99} -> Denver, Colorado, America/Denver.

Complete temporal context for a location: local time, timezone, 14-day calendar with day names and Today/Tomorrow offsets, sunrise/sunset/solar times (from the weather pipeline's astro product), and moon phase. Use whenever you need to reason about dates, times, or daylight for a location -- including "what time is sunset?", "is it dark there now?", or "what day of the week is the 4th-day forecast?". Accepts a place name directly. Example: {"location": "Seattle"}.

Active NHC (National Hurricane Center) tropical systems: forecast cones, track lines, forecast points, coastal watches/warnings, and 7-day Tropical Weather Outlook formation areas -- Atlantic + East Pacific. Each feature carries a kind (cone | track | points | watch_warning | outlook_area) plus storm name, intensity, and timing properties. include_geometry=true adds full GeoJSON geometries (large). An empty result means no active tropical activity. Example: {} or {"include_geometry": true}.