MCP server exposing Fahsai's public fire, air quality, weather, and power plant data for mainland Southeast Asia and surrounding areas
Strong foundation with 14 well-named, verb-prefixed tools and comprehensive descriptions (avg 180 chars, well above 34-char baseline). All tools have input/output schemas with Zod validation. Descriptions are LLM-optimized with dependency hints ('call get_latest_date first'), confidence filters explained, and data-source distinctions (model vs measured). However, parameter descriptions are sparse, most params lack inline guidance on format/range/constraints. Error handling is minimal (no recovery guidance, no actionable error messages visible). No tool annotations (readOnlyHint, destructiveHint). Output schemas documented but response shaping could be more explicit about field semantics.
Resolves a place name to coordinates via Nominatim (OpenStreetMap) — returns the matched lat/lng, the place name as Nominatim itself resolved it (so you can confirm it matched what the user meant), and a default bounding box around it. Every other location-aware tool already resolves a `place` param internally, so only use this directly when you need the coordinates themselves rather than data for that location.
Gridded PM2.5 estimates from the CAMS atmospheric model (via Open-Meteo) for a single date, filtered by place name or bounding box — this is a MODEL ESTIMATE, not a ground-station measurement (see get_station_readings for actual measured values; don't conflate the two). By default returns an area summary (mean, median, and p95 PM2.5 across the grid, each with its own AQI category), not the raw grid — set `include_raw_grid` to also get individual grid points (evenly, spatially sampled; capped at 500 points even when requested, with a note if the grid was truncated). Omitting both `place` and `bbox` summarizes Fahsai's full coverage area, not any one place. If the caller means "today"/"latest" without a specific date, call `get_latest_date` first — a 404 here often means the date hasn't finished ingesting yet, not that no data exists.
Daily PM2.5 time series from the CAMS atmospheric model (via Open-Meteo) for a date range — this is a MODEL ESTIMATE, not a ground-station measurement (see get_station_readings for actual measured values; don't conflate the two). NATIONWIDE ONLY: this route has no place/bbox scoping on the Fahsai API side, so do not pass a location to this tool — results always cover the full default coverage area. Each day's value is the CAMS daily p95 PM2.5, with its own AQI category. Max 30 days per request (enforced client-side); narrow the range and retry if rejected. If the caller means "today"/"latest" without a specific date, call `get_latest_date` first — a 404 here often means the date hasn't finished ingesting yet, not that no data exists.
Parameter descriptions missing or minimal. Tools like get_fires accept 'confidence', 'place', 'bbox', 'limit' but descriptions do not explain valid values (low/nominal/high for confidence), format (lat/lng bounds for bbox), or constraints (max limit). LLMs cannot infer these from names alone.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). All 14 tools are read-only but this is not declared in the tool definition. Agents cannot distinguish safe tools from destructive ones without explicit hints.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 43 | 2025-06-18+ | v2 |
Active fire detections (NASA FIRMS) for a single date, filtered by place name or bounding box. Use the `confidence` filter (`low`/`nominal`/`high`) to cut FIRMS noise — this is the field for that, not FRP. Returns a total count, a confidence breakdown, and either the full point list or the top fires by fire radiative power (FRP) when the result is large. If the caller means "today"/"latest" without a specific date, call `get_latest_date` first — a 404 here often means the date hasn't finished ingesting yet, not that no data exists.
Active fire detections (NASA FIRMS) across a date range (max 30 days), filtered by place name or bounding box. Use the `confidence` filter (`low`/`nominal`/`high`) to cut FIRMS noise — this is the field for that, not FRP. Returns a total count, a confidence breakdown, and either the full point list or the top fires by fire radiative power (FRP) when the result is large. If the caller means "today"/"latest" without a specific date, call `get_latest_date` first — a 404 here often means the date hasn't finished ingesting yet, not that no data exists.
The most recent date with complete data across fires, CAMS, and ground station readings. Call this first when a caller means "today"/"latest" without a specific date — a 404 from `get_fires`, `get_weather`, or `get_cams` on a bare "today" often just means that date hasn't finished ingesting yet, not that no data exists.
Power plants (WRI dataset: fuel type, capacity in MW, owner, commissioning year) across Fahsai's coverage area (mainland Southeast Asia and surrounding border regions), filtered by place name or bounding box. The underlying dataset is global, so results are always scoped to Fahsai's coverage area even when neither `place` nor `bbox` is given. By default returns a summary (count, breakdown by country and fuel type, and the top plants by capacity) rather than the full list — set `include_all` to get every matching plant instead (capped).
Structured scientific context (transport trajectory, upwind fire pressure, peer stations, 7-day trend, seasonal baseline) for the nearest station's PM2.5 reading near a place or bounding box, for reasoning about *why* a reading looks the way it does. If `station_id` is given (e.g. from a prior get_stations or get_station_history call), it is used as an exact match with no distance/cutoff logic, taking precedence over `place`/`bbox` if both are somehow given. This is not a pre-written explanation — it is data for the calling LLM to reason over. Several fields are conditionally null, not missing data: `transport` is null for OUTLIER_HIGH/OUTLIER_LOW `explainCase` values (an outlier reading has no plausible transport path by definition); `stationBaseline` is null when the reading is unremarkable for the season; `trend` is null when there is not enough history; `peers`/`outlier` are null depending on data availability. Do not treat any of these nulls as a bug — they are meaningful signal.
A station's calendar-day PM2.5 climatology (median/p25/p75 per day-of-year, aggregated across years, each value with its own EPA AQI category) — use this to answer "is this normal for the time of year" questions. Requires a `station_id` from `get_stations` or `get_station_readings` — a place name is not valid input here. By default returns the current season's min/median/max plus today's specific day-of-year stats; pass `month`+`day` together for a specific day instead, or `full: true` for the entire 365-day curve (large — only ask for this if you need the whole shape). Any day/aggregate built from fewer than 30 samples is flagged `thin: true` — treat it as statistically weak.
Daily rollup for a single station — one row per calendar day with mean PM2.5 (EPA AQI category attached), reading count, weather (wind formatted as compass labels, precipitation, humidity), and that day's baseline stats (median/p25/p75, each with its own AQI category), where available. Requires a `station_id` from `get_stations` or `get_station_readings` — a place name is not valid input here. `days` (default 7, max 30) sets how many days back from `date` (default: latest available) to return. A day with no PM2.5 reading ingested yet reports `pm25`/`aqiCategory` as null rather than a misleading zero; `weather`/`baseline` are null when not available for that day.
Latest PM2.5 ground-station measurements (OpenAQ) for a place or bounding box, each with an EPA AQI category. This is a measurement, not a model estimate — see `get_cams` for the gridded model instead. Per-station `attribution`, when present, carries source-specific requirements beyond the blanket OpenAQ CC BY 4.0 attribution and must not be dropped.
Raw PM2.5 reading time series for a single station, each point with an EPA AQI category — use this for "when exactly did it spike" questions that a daily rollup can't answer. Requires a `station_id` from `get_stations` or `get_station_readings` — a place name is not valid input here. `hours` (default 24, max 720) sets how far back to look. Note: intraday granularity depends on the station's data provider and isn't guaranteed — some stations report at most once per day.
All PM2.5 ground-monitoring stations (OpenAQ) for a place or bounding box, with id, name, coordinates, country, and provider. This is the way to discover the `station_id` values needed by `get_station_readings_history`, `get_station_history`, and `get_station_baseline` — call this first to find a station near a location, then pass its `id` as `station_id` to those tools. `name` does a case-insensitive substring match on station name; combined with `place`/`bbox` it narrows within that area, used alone it searches Fahsai's full coverage area.
Gridded weather (wind, relative humidity, precipitation) for a single date, filtered by place name or bounding box. Wind direction is always the direction wind is coming FROM — see the `wind` field's `fromLabel`/`toLabel`. The underlying grid can span thousands of points, so by default this returns a coarse regional summary (a bbox-wide `summary` plus per-region `cells`), not every raw point — set `include_raw_points` to get individual grid points instead (capped, with a note if truncated). If the caller means "today"/"latest" without a specific date, call `get_latest_date` first — a 404 here often means the date hasn't finished ingesting yet, not that no data exists.
Error handling not visible in source. No recovery guidance, no actionable error messages, no categorization of retryable vs fatal errors. When a tool returns 404 or validation error, LLM has no guidance on next steps.
Output schema documentation incomplete. While Zod schemas exist, the response shaping (buildToolResponse) does not document field semantics, nullable fields, or pagination structure. LLMs cannot plan downstream calls without knowing what fields are present.
Pagination not visible in tool definitions. Tools like get_fires, get_stations, get_power_plants may return large result sets but no limit/offset/cursor parameters are documented. Risk of context window exhaustion.