atlas-tx

name	atlas-tx
description	Safe agent workflow for Texas drinking-water-risk and environmental-justice context across state and federal public data.

Atlas TX

Atlas TX surfaces Texas Public Water System drinking-water risk and environmental-justice burden by joining TCEQ permits + Texas Water Districts with EPA SDWIS, EPA EJScreen, and Census ACS. The skill is an agent-safe wrapper around the Atlas TX MCP server.

When to activate

Use when the user asks about:

• Texas drinking-water risk, violations, or PWS-level water quality context
• environmental-justice burden / exposure indicators in Texas (county or block-group)
• Texas county-level environmental, infrastructure, or permit context

Do not activate for:

• Non-Texas geographies
• Health-outcome causation claims
• Investment / municipal-debt advice
• Personally identifying information about individuals

Guardrails

• Indicator, not outcome. Drinking Water Risk Score is a risk indicator derived from violation history, not a measurement of present harm. EJ Burden Overlap is burden / exposure, not proof of harm.
• Cite every source. Surface every dataset_id + publisher + URL from each tool's sources envelope.
• Pass caveats verbatim. Every entry in caveats must reach the user without paraphrasing away the qualifier.
• Stamp the snapshot date. When cache_state: "snapshot", name the snapshot date.
• Bound numeric claims. Never extrapolate beyond returned rows. If a comparison isn't supported by data, say so.
• No medical / diagnostic / causal claims linking a water system or facility to specific health outcomes.
• No investor-grade or rating-agency claims about municipal entities (fiscal/debt is out of scope at v0.1).
• No naming of individuals from any dataset.

MCP tools

This skill expects the Atlas TX MCP server at packages/mcp-server/ with the v0.1.0 tool surface defined in docs/contracts/mcp-tools.md:

• discover_datasets
• get_dataset_schema
• score_pws_drinking_water_risk
• overlay_ej_burden
• summarize_water_risk_for_county

If a required tool is missing, fall back to discover_datasets + get_dataset_schema and decline the user's request with a clear explanation.

Example invocation

User: "What Texas counties have the most overlooked drinking-water risk?"

1. Call score_pws_drinking_water_risk with { limit: 25 }.
2. Aggregate scores by county. Identify counties whose mean score is high but who do not appear in well-known rankings (e.g. NRDC, EPA enforcement annual report).
3. For the top 3 counties, call summarize_water_risk_for_county with { county }.
4. For at least one of those counties, call overlay_ej_burden with { county, buffer_miles: 1 } to add demographic context.
5. Return a sourced response that:
   • names the counties, scores, and population served
   • quotes the caveats from each tool envelope
   • lists every Source (publisher + URL + retrieved_at)
   • stamps the snapshot date
   • frames the finding as risk indicators, not harm

References

• TCEQ Water Quality Permits — https://www.tceq.texas.gov/permitting/wastewater
• Texas Water Districts (Office of Water) — https://data.texas.gov/dataset/Texas-Water-Districts/hr84-s96f
• EPA SDWIS — https://www.epa.gov/ground-water-and-drinking-water/safe-drinking-water-information-system-sdwis-federal-reporting
• EPA EJScreen — https://www.epa.gov/ejscreen
• EPA ECHO — https://echo.epa.gov/
• Census ACS — https://www.census.gov/programs-surveys/acs

Stable URLs only. Don't link results pages or filtered queries — those rot.

Reproducing the example

From a clean clone:

npm install
# (M2+) start MCP server
node packages/mcp-server/src/index.js
# in another shell, run the example invocation against the local MCP

If this doesn't reproduce, the skill is not done. See docs/contracts/skill-protocol.md.