Us Drought Monitor Mcp Server

Overview

Weekly drought statistics from the U.S. Drought Monitor (USDM) data services. Snapshot an area's current or dated severity, follow weekly history, rank states and counties, and count county weeks in drought across national, state, county, HUC watershed, tribal, climate-division and River Forecast Center geographies. Large results stage as tables for read-only SQL. No API key is required.

The caller guide covers interpretation and complete analytical retrieval; the server also serves it as a skill resource.

Tools

Capability reference

drought_list_reference _tool

• topic: categories, dsci, area_types, coverage, states, counties, tribal, climate_divisions, river_forecast_centers or huc; nameContains token filter, state for counties or climate divisions, hucLevel (2, 4, 6 or 8) required for huc. Optional parentHuc selects descendants of an existing shorter HUC at level 2, 4 or 6
• Default 25 records per page, clamped to 200; continued via nextCursor with unchanged filters. Explicit pages return the effective count, fewer only at exhaustion; totalCount counts every match
• Records keep leading-zero identifiers, climate-state postal codes and source vintages; bundleVersion and topic-dependent sourceMetadata (source URLs and SHA-256 hashes) date the offline bundle. provenance.sourceUrls uses the same source selection

drought_get_current_severity _tool

• One area (areaType + aoi, plus hucLevel for HUC; aoi takes an exact identifier or exact state or "County Name, ST" name); unique suffix-free county names such as "King, WA" are accepted after exact-name matching. Optional date defaults to the latest published map; us selects both CONUS and Total
• Each snapshot carries cumulative and categorical area percent, cumulative square miles, cumulative population (populationStatus unsupported outside US/state/county), dsci, and dsciChange4Weeks / dsciChange52Weeks against exactly 28 and 364 days earlier, each with an available / missing / before_coverage status
• A date outside published coverage returns found: false with guidance
• River Forecast Centers use areaType: "river_forecast_center" with an official code (for example CBRFC) or exact name (Colorado Basin); area percentages, square miles and DSCI are supported. Unknown RFCs fail local reference resolution

drought_get_timeseries _tool

• 1–100 areas sharing one areaType and HUC level; inclusive startDate / endDate; metric area_percent (default), area_sq_miles, population_percent or population_count (US/state/county only); statisticsType cumulative (default) or categorical
• Rows carry the metric buckets plus dsci; missingAreas, missingObservations (date runs per metric and DSCI component) and per-area summary (every tied peak) cover the complete series
• Rows past the inline budget stage behind a preview, with the table in staging; optional canvas_id stages into an existing canvas. Metadata remains complete inline whenever it fits. Overflow automatically retains complete rows and metadata, including zero-observation results; metadata_pages gives each array's total, returned count and next offset

drought_compare_areas _tool

• scope states or counties (optional state, else nationwide); metric dsci (default), d2_d4 or d3_d4; optional date (default latest) and baselineDate to rank the change; direction desc (default) or asc
• Competition ranks share ties; excludedAreas names each area missing its target or baseline value; changeUnit is index or percentage_points
• limit (default 20, clamped to 200) sizes the inline preview; the complete ranking stages, with optional canvas_id reuse

drought_get_weeks_in_drought _tool

• startDate, endDate, minCategory (D0–D4) and minWeeks required; mode consecutive (default) or non_consecutive; states or one county, never both, or neither for nationwide counties
• Rows give fips, county, state, and either startDate / endDate / consecutiveWeeks or nonConsecutiveWeeks; a county qualifies when any part of it reaches the category or worse, and streaks are bounded by the requested window
• Consecutive startMayExtendEarlier and endMayContinue flag boundaries touching the effective window; they express uncertainty about continuation. The Monday end is compared with the final map Tuesday plus six days. Non-consecutive rows omit these flags
• Default 25 rows per page, clamped to 200; nextCursor stays valid five minutes with unchanged filters. Explicit pages return the effective count, fewer only at exhaustion. Counts are duration statistics, not program eligibility

drought_dataframe_describe _tool

• canvas_id from a staging response; tables with SQL column types, row counts and expiry
• With table_name, that table's full producer metadata: requested and effective ranges, units, bucket-column mapping, missing data, summaries and provenance
• Oversized metadata returns first-page prefixes with metadata_pages. To read a complete retained collection, supply table_name, metadata_field (areas, summary, missingAreas, missingObservations), metadata_offset and metadata_limit (default 20, clamped to 100); follow metadata_page.next_offset until null. Explicit pages return the effective count, fewer only at exhaustion, from the same snapshot without fetching upstream again

drought_dataframe_query _tool

• canvas_id plus one DuckDB read expression with an optional terminal semicolon; writes, multiple statements, external reads and system catalogs are refused
• A column named current_schema must be table-qualified (t.current_schema); unqualified references are refused because DuckDB can bind them as an engine-settings function
• limit (default 25, clamped to 500) and offset up to 250,000; continue next_offset with identical SQL until null. Explicit pages return the effective count, fewer only at exhaustion, regardless of cell width. SELECT * FROM <table_name> ORDER BY row_id returns the original rows in order
• countIsExact is true only when one page from offset 0 holds every row; use SELECT COUNT(*) for totals

drought_dataframe_drop _tool

• canvas_id and table_name; removes the analytical copy only, returning dropped: false when the table is already gone; an expired canvas fails as canvas_not_found
• Opt-in: absent from tools/list unless DROUGHT_DATAFRAME_DROP_ENABLED=true

Features

Built on @cyanheads/mcp-ts-core: stdio and Streamable HTTP transports, pluggable auth (none / jwt / oauth), swappable storage (in-memory, filesystem, Supabase, Cloudflare KV/R2/D1), structured logging with optional OpenTelemetry tracing.

USDM-specific:

• Fixed-origin client for the USDM data services (usdmdataservices.unl.edu): no key, at most 64 upstream requests and 64 MiB per call, three concurrent upstream requests, one retry that honors Retry-After
• Dates snap to the containing Tuesday–Monday map week; coverage starts 2000-01-04, and the latest map (usually published Thursday) is discovered from upstream
• Cumulative buckets (d1_d4 = D1 or worse) overlap; categorical buckets (d1_only) are exclusive. DSCI is upstream's experimental 0–500 index, never recomputed
• Offline geography bundle: 52 states and jurisdictions, 3,230 counties (Connecticut planning regions included), 323 tribal areas, 357 climate divisions, 13 River Forecast Centers and 2,916 HUCs, refreshed by scripts/update-reference-data.ts
• Staged tables live in DuckDB with a 30-minute sliding lifetime renewed by queries; canvas tokens are tenant-scoped, shareable capabilities, so treat them as credentials
• At the four-canvas process cap, omitted-token staging reuses the tenant's most recently used service-owned canvas with a free slot (eight tables per canvas). Explicitly supplied canvases remain caller-owned; live tables are never evicted. staging_capacity_exhausted advises token reuse or expiry, while staging_busy advises waiting for active work. The two-workflow admission bound applies before fetching, including inline reads
• Connecticut's eight legacy county FIPS remain discoverable but are omitted from comparison populations. Their empty observation lookups explain that USDM serves Connecticut by planning region; call drought_list_reference with topic: counties, state: CT. Counties are never mapped onto different planning-region boundaries; currentUsdm: false alone is not a historical-coverage rule
• Node acquisition uses a private reusable Undici HTTP/1.1 dispatcher, released during server teardown; Bun retains native fetch

Agent-friendly output:

• Reference records, query rows and describe columns use compact tables with JSON cells; empty cells distinguish absent optional fields from null values. Metadata, nested values and clickable URLs remain complete on both response surfaces
• Positive integer page/preview limits above 200 (references, comparisons, duration), 500 (SQL), or 100 (metadata) are clamped; effectiveLimit and limitNotice disclose the result. Set DROUGHT_CLAMP_LIMITS=false to advertise and enforce these maxima instead. Clamps emit a request log and the drought.limit_clamps counter. Continuation retains every record. Nonpositive and fractional limits are rejected; count ceilings do not bound response bytes
• Date inputs advertise YYYY-MM-DD and validate real calendar dates; blank optional snapshot and comparison dates retain their defaults. A duration threshold beyond its window advises lowering minWeeks or widening the dates
• Provenance on every response: sourceUrls, retrievedAt, the USDM credit line and the experimental-DSCI citation
• Supported population snapshots and population series add provenance.populationInterpretation, retained through dataframe describe/query: USDM population figures are rough census-based estimates assuming uniform population within counties, with artificial precision rather than exact contemporary headcounts. Source values remain unchanged; no census vintage is inferred
• Missing data stays explicit: nulls with dsciStatus / populationStatus, missingAreas, missingObservations date runs and excludedAreas, never zero-filled
• Range and identity echo: requestedRange / effectiveRange, latestMapDate, resolvedAreas, and staging handles that name drought_dataframe_describe and drought_dataframe_query

Getting started

Connect to the hosted instance:

claude mcp add --transport http us-drought-monitor-mcp-server https://us-drought-monitor.caseyjhand.com/mcp

Or run it locally:

claude mcp add us-drought-monitor-mcp-server -- npx -y @cyanheads/us-drought-monitor-mcp-server

For other clients, add this to the MCP configuration file:

{  "mcpServers": {    "us-drought-monitor-mcp-server": {      "command": "npx",      "args": ["-y", "@cyanheads/us-drought-monitor-mcp-server"]    }  }}

No API key or account is required. Set DROUGHT_DATAFRAME_DROP_ENABLED=true to expose table removal. For Docker or a source build, see Running the server.

Configuration

See .env.example for every server setting and the common framework overrides.

Running the server

Local development

Requires Bun v1.4.0 or higher (or Node.js v24+).

git clone https://github.com/cyanheads/us-drought-monitor-mcp-server.gitcd us-drought-monitor-mcp-serverbun installcp .env.example .env# optional: set DROUGHT_DATAFRAME_DROP_ENABLED or canvas limits

Staging and SQL need native DuckDB, included in npm and Docker installs. The portable Claude Desktop .mcpb bundle fixes CANVAS_PROVIDER_TYPE=none: producers return complete inline results within acquisition limits. Explicit canvas tokens and dataframe operations still require an analytical provider; invalid, expired or foreign tokens never fall back to inline success.

• Build and run the production version:

  sh

  # One-time buildbun run rebuild
  # Run the built serverbun run start:http# orbun run start:stdio

• Run checks and tests:

  sh

  bun run devcheck  # Lints, formats, type-checks, and morebun run test      # Runs the test suite

Docker

docker run --rm -p 127.0.0.1:3010:3010 ghcr.io/cyanheads/us-drought-monitor-mcp-server:latest

The image runs on Bun with native DuckDB, defaults to HTTP on port 3010 in stateless session mode, and logs to /var/log/us-drought-monitor-mcp-server. OpenTelemetry peer dependencies are installed by default; build with --build-arg OTEL_ENABLED=false to omit them.

To build locally, run docker build -t us-drought-monitor-mcp-server .. For stdio clients, use "command": "docker" with "args": ["run", "-i", "--rm", "-e", "MCP_TRANSPORT_TYPE=stdio", "ghcr.io/cyanheads/us-drought-monitor-mcp-server:latest"].

Project structure

Development guide

See CLAUDE.md for development guidelines and architectural rules. The short version:

• Handlers throw, framework catches — no try/catch in tool logic
• Service errors carry the reason declared in each tool's inline errors[] contract
• Register new tools in the createApp() arrays in src/index.ts
• Wrap external API calls: validate raw → normalize to domain type → return output schema; never fabricate missing fields

Contributing

Issues are welcome. Run checks and tests before submitting:

bun run devcheckbun run test

License

This project is licensed under the Apache 2.0 License. See the LICENSE file for details.

The U.S. Drought Monitor is jointly produced by the National Drought Mitigation Center at the University of Nebraska-Lincoln, the United States Department of Agriculture, the National Oceanic and Atmospheric Administration and the National Aeronautics and Space Administration. Map courtesy of NDMC.

DSCI: Akyuz, F. A. 2017. Drought Severity and Coverage Index. United States Drought Monitor.