Tools

Five tools for querying USGS water data, plus two for SQL analytics over the DuckDB-backed canvas dataframes that water_get_series and water_find_sites materialize:

water_list_parameters

Static lookup of well-known USGS parameter codes — no network call, instant response.

• Discover that 00060 = Discharge (ft³/s), 00065 = Gage height (ft), 00010 = Temperature (°C), 72019 = Depth to water level (ft), and more
• Filter by thematic domain: streamflow, groundwater, temperature, meteorological, water-quality, or all
• Use this first — parameter codes are required by every other water tool

water_find_sites

Discover USGS monitoring sites before calling data tools — all other tools require a site number.

• Geographic scoping: bounding box ("west,south,east,north" decimal degrees), 2-letter state code, bare 5-digit FIPS county code (e.g. 51013), or HUC watershed code — either a 2-digit major HUC (02) or an 8-digit minor HUC (02070008), the only two lengths NWIS accepts
• Site type filtering: ST (stream), GW (groundwater well), LK (lake/reservoir), SP (spring), and more
• Parameter filter: only return sites that have data for a specific parameter code — comma-separate to require several (e.g. 00060,00065)
• Data type filter: require sites with real-time (iv), daily (dv), or groundwater (gw) data
• Returns site number, name, coordinates, type, state/county/HUC codes, and drainage area (expanded mode only) — altitude is included in both modes when USGS records it
• Bounded result set: capped at 500 sites inline, with a truncated flag and upstreamTotal (the full upstream count) so an oversized query never overflows the response
• DataCanvas spillover: when the result is truncated and CANVAS_PROVIDER_TYPE=duckdb is set, the full match set is staged to a DuckDB-backed canvas — the response includes canvas_id and table_name to retrieve every match past the 500 cap via water_dataframe_query. Without DataCanvas, narrow the query with additional filters (county, HUC, bbox, parameter, data type) to bring the result under the cap

water_get_readings

Get the latest instantaneous (~15 min) values for one or more USGS monitoring sites.

• Batch up to 100 site numbers in a single call
• Accepts any parameter code discoverable via water_list_parameters
• Configurable lookback period via ISO 8601 duration (e.g. PT2H = last 2 hours, P7D = last 7 days)
• Returns per-site, per-parameter records with timestamp, value, unit, and provisional/approved qualifier
• Bounded by design: each series carries its 10 most recent records, with totalValues reporting how many the period actually held and truncated flagging the cap. Use water_get_series when you need the full series
• Partial batches are explicit: requested sites NWIS returns no data for are named in missingSites rather than dropped silently
• Groundwater depth available via parameterCd=72019 (the legacy gwlevels endpoint was decommissioned November 2025 — use the IV service instead)

water_get_series

Get a historical time series for a site and parameter over a date range.

• Daily values (DV service, one value per day) or instantaneous values (IV service, ~15 min resolution)
• Returns site name, parameter name, unit code, and time-ordered value records with qualifiers
• DataCanvas spillover: large date ranges (>500 records) automatically spill to a DuckDB-backed canvas when CANVAS_PROVIDER_TYPE=duckdb is set — response includes canvas_id and table_name for follow-up SQL via water_dataframe_query
• Without DataCanvas, returns the most recent 500 records with a truncated flag and totalRecords count
• Supports chaining: pass a prior canvas_id to append data to an existing canvas

water_get_conditions

Get current hydrologic conditions placed in full historical context.

• Fetches the current IV reading and the full daily percentile table in parallel
• Classifies the reading: record-high (≥ p95), above-normal (p75–p95), normal (p25–p75), below-normal (p10–p25), low (p05–p10), record-low (< p05)
• Pairs each class with a percentileLabel spelling out the threshold — record-high and record-low mark percentile-of-record extremes, not verified all-time records, and the label says so where the class name does not
• Ranks against the observation's own calendar day, so a reading near midnight is not compared against the neighboring day's percentiles
• Discloses the granularity approximation in comparisonBasis: the reading is instantaneous while the percentiles are approved daily-mean values, so the class is a "how unusual is this" ranking — not a flood-stage or drought determination, which need authoritative thresholds this tool does not fetch
• Validates site and parameterCd at the schema edge; a well-formed value NWIS still rejects surfaces as the typed invalid_request reason rather than an opaque upstream error
• Gracefully degrades when historical context is missing: returns the current reading with historicalContext: null and a historicalContextStatus saying why — no_record (new/short record), no_matching_day (no row for the date), or unavailable (stat call failed — transient and retryable, kept distinct from a sparse record)

water_dataframe_describe / water_dataframe_query

In-conversation SQL analytics over the dataframes that water_get_series and water_find_sites materialize on a DuckDB-backed canvas — time-series tables from the former, full site match sets from the latter.

Workflow:

1. Call water_get_series with a large date range, or water_find_sites with a query that matches more than 500 sites — when DataCanvas is enabled, the response includes canvas_id and table_name
2. Call water_dataframe_describe with the canvas_id to confirm the table schema — series tables carry date_time, value, qualifiers, site_number, parameter_cd, unit_code; site tables carry site_number, site_name, site_type, latitude, longitude, huc_cd, and the expanded fields
3. Call water_dataframe_query with the canvas_id and a SELECT statement to run aggregates, filter, or join

Read-only by default — only SELECT statements are permitted. Results are capped at 10,000 rows; a query matching more comes back with truncated: true. Requires CANVAS_PROVIDER_TYPE=duckdb in the server environment.

Resources and prompts

All resource data is also reachable via tools. Use water_find_sites for geographic site discovery.

Features

Built on @cyanheads/mcp-ts-core:

• Declarative tool and resource definitions — single file per primitive, framework handles registration and validation
• Unified error handling — handlers throw, framework catches, classifies, and formats
• Pluggable auth: none, jwt, oauth
• Swappable storage backends: in-memory, filesystem, Supabase, Cloudflare KV/R2/D1
• Structured logging with optional OpenTelemetry tracing
• STDIO and Streamable HTTP transports

USGS NWIS–specific:

• Wraps NWIS IV (instantaneous), DV (daily), site, and stat endpoints — no API key required, fully public
• Input formats checked at the edge against what NWIS actually accepts — site numbers, parameter codes, ISO 8601 periods, HUC, FIPS county, state, and bbox all carry a validated pattern that is advertised in each tool's JSON Schema, so malformed values fail with a pointed message instead of an opaque upstream 400
• HTML error detection: NWIS returns 400 with an HTML body for bad inputs; the service layer extracts NWIS's own message — which names the field it rejected — and maps it to a typed failure
• Multi-site batching: water_get_readings accepts up to 100 site numbers in one call
• Provisional vs. approved data qualifiers surfaced on every reading — not hidden from callers
• DataCanvas spillover: water_get_series (long date ranges) and water_find_sites (match sets past the 500-site cap) stage the full result as a DuckDB-backed table queryable via water_dataframe_query
• Groundwater via the IV service using parameter 72019 — the legacy gwlevels endpoint was decommissioned November 2025

Agent-friendly output:

• Percentile classification on every conditions response — callers get a percentileClass string (record-high, normal, record-low, etc.) they can act on directly without parsing numeric thresholds, plus a percentileLabel stating the threshold in plain language so the record-* classes are not mistaken for verified all-time records
• Partial success on conditions: when percentiles are missing, the current reading still returns with historicalContext: null and a historicalContextStatus that separates an empty stat table (no_record / no_matching_day) from a failed stat call (unavailable, transient), rather than collapsing both into an error
• Partial success on batches: water_get_readings returns the series it got and names the rest in missingSites, so a silently dropped site never reads as a complete answer
• Truncation signals: water_get_series reports totalRecords and truncated, water_find_sites reports upstreamTotal and truncated, and water_get_readings reports per-series totalValues plus truncated, so callers know when a preview is incomplete. canvas_id / table_name tell them exactly how to retrieve the rest
• Structured content and rendered text agree: every cap and count a tool applies is reported identically in structuredContent and in the markdown, so neither class of client sees a different answer

Getting started

Public Hosted Instance

A public instance is available at https://usgs-water.caseyjhand.com/mcp — no installation required. Point any MCP client at it via Streamable HTTP:

{
  "mcpServers": {
    "usgs-water-mcp-server": {
      "type": "streamable-http",
      "url": "https://usgs-water.caseyjhand.com/mcp"
    }
  }
}

Self-Hosted / Local

Add the following to your MCP client configuration file.

{
  "mcpServers": {
    "usgs-water-mcp-server": {
      "type": "stdio",
      "command": "bunx",
      "args": ["@cyanheads/usgs-water-mcp-server@latest"],
      "env": {
        "MCP_TRANSPORT_TYPE": "stdio",
        "MCP_LOG_LEVEL": "info"
      }
    }
  }
}

Or with npx (no Bun required):

{
  "mcpServers": {
    "usgs-water-mcp-server": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@cyanheads/usgs-water-mcp-server@latest"],
      "env": {
        "MCP_TRANSPORT_TYPE": "stdio",
        "MCP_LOG_LEVEL": "info"
      }
    }
  }
}

Or with Docker:

{
  "mcpServers": {
    "usgs-water-mcp-server": {
      "type": "stdio",
      "command": "docker",
      "args": [
        "run", "-i", "--rm",
        "-e", "MCP_TRANSPORT_TYPE=stdio",
        "ghcr.io/cyanheads/usgs-water-mcp-server:latest"
      ]
    }
  }
}

To enable DataCanvas for SQL analytics over large result sets (time series and site match sets), add CANVAS_PROVIDER_TYPE=duckdb to the env block in any of the configs above.

For Streamable HTTP, set the transport and start the server:

MCP_TRANSPORT_TYPE=http MCP_HTTP_PORT=3010 bun run start:http
# Server listens at http://localhost:3010/mcp

Prerequisites

• Bun v1.4.0 or higher (or Node.js v24+).
• No API key required — USGS NWIS is a free, public API.

Installation

1. Clone the repository:

git clone https://github.com/cyanheads/usgs-water-mcp-server.git

2. Navigate into the directory:

cd usgs-water-mcp-server

3. Install dependencies:

bun install

4. Configure environment:

cp .env.example .env
# Edit .env to set any optional overrides

Configuration

See .env.example for the full list of optional overrides.

Running the server

Local development

• Build and run:

  sh

  Copy

  # One-time build
  bun run rebuild
  
  # Run the built server
  bun run start:stdio
  # or
  bun run start:http
  

• Run checks and tests:

  sh

  Copy

  bun run devcheck   # Lint, format, typecheck, security
  bun run test       # Vitest test suite
  bun run lint:mcp   # Validate MCP definitions against spec
  

Docker

docker build -t usgs-water-mcp-server .
docker run --rm -p 3010:3010 usgs-water-mcp-server

The Dockerfile defaults to HTTP transport, stateless session mode, and logs to /var/log/usgs-water-mcp-server. OpenTelemetry peer dependencies are installed by default — build with --build-arg OTEL_ENABLED=false to omit them.

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
• Use ctx.log for request-scoped logging, ctx.state for tenant-scoped storage
• Register new tools and resources via the barrels in src/mcp-server/*/definitions/index.ts
• Wrap external API calls: validate raw → normalize to domain type → return output schema; never fabricate missing fields

Contributing

Issues and pull requests are welcome. Run checks and tests before submitting:

bun run devcheck
bun run test

License

Apache-2.0 — see LICENSE for details.