Every published drinking water result, AI native.

Our agent or yours. Connect Muriel to Claude, ChatGPT, Cursor or any AI assistant as an MCP connector, build it into your software with the API, or pull the record for your own analysis: millions of results from dozens of states, back to the 1970s, updated as states publish.

Get started Read the docs

Not a developer? Try the Muriel agent, no code needed.

Use Muriel in your AI harness

Add Muriel to the AI tools you already use.

Two steps to bring near real-time water quality results, trends, and compliance intelligence into Claude, Cursor, VS Code, ChatGPT, or any assistant that supports remote MCP servers.

The quickest way: paste it into your AI coding assistant and it sets up the Muriel connector for you. You just sign in and click Allow. Or follow the two steps below.

Preview the prompt

Set up the Muriel drinking water MCP server for me, in the AI tool you are running in.

Server details:
- Name: muriel
- URL: https://mcp.murielai.io/mcp
- Transport: remote MCP over Streamable HTTP
- Sign-in: OAuth. I will sign in with my Muriel account in a browser window and click Allow. There is no API key, so do not ask me for one or write one into any file.

1. Work out which tool you are and add the server the way it expects:
- Claude Code: run `claude mcp add --transport http --scope user muriel https://mcp.murielai.io/mcp`, so it is available in every project. Then, inside Claude Code, run /mcp, pick muriel and choose Authenticate.
- Cursor: add "muriel": {"url": "https://mcp.murielai.io/mcp"} under "mcpServers" in ~/.cursor/mcp.json.
- VS Code (GitHub Copilot agent mode): add "muriel": {"type": "http", "url": "https://mcp.murielai.io/mcp"} under "servers" in .vscode/mcp.json.
- Any other tool: follow its documentation for adding a remote (Streamable HTTP) MCP server with OAuth. If it can only run local (stdio) servers, use the command `npx -y mcp-remote https://mcp.murielai.io/mcp`.
- If you cannot edit your own configuration (for example the Claude app or claude.ai), do not try. Tell me to open Settings, then Connectors, choose Add custom connector, name it Muriel and paste the URL.

2. When you edit a config file, merge with what is already there. Keep the existing JSON valid and do not change or remove any other MCP servers.

3. Tell me exactly what is left for me to do. Usually that is reloading or restarting the tool, then signing in to Muriel in the browser window that opens and clicking Allow.

4. Once it is connected, check it works: ask Muriel what data it has for Illinois (it should use the data_coverage tool) and show me a short summary of the answer. If the connection fails, show me the exact error and what you tried.

Add the connector

Add one URL to your assistant, then sign in to your Muriel account and click Allow. No API key to copy.

# Settings → Connectors → Add custom connector
Name   Muriel
URL    https://mcp.murielai.io/mcp

claude mcp add --transport http muriel https://mcp.murielai.io/mcp
# then, inside Claude Code: /mcp → muriel → Authenticate

// ~/.cursor/mcp.json
{
  "mcpServers": {
    "muriel": { "url": "https://mcp.murielai.io/mcp" }
  }
}

// .vscode/mcp.json
{
  "servers": {
    "muriel": { "type": "http", "url": "https://mcp.murielai.io/mcp" }
  }
}

# Any client that supports remote MCP servers with OAuth:
https://mcp.murielai.io/mcp

# A client that only runs local (stdio) servers:
npx -y mcp-remote https://mcp.murielai.io/mcp

Ask in your assistant

It calls Muriel when a question needs drinking water data, and answers from what Muriel returns.

Tool call

get_notable_events({
  "pwsid": "IL0316000",
  "since": "2026-09-29"
})

Muriel returns (trimmed)

{
  "rows": [{
    "event_type":
      "result.prior_comparison",1
    "severity": "watch",
    "system_name": "CHICAGO",
    "pwsid": "IL0316000",2
    "summary": "NITRATE at ENTRY
      POINT 01 for IL0316000:
      0.31 → 0.42 mg/L (+35.5%)",3
    "federal_limit_mgl": 10.0,
    "exceeds_federal_limit": false,
    "collection_date": "2026-09-24"
  }],
  "returned": 14
}

Your AI assistant

Anything new with Chicago’s water this week?

Used Muriel get_notable_events

One change worth knowing about:

Nitrate moved from its previous result1

Chicago · IL03160002

Nitrate at entry point 01 rose from 0.31 to 0.42 mg/L (collected Sep 24). That is still far below the 10 mg/L federal limit.3

1 event since Sep 294

Ask a follow-up…

Numbers match each field Muriel returns to where it appears in the answer. Example only: the response is the real shape, but the event and its figures are illustrative. Everything else you need is in the MCP server reference below.

Build it in

One call puts Muriel in your software.

Two steps to instant integrations of real-time water quality results, trends, and compliance intelligence. Create an API key on your account page first. For Python, pip install muriel (PyPI); for JavaScript and TypeScript, npm install @muriel-ai/sdk (npm).

Send a request

Ask for a water system’s results by its ID, from any application. Filter by contaminant, date and detections.

# pip install muriel
import muriel  # reads MURIEL_API_KEY from the environment

results = muriel.get_results(
    pwsid="IL0316000",
    analyte="Lead",
    limit=12,
)

// npm install @muriel-ai/sdk
import { getResults } from '@muriel-ai/sdk'  // reads MURIEL_API_KEY

const results = await getResults({
  pwsid: 'IL0316000',
  analyte: 'Lead',
  limit: 12,
})

curl "https://api.murielai.io/api/v1/records/system/IL0316000?result_analyte_name=Lead&limit=12" \
  -H "X-API-Key: $MURIEL_API_KEY"

Show the results in your software, or use them for analysis

Muriel returns structured results, ready to display.

Response (trimmed)

{
  "pwsid": "IL0316000",
  "system_name": "CHICAGO",1
  "record_count": 12,2
  "records": [{
    "result_analyte_name": "LEAD",
    "concentration_value": 2.1,
    "concentration_units": "UG/L",3
    "sample_collection_date":
      "2026-09-24",4
    "result_non_detect": false
  }, …]
}

Your application

DashboardSystems ComplianceReports

Samples this month1,284

Open work orders37

Active sites412

Murieljust now

Chicago · IL03160001

12 recent lead results2

Latest: 2.1 µg/L3

Collected September 24, 20264

View resultsDismiss

Numbers match each field in the response to where it appears in the app. Example only: the response is the real shape of /api/v1/records/system/{pwsid}, but the figures are illustrative.

Try it Paste it into any AI coding assistant to get a working Python script for the Muriel API.

Preview the prompt

I want to connect a Python application to the Muriel drinking water quality API.

1. Read the API documentation at https://api.murielai.io/docs and the OpenAPI spec at https://api.murielai.io/openapi.json. Use only the endpoints, parameters and response fields that appear there. If something I ask for below is not in the docs, say so rather than guessing.

2. Write one Python 3 script, muriel_client.py, that:
   - reads the API key from the MURIEL_API_KEY environment variable and sends it in the X-API-Key header (never hard-code the key);
   - has a function get_system_results(pwsid, analyte=None, limit=100, detections_only=False) that calls GET /api/v1/records/system/{pwsid}, passing analyte as result_analyte_name;
   - as a working example, fetches the 12 most recent lead results for Chicago (PWSID IL0316000) and prints one line per record: sample date, analyte, concentration value and units, or "not detected";
   - handles failures with clear messages: missing key, 401 (invalid key), 404 (no records for that PWSID), 429 (rate limited), timeouts, and 5xx errors;
   - uses only the standard library and the requests package.

3. Keep it short and readable, with a comment at the top explaining how to set MURIEL_API_KEY and run the script. When you are done, tell me the exact commands to install requests and run it.

For research and analysis

The whole record, in one shape, ready for your own analysis.

Every state’s results in the same units and terms, so a comparison across states or decades starts with the question rather than the cleanup.

History back to the 1970s

Decades of results for most states, alongside the current feed, in one dataset.

Exactly the rows you need

Use since to pull only what is new, and schema to get the columns your notebook expects.

CSV when you want it

Ask the agent for any table and export it as a CSV, no code required.

Limits stated, not hidden

Muriel says where each state’s data stops, so a gap is never mistaken for a clean record.

API access and pricing are set up with each team for now. Talk to us →

MCP server reference

Everything about the Muriel connector, in one place.

For people adding Muriel to an assistant, and for developers building an MCP client that talks to it.

At a glance

Connector URL	`https://mcp.murielai.io/mcp`
Transport	Streamable HTTP (MCP specification 2025-03-26 or later). Stateless, JSON responses.
Sign-in	Your Muriel account, through OAuth 2.1. The assistant registers itself; there is no key to copy.
Access	Read only: 16 tools over published sample results, events, trends and the federal rules.
Cost	Your Muriel credits, on the same price list as the app and the API: 1 credit per 1,000 records returned, added up across calls. A call that returns nothing is free.
Limits	30 tool calls a minute and 2,000 a day per account.
Disconnect	Account page → Connected apps → Disconnect. Takes effect within a minute.

Set up

You need a Muriel account. If you don’t have one, the sign-in screen creates it with an emailed link.

Claude (claude.ai and the Claude desktop app)

Open Settings → Connectors and choose Add custom connector.
Name it Muriel and paste https://mcp.murielai.io/mcp as the URL.
Click Connect. A Muriel page opens: sign in, then click Allow.
In a chat, make sure Muriel is turned on in the tools menu, then ask a question.

On Team and Enterprise plans an owner may need to add the connector for the organization first.

Claude Code

Run claude mcp add --transport http muriel https://mcp.murielai.io/mcp. Add --scope user to use it in every project.
Inside Claude Code, type /mcp, choose muriel, then Authenticate. Your browser opens to sign in and allow.

Cursor

Add Muriel to ~/.cursor/mcp.json (all projects) or .cursor/mcp.json (one project): {"mcpServers": {"muriel": {"url": "https://mcp.murielai.io/mcp"}}}
Open Settings → MCP, find Muriel, and click to sign in.

VS Code (GitHub Copilot agent mode)

Add to .vscode/mcp.json: {"servers": {"muriel": {"type": "http", "url": "https://mcp.murielai.io/mcp"}}}, or run MCP: Add Server from the command palette and choose HTTP.
Start the server when prompted and sign in in the browser window that opens.

ChatGPT

Where your plan allows custom connectors (developer mode), create a connector with the URL above and OAuth authentication, then sign in when ChatGPT asks. Availability varies by plan and workspace settings.

Other assistants

Any MCP client that supports remote servers and OAuth works with the URL alone. For one that only runs local (stdio) servers, use npx -y mcp-remote https://mcp.murielai.io/mcp as the command; it handles sign-in in your browser.

Test it without an assistant

Run npx @modelcontextprotocol/inspector, choose Streamable HTTP, enter the URL and connect. You can sign in, list the tools and call each one by hand.

The first time your assistant uses Muriel, it opens a Muriel page. If you are signed out you sign in first (Google, an emailed link or a password), then see what the assistant is asking for and click Allow or Cancel.

It can

Look up published sample results, trends and notable events
Find water systems and check results against federal standards
Search the federal drinking water rules
Use your credits for the data it looks up (1 per 1,000 records)

It can’t

See your Muriel conversations or documents
Change your account or buy anything
Act as anyone but you, or work after you disconnect it

Each assistant gets its own connection, and you can disconnect any of them from Connected apps on your account page. A free trial session can’t approve a connection; sign in with a real account first.

Tools

All tools are read-only and take JSON arguments. Your assistant chooses them for you; the parameters are listed for anyone writing prompts or a client. pwsid is a water system’s federal ID (for example IL0316000); state is a two-letter code; dates are ISO (2026-09-01).

Tool	Use it for	Parameters
Find and check coverage
`find_water_system`	Turning a utility name into a PWSID, with how many results it has.	`query`*, `state`, `limit`
`data_coverage`	What data exists for a state or system: row counts, analytes, date range.	`state`, `pwsid`
`historical_depth`	How far back the archive goes, before asking a long-range question.	`state`, `pwsid`
Results
`get_sample_results`	Individual results with value, unit, non-detect flag and normalized mg/L.	`pwsid`, `contaminant`, `state`, `start`, `end`, `only_detections`, `limit`
`summarize_sample_results`	Per-contaminant counts, date range and min/max/mean for a system or state.	`pwsid`, `state`, `contaminant`, `limit`
`rank_systems_by_size`	A state-wide question with no system named: one row per system, largest first.	`state`, `contaminant`, `only_exceeding`, `start`, `end`, `limit`
`deep_history_search`	Results from the full archive, decades back, beyond the current window.	`pwsid`, `state`, `analyte`, `since`, `until`, `limit`
What changed
`get_notable_events`	Recent changes worth a look: moves from a previous result, unusually high values, sample-point changes. Severity `info`, `watch` or `high`.	`pwsid`, `state`, `since`, `min_severity`, `event_type`, `limit`
`get_record_changes`	Results that appeared, disappeared or were amended between updates.	`pwsid`, `state`, `since`, `kind`, `limit`
`analyte_trend`	Whether a contaminant is rising or falling at a system’s sampling points, or across a state.	`contaminant`, `pwsid`, `state`, `min_span_days`
Compliance math
`check_compliance_from_data`	The correct statistic from real results: running annual average, locational RAA, 90th percentile.	`pwsid`, `contaminant`, `start`, `end`
`check_monitoring_gaps`	Candidate missed samples against the monitoring schedule the state publishes.	`pwsid`, `state`, `analyte`, `only_assessable`, `limit`
Rules and standards
`lookup_standard`	The federal MCL or action level, MCLG, units, averaging basis and CFR citation.	`contaminant` or `category`
`evaluate_result`	A measured value against the federal standard, with unit conversion.	`contaminant`, `value`, `unit`
`resolve_analyte`	How a lab’s analyte name maps to a regulated contaminant.	`analyte_name`*
`search_regulations`	Passages from the federal rules: monitoring, lead and copper, DBPs, PFAS, public notice.	`query`*, `top_k`

* required. The full JSON schema for each tool comes back from tools/list.

Example questions

“Anything new with Chicago’s water this week?”
“Which California systems had a high-severity event in the last two weeks?”
“Has nitrate been rising at Los Angeles DWP over the last ten years?”
“What is the running annual average for TTHM at IL0316000 this year?”
“Which of California’s largest systems have detected PFOA or PFOS, and at what levels?”
“We measured lead at 18 ppb at one tap. What does the rule require next?”
“Were any of Georgia’s published results revised since last week?”

Getting good answers

Name the system or the state. “Chicago” or “IL0316000” gets a specific answer; “my water” does not.
An empty result is not a clean result. Most often the state doesn’t publish that data, or stopped publishing. Ask your assistant to check coverage first, and see which states are current on the coverage table.
One result over a limit is not a violation. Most limits apply to an average. Ask for the compliance statistic (“running annual average”, “90th percentile”) rather than comparing one sample to the MCL.
Narrow wide questions. One system, one contaminant or a shorter date range is faster and more complete than a whole state over decades.
These are published monitoring results, not compliance determinations. The state primacy agency decides violations, and state rules can be stricter than federal ones.

Limits

30 tool calls a minute and 2,000 a day per account. Past that, the tool returns a message saying so; wait and ask again.
A single call can run for up to two minutes. A very wide search may time out with a message asking for a narrower question.
Large results are truncated, and the response says so rather than presenting a partial list as complete. Need the whole set? Muriel’s own agent exports it as CSV.

For client developers

Muriel follows the MCP authorization specification. A compliant client needs nothing but the connector URL.

MCP endpoint	`POST https://mcp.murielai.io/mcp` — stateless; each request stands alone, responses are `application/json`
Protected resource metadata	`https://mcp.murielai.io/.well-known/oauth-protected-resource/mcp` (RFC 9728)
Authorization server	`https://murielai.io` — metadata at `/.well-known/oauth-authorization-server` (RFC 8414)
Registration	Dynamic client registration, `POST /oauth/register` (RFC 7591). `token_endpoint_auth_method`: `none`, `client_secret_post` or `client_secret_basic`
Redirect URIs	HTTPS; `http://localhost` or `127.0.0.1` on any port (RFC 8252); or a private-use scheme such as `cursor://`. Matched exactly.
Grant	Authorization code with PKCE, `S256` only. The authorization response carries `iss` (RFC 9207).
Scope	`muriel.read`
Resource	Send `resource=https://mcp.murielai.io/mcp` (RFC 8707). Tokens work only at the MCP server.
Tokens	Bearer access token, one hour. Refresh token, 60 days, rotated on every use; reusing a spent one revokes the connection.
Revocation	`POST /oauth/revoke` (RFC 7009)
CORS	Allowed from any origin on the MCP endpoint and on the token, registration and metadata endpoints, for browser-based clients.

The flow: POST /mcp without a token returns 401 with WWW-Authenticate: Bearer resource_metadata="…". The client reads that metadata, finds murielai.io, registers, sends the person to /oauth/authorize, exchanges the code at /oauth/token, and retries with Authorization: Bearer. A 401 later means the token expired or the person disconnected; refresh, or start sign-in again.

Agent-to-agent (A2A)

Muriel publishes an A2A 1.0 Agent Card at /.well-known/agent-card.json and answers JSON-RPC at POST https://murielai.io/a2a. Each skill on the card is one of the tools above, with the same sign-in, limits and cost (none). Send one skill call per message as a data part, and the reply comes back at once as a message:

{"jsonrpc": "2.0", "id": 1, "method": "SendMessage",
 "params": {"message": {"messageId": "m1", "role": "ROLE_USER",
   "parts": [{"data": {"skill": "get_notable_events",
                       "arguments": {"state": "CA", "min_severity": "high"}},
              "mediaType": "application/json"}]}}}

Send the bearer token in Authorization. A2A 0.3 clients (message/send) work too. Muriel keeps no long-running tasks, and doesn’t stream or send push notifications.

Headless agents: sign up by email (auth.md)

An agent with no browser to open can register for a person by email instead. It posts the person’s address to /agent/auth. Muriel emails them a link and a short code, and the agent polls /oauth/token with the device-code grant until they sign in and click Allow. It then gets the same read-only tokens. Every request and response is in auth.md. Anonymous registration is not supported.

Troubleshooting

The sign-in page says the app “can’t connect to Muriel”.: The assistant sent a return address it didn’t register, or its registration is stale. Remove the connector and add it again.
I signed in but went back to the assistant without approving.: Start the connection again from the assistant. You’ll go straight to the Allow screen, since you’re already signed in.
It worked yesterday and now says my sign-in expired.: The connection was disconnected (perhaps from another device), or the assistant lost its refresh token. Reconnect from the assistant’s connector settings.
Claude Code shows Muriel as “needs authentication”.: Run /mcp, choose muriel, then Authenticate.
The assistant says Muriel found nothing.: Check that the state publishes the data you asked about (ask “what data does Muriel have for Ohio?”), and try the system’s PWSID instead of its name.
“Rate limit reached” or “Muriel is busy”.: Wait a minute and ask again. If you need more than the daily limit, tell us.
My company network blocks it.: The assistant needs to reach mcp.murielai.io and your browser needs murielai.io, both over HTTPS. Ask IT to allow both.

Still stuck? Write to us with the assistant you’re using and what it said.

Every published drinking water result, AI native.

Use Muriel in your AI harness

Add the connector

Ask in your assistant

Nitrate moved from its previous result1

Build it in

Send a request

Show the results in your software, or use them for analysis

Chicago · IL03160001

For research and analysis

History back to the 1970s

Exactly the rows you need

CSV when you want it

Limits stated, not hidden

MCP server reference

At a glance

Set up

Claude (claude.ai and the Claude desktop app)

Claude Code

Cursor

VS Code (GitHub Copilot agent mode)

ChatGPT

Other assistants

Test it without an assistant

Sign-in and permissions

Tools

Example questions

Getting good answers

Limits

For client developers

Agent-to-agent (A2A)

Headless agents: sign up by email (auth.md)

Troubleshooting