gribstream.com

Use tools labeled "NO AUTH / PUBLIC / READ-ONLY." for discovery, metadata, selector lookup, expression help, validation, and request building only; those tools do not return weather values. Use tools labeled "AUTH REQUIRED / DATA QUERY." when the user asks for actual GribStream weather values, CSV/JSON/NDJSON/Parquet data, plots, heatmaps, or data analysis. When a live query returns a resource_link, the exact signed HTTPS URL is the result: download it directly once with an available HTTPS or file-download capability, without an API token. This consumes the selected GribStream MCP result; it is not a direct GribStream API call or a connector bypass. Do not look for an MCP resources/read tool, and do not repeat the query merely to obtain the file bytes. If the client reports that network or allowlist policy blocks the returned resource_link, do not retry URL delivery. For CSV, JSON, or NDJSON whose reported size is at most 10000000 bytes, retry the same request at most once with delivery=inline. Parquet cannot use inline delivery. If the artifact is larger, reduce the request using the documented split order before one inline CSV retry. Time ranges are half-open: the start is inclusive and the end is exclusive. The start must be earlier than the end; equal bounds select no times. Use timesList for one exact valid time on /timeseries or one exact model run time on /runs. For a common natural-language weather concept such as 2 metre temperature or 10 metre wind, call gribstream_resolve_shared_parameter with the matching shared code before guessing a dataset-native parameter code. Use gribstream_list_shared_parameters only when the shared code is uncertain. Do not set asOf unless the user explicitly asks for backtesting, time travel, what was known then, or a historical model-run cutoff. Omit asOf for latest or best-available forecasts because it filters out forecast runs newer than the cutoff and can make recent data look missing. For dense grids or large tabular results, use delivery=url so MCP returns a short-lived signed HTTPS resource link instead of putting the full dataset into tool output. Prefer application/vnd.apache.parquet only when a compatible Parquet reader is already known to be available. If reader availability is unknown, do not spend tool calls probing or installing packages; use text/csv with delivery=url immediately. Use one grid request, process the downloaded file locally, and never print or serialize the full file into chat or tool output. Do not install a dependency or write a Parquet parser solely for one result. When a real API request limit requires splitting, split by time range first, split variables second, and split coordinates/grid/space last; spatial tiling creates the most extra work. Live query responses are streamed from parallel extraction and are not guaranteed to be sorted; sort rows before charting, building complete time series, or detecting missing timestamps. This hosted mixed-auth connection advertises live /timeseries and /runs tools, but requires MCP OAuth before executing them.