Epa Drinking Water Mcp Server

Overview

EPA drinking-water records from SDWIS DMAP and the UCMR 5 publication. Resolve water systems, inspect historical violations and enforcement, and retrieve Lead and Copper Rule summaries or unregulated-contaminant occurrence measurements. Runs over stdio or Streamable HTTP without an API key.

Tools

The dataframe tools require npm or Docker with CANVAS_PROVIDER_TYPE=duckdb; drop also requires SDWIS_DATAFRAME_DROP_ENABLED=true. The Claude Desktop bundle provides the six EPA-reading tools.

The published caller guide covers evidence chains and staged analysis. Plugin, npm, Docker, and MCPB installs include it; MCP clients can read skill://epa-drinking-water-guide/SKILL.md through the Skills extension.

Capability reference

sdwis_list_reference _tool

• Required topic: identifiers, system_types, sources, contaminants, violation_types, enforcement_types, treatment, service_areas, standards, or coverage; optional substring query
• Interpretation topics are bundled; dictionaries are live. SDWIS codes and UCMR analytes remain separate namespaces; standards explain MCL, MCLG, and action levels without classifying historical records as exceedances

sdwis_search_water_systems _tool

• At least one narrowing selector: pwsid, name_contains, state, county (requires state), population_min / population_max, source, system_type, school_or_daycare, primacy_agency, or pwsid_prefix; names match literal case-insensitive substrings and counties match normalized names exactly, including punctuation
• activity defaults to active; choose inactive or all explicitly. State means reported legal-entity state; county means county served, and neither establishes complete physical service boundaries
• school_or_daycare:true selects reported Y, false selects N; unknown indicators match neither. Discovery and profile summaries preserve school/daycare and wholesaler source indicators; this is not an inventory of every school receiving public water

sdwis_get_system_profile _tool

• Required nine-character alphanumeric pwsid; section is summary (default), geography, service_areas, facilities, or treatment, independently paged
• Summary returns system with empty items and returned:0; child sections return items. found tracks inventory existence: an empty child section can still return found:true with an empty-section notice. found:false is not a compliance finding; operator contacts and mailing addresses are excluded
• Facilities resolve SDWIS treatment/violation facility_id values within the same PWSID; optional exact facility_id is valid only for this section. SDWIS and UCMR facility IDs are separate namespaces

sdwis_get_violations _tool

• Require pwsid or state; optional county, contaminant, category, health_based, and compliance_status
• Inclusive date_from / date_to select compliance-period overlap; a known endpoint can prove overlap despite a missing endpoint. Indeterminate periods are excluded with a count of affected rows examined during the call. Monitoring/reporting violations may lack concentrations
• contaminant:lead selects 1030 and Lead and Copper Rule code 5000, including copper and monitoring records; either exact code selects only itself
• stage:true requests a partition; stage_max_rows defaults to 5,000 and clamps at 25,000. Supply a held canvas_id to add another partition

sdwis_search_contaminants _tool

• Require dataset (violations, lcr, or ucmr5) and pwsid or state; UCMR also accepts zip alone without SDWIS membership requests. Additional selectors intersect; county still requires state. Optional contaminant and inclusive dates select compliance periods, LCR sample periods, or UCMR collection dates respectively
• LCR lead/copper select PB90/CU90 summaries, not household samples. UCMR PFAS covers 29 analytes; lithium is separate. zip and detected_only are UCMR-only; service ZIP is not sample location, and censored values remain unknown below reporting limits
• Supports stage, canvas_id, and stage_max_rows (5,000 default, 25,000 maximum). Cold UCMR calls may need a 180-second client timeout; broad state/county scopes resolve reported-state membership through resumable pages before returning occurrence rows

sdwis_get_enforcement_actions _tool

• Actions require pwsid or reported legal-entity state, with optional state-qualified county; geographic membership resolves before action retrieval. section:associations and any violation_id filter require pwsid; supplied geography still intersects that system
• action_type, originator, and inclusive enforcement-date filters apply only to actions. Associations page separately; source action text does not imply a federal order or monetary penalty
• Action types match the complete published code exactly and case-insensitively, including punctuation; missing dictionary labels remain null
• Supports stage, canvas_id, and stage_max_rows (5,000 default, 25,000 maximum)

sdwis_dataframe_describe _tool

• Required caller-held canvas_id; optional exact table_name; limit defaults to 10 with a cap of 500 and cursor pagination
• One whole table per item includes its columns array, row count, expiry, filters, provenance, completeness, and continuation; counts and limits are in tables. There is no global canvas listing
• Describe continuations retain their original deadline (at most 15 minutes or table expiry when issued) and reject changed membership, rows, schema, or lineage. Query-driven TTL extensions preserve cursor identity; responses report current authoritative expiry
• selection_coverage reports which source partitions and required companion relations remain staged, separately from SQL results

sdwis_dataframe_query _tool

• Required canvas_id and one read-only SELECT, WITH, or FROM statement in sql (at most 20,000 characters); describe the canvas first
• Retains at most CANVAS_DEFAULT_ROW_LIMIT (framework default 10,000), bounded by 25,000. row_count, truncated, and has_more distinguish retained rows from omitted rows; continuations read an immutable result for up to 15 minutes
• File, network, write, and system-catalog access are refused. source_complete conservatively includes every live source partition, including tables the SQL does not reference
• Known caller-correctable SQL errors preserve path-redacted diagnostics and structured details; unknown engine/database failures remain generic
• selection_coverage.chain_complete confirms the canvas holds a complete source retrieval chain; it does not establish which tables the SQL uses or override truncated. Query continuations retain the original coverage snapshot

sdwis_dataframe_drop _tool

• Required canvas_id and exact table_name; requires CANVAS_PROVIDER_TYPE=duckdb and SDWIS_DATAFRAME_DROP_ENABLED=true
• Idempotent: missing tables or expired canvases return dropped:false; other staged partitions remain available

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.

EPA-specific:

• Keyless SDWIS reads over EPA DMAP, paced at one request per second, with every returned row rechecked against the requested predicates
• Contact-free inventory and service membership; source codes kept alongside EPA dictionary labels, which stay null when EPA has none
• Quarterly SDWIS records, LCR summaries, and limited-participation UCMR 5 monitoring kept distinct; empty results establish neither current compliance, tap-water safety, nor contaminant absence
• UCMR 5 archive downloaded and indexed once per published revision on first use; censored results keep a null value with the reporting limit

Agent-friendly output:

• Provenance — effective_filters, source URL, retrieved_at, publication/HTTP validators when available, and interpretation caveats; source dates keep their calendar day, with non-midnight timestamps in a _raw field
• Completeness — source_complete:false can stay false with has_more:false once earlier rows went to other pages or partitions; aggregation does not restore it
• Source coverage — selection_partition identifies a retrieval and its delivered partition; canvas selection_coverage distinguishes a fully staged chain from missing, dropped, or expired partitions and companions
• Whole-record paging — every limit defaults to 10 with a cap of 500. Ready pages return the requested number of available whole rows; bounded source work can checkpoint earlier. Follow next_cursor with unchanged selectors, including empty phase:membership pages; source cursors expire after 15 idle minutes
• Lossless output — content[] and structuredContent carry the same values, including large rows. Byte size does not shorten pages. SDWIS_CLAMP_LIMITS=true clamps over-cap requests with a notice, log entry and counter increment; false advertises and enforces a strict maximum of 500

Getting started

No EPA account or API key is required. Add the server to Claude Code:

claude mcp add epa-drinking-water-mcp-server -- npx -y @cyanheads/epa-drinking-water-mcp-server

For other MCP clients, add this stdio configuration:

{  "mcpServers": {    "epa-drinking-water-mcp-server": {      "type": "stdio",      "command": "npx",      "args": ["-y", "@cyanheads/epa-drinking-water-mcp-server"],      "env": {        "MCP_TRANSPORT_TYPE": "stdio"      }    }  }}

For SQL staging, add "CANVAS_PROVIDER_TYPE": "duckdb" to env (Docker: -e CANVAS_PROVIDER_TYPE=duckdb), call a source tool with stage:true, and pass the returned dataset.canvas_id to sdwis_dataframe_describe and sdwis_dataframe_query. The Claude Desktop bundle fixes CANVAS_PROVIDER_TYPE=none; staging there returns canvas_disabled.

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+).
• Network access to data.epa.gov and www.epa.gov; allow a 180-second client timeout for the first UCMR call, which downloads and indexes the archive.

Installation

1. Clone the repository:

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

2. Navigate into the directory:

cd epa-drinking-water-mcp-server

3. Install dependencies:

bun install

4. Configure environment (optional):

cp .env.example .env# set CANVAS_PROVIDER_TYPE=duckdb to enable SQL staging

Configuration

See .env.example for the full overrides.

Running the server

Local development

bun run rebuildbun run start:stdio# or bun run start:httpbun run devcheckbun run test

Docker

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

The image defaults to Streamable HTTP on port 3010 with stateless sessions. Add -e CANVAS_PROVIDER_TYPE=duckdb for SQL staging.

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 logging; source results never go to ctx.state or user storage
• Register new tools in the createApp() array in src/index.ts
• Wrap external API calls: validate raw → normalize to domain type → return output schema; never fabricate missing fields, and never infer compliance, household exposure, or exceedances

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.