mcp-epa-echo

An MCP server that exposes the EPA ECHO (Enforcement and Compliance History Online) water-quality data as tools — NPDES-permitted facility search, permit limits, measured discharges, compliance violations, and formal enforcement cases.

It is the ECHO sibling of mcp-usgs-water-data and follows the same philosophy: ECHO's raw API is built for browsers and bulk downloads, and handing its responses straight to a model goes wrong in specific, easy-to-miss ways. This server collapses four of those traps into single, self-describing tool calls:

1. The QID two-step. Every ECHO search (get_facilities, get_cases, ...) returns a QueryID you must re-fetch with get_qid to get actual rows. Every tool here does that internally — one call in, rows out.

2. ID soup. A facility has an NPDES permit number, a separate Registry ID, and (for enforcement cases) is looked up by yet another param keyed on the Registry ID. resolve.ts hides all of it: pass a name or an NPDES permit number, get a facility back.

3. Deeply nested effluent JSON. The raw DMR (Discharge Monitoring Report) response buries values under Results → PermFeatures → Parameters → DischargeMonitoringReports. Flattened here to one tidy row per outfall × parameter × report.

4. The windowing trap. Calling ECHO's effluent-chart endpoint without a date range silently returns only the facility's current permit window — which can start as recently as 2023, making older discharges and consent orders invisible with no error, no warning, just an emptier-than-expected answer. Every tool that can hit this trap says so explicitly, every time, whether or not the current window happens to have rows.

Install / build / test

npm install
npm run build   # compiles TypeScript to dist/
npm test        # 152 tests, no network access required

Related MCP server: socrata-mcp

Configure your MCP client

The server speaks MCP over stdio. Point your client at the built entry file — .mcp.json in this repo already does:

{
  "mcpServers": {
    "epa-echo": {
      "type": "stdio",
      "command": "node",
      "args": ["/absolute/path/to/dist/src/index.js"],
      "env": {}
    }
  }
}

The config runs dist/, not the TypeScript source. Run npm run build after any code change, or the server keeps running the old build.

No API key or authentication is required — ECHO's web services are public EPA data.

The 6 tools

Every facility-taking tool accepts npdesId OR name, never both — resolution happens internally (resolve.ts), and an ambiguous name returns a note asking you to call find_facility first to get a specific npdesId.

Every tool that returns a list caps results at maxResults (default 50, max 500) and reports truncated / totalMatched so a capped response is never mistaken for the complete set.

find_facility

At least one identifying filter is required: name, city + state together, or npdesId. city alone is rejected (too broad to identify a facility). nameContains is a separate, client-side substring filter applied after ECHO's own server-side name/city search — it narrows, it does not broaden a search that came back empty. A name/city search can match more than one facility; results are never guessed down to one, and the response's note says when they're ambiguous. permitStatus ("Effective", "Terminated", "Expired", ...) is returned per facility — appearing in results does not mean a facility currently holds an active permit.

facilities_near

Exactly one spatial filter: bBox ({west, south, east, north} in decimal degrees) or lat + lon + radiusMiles together (radius capped at 25 miles). Zero, both, or a partial point set is rejected before any network call. Results include majors and minors — a dense urban area within even a 2-mile radius can return hundreds of stormwater/general-permit facilities.

get_permit_limits

Omit startDate/endDate for the facility's current permit limits (the common case). A permit reissuance changes limits, so a past date range returns a prior permit's limits, not the current one. Rows are deduped to the distinct limit set — one row per (outfall, parameter, statistical basis, limit type), since the raw data has one row per monitoring report (e.g. 24 monthly reports sharing the same limit). limitValue:null means monitoring is required but no numeric limit is assigned — not "no data."

get_dmr_values

The trap, solved. Omitting startDate/endDate returns only the current permit window, and the response always notes this — even when current-window rows come back, so a populated response is never mistaken for the full history. Passing a date range routes the same underlying ECHO call (eff_rest_services.get_effluent_chart with start_date/end_date) into returning real historical data from older permits — there is no separate historical endpoint. parameter filters by exact code ("00530") or a case-insensitive name substring ("suspended"). dmrValue:null with a populated nodiFlag means ECHO recorded a specific no-data reason (e.g. "monitoring not required"); with nodiFlag:null too, the value is simply missing. Values may be provisional and subject to revision.

get_violations

Same windowing trap, same fix, as get_dmr_values — reuses its exact note logic rather than a forked copy. A row appears only when ECHO's own compliance determination (NPDESViolations) flagged it — a NODI code alone is never treated as a violation. Two kinds show up, distinguished by violationCode/severity: reporting/monitoring violations (e.g. D80, a report submitted late or not at all) and effluent/numeric violations (e.g. E90, a reported value that exceeded its limit; exceedancePct is populated when ECHO computed one). If a date range returns rows but none are violations, that's reported as a genuinely clean result, not an error or missing data.

get_enforcement_actions

Looked up by the facility's ECHO Registry ID, not its NPDES permit number — resolved internally the same way find_facility resolves an npdesId or name. If the Registry ID couldn't be confirmed, this tool says so rather than guessing. Formal ICIS-NPDES cases: consent orders, administrative orders, penalties. Not the same record as get_violations — no 1:1 relationship either direction. startDate/endDate filter client-side against either dateFiled or settlementDate (ECHO's case search has no server-side date filter); a case with no date on file at all is excluded, not silently included, when a range is given.

ECHO quirks (hard-won, don't re-learn these)

Everything below was verified against live ECHO responses, not documentation — several plausible-looking parameter names silently no-op (return the whole database, or an unrelated empty result) instead of erroring, so getting these wrong fails quietly.

• get_qid returns a DEFAULT column set that silently OMITS fields callers read by name — RegistryID, FacLong, CWPMajorMinorStatusFlag, and CWPTotalPenalties on facility rows; StateLocPenaltyAmt (and others) on case rows. It does not error — those fields just come back missing, indistinguishable from a genuinely null value, so a tool can look like it's working while quietly returning registryId: null / lon: null for every real facility. You must pass qcolumns (a comma-separated list of ECHO column IDs) on every get_qid call to get them. This server does, via the shared CWA_FACILITY_QCOLUMNS / CASE_QCOLUMNS constants — single-sourced so every call site stays in sync. This was the single most expensive lesson of this build: it passed every offline test (fixtures were captured WITH qcolumns) and only broke live.

• Exact NPDES permit lookup: p_pid. Not p_id, p_permit, p_permitnumber, or p_npdesid — all of those silently fall through to ECHO's whole-database row-limit error.

• City filter: p_ct. Not p_city (also silently ignored).

• Name filter: p_fn — a genuine server-side substring match, not a broader "any word" match.

• Spatial bounding box: ECHO has a native bbox filter on cwa_rest_services.get_facilities — p_c1lat/p_c1lon (southwest corner) + p_c2lat/p_c2lon (northeast corner). No lat/long/radius approximation needed.

• Point-radius search: p_lat / p_long / p_radius (miles).

• Enforcement case → facility link: p_facility_id on case_rest_services.get_cases, and the value it wants is the Registry ID, not the NPDES permit number. p_id, p_pid, p_reg, p_facid, and several other guesses all silently no-op here too.

• ExceedencePct carries a literal % in the string (e.g. "32%"), not a bare number — Number("32%") is NaN. Parsed correctly here; worth knowing if you ever touch ECHO's raw JSON directly.

• Case penalties are frequently $0 in the federal FedPenalty field even for real cases — many enforcement actions are state-level, and the real penalty is in StateLocPenaltyAmt instead. Both are surfaced (penalty / stateLocalPenalty); checking only one can make a facility's entire enforcement history look penalty-free when it wasn't.

• The windowing trap is real and undocumented: eff_rest_services.get_effluent_chart called without start_date/end_date returns only the current permit window (verified: a facility with a 2023–2026 permit and zero current discharges returns empty PermFeatures, even though it has real 2011-era historical DMR data). Passing the date range through is the entire fix — there is no separate historical endpoint.

• get_qid has no default page cap observed up to at least 1815 rows with no pageno/responseset supplied — find_facility, facilities_near, and get_enforcement_actions (the tools that go through get_qid) source totalMatched from the actual rows returned rather than the summary QueryRows field, and every live check so far has shown the two agreeing.

Development

npm run build   # tsc -> dist/
npm test        # build, then node --test (no network)

Tests are hermetic: no live network calls. Fixtures under test/fixtures/ are real captured ECHO responses (a handful of specific rows are disclosed as hand-edited via an inline _fixtureNote where a live example didn't surface one), not synthesized from scratch. See DESIGN.md for the full design rationale and the field research behind it.

License

MIT.