Rain Check

01

Connect it while you read this.

Streamable HTTP, stateless, no key. It is POST only: the server holds no session, so GET and DELETE answer 405 by design rather than opening a stream that never closes.

claude mcp add --transport http rain-check https://rain-check-mu.vercel.app/mcp

For anything that does not speak MCP, the identical operations sit at /v1/<operation> and are described by an OpenAPI document generated from the same constants the tools use, so a custom GPT Action and a Claude client cannot get different answers to the same question.

02

Six tools.

The judgement is deliberately visible. Every threshold that produced a verdict comes back with it and can be overridden per call, so nothing is hidden from the model.

01

check_workability

For a place and a set of dates, returns a GO, MARGINAL or NO GO per day for a given trade, with the rain, wind, temperature and daylight behind the call. Weather and access are reported separately, because a day can be fine to work and still fall on a public holiday.

Will not present a verdict as a safety clearance. These are planning thresholds, they are printed in every answer, and they are overridable.

02

next_work_window

For when the question is "when can we book it" rather than "what about Thursday". Returns the next run of days that suits the trade.

03

list_trade_profiles

The thresholds each trade is judged against, so you can see what you are changing before you override anything. Concreting, painting, roofing and the rest each break on different numbers.

04

find_location

Resolves an Australian place name to coordinates, a timezone and a state, which is what decides whose public holidays apply.

Will not pick between candidates. An ambiguous suburb comes back as a question with the options listed, because Richmond exists in five states and guessing produces a confident answer about the wrong site.

05

public_holidays

State holidays for a date range, returned alongside the official state page every time.

Will not claim to cover regional show days, and says so in the answer. The data is third party and covers states. Anyone about to use it for pay needs to know that before they do.

06

business_days

Counts working days between two dates for a given state, holidays excluded.

03

Two calls I made, rather than bugs I fixed.

These probably say more about how I would work on someone's client build than the bug list does.

The obvious public holiday dataset is officially inactive and still serving plausible dates, so anything built on it keeps working and quietly goes stale. I moved off it, and made the limitation part of every response rather than a footnote, with the official state page returned alongside the data.
The Bureau of Meteorology is the source an Australian tool should want, and its own terms say you must not use, copy or share it. So the server runs on Open-Meteo under CC BY 4.0 instead.

04

What broke.

Twelve defects are logged in the repository. These are the three worth reading.

Every local test passed, and every deployed call returned 500

I wrote the transport layer against the Web Request object. It typechecked and the tests passed, because the tests built their own Request objects. The platform hands the handler a Node IncomingMessage. The fix to the code was small; the useful fix was to the test, which now stands up a real HTTP server and talks to it over a socket.
A silently ignored parameter, surfacing as a confident wrong answer

The geocoder takes countryCode, not country. The wrong name is accepted without complaint and ignored, so a search for Richmond returned Virginia, British Columbia and California, my Australia filter discarded every row, and the server told the caller no such place existed. Fixed the parameter, widened the page size, and filtered to populated-place codes so Richmond Oval stops presenting itself as a suburb.
A stateless server hanging on a request it should have refused

A GET on the MCP endpoint opens the server-to-client notification stream. Stateless, there is no session and nothing to push, so the transport returned an SSE body that never closed and the function ran to its timeout. GET and DELETE now answer 405 with the reason, and the test asserts a prompt refusal rather than a status code, so a regression shows up as a hang instead of a silent pass.

05

Check it without trusting this page.

There are no mocks anywhere in this project. npm run prove connects with the official client SDK, then re-fetches every number straight from the upstream and compares field by field, printing the URL it used. Last run: 30 of 30 values identical.

# Anthropic's own inspector, not my code
npx @modelcontextprotocol/inspector --cli \
  https://rain-check-mu.vercel.app/mcp --transport http --method tools/list

# raw JSON-RPC, no client, no session, because it is stateless
curl -s -X POST https://rain-check-mu.vercel.app/mcp \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

Or the version that needs only a browser. Open both and compare the dates: rain_mm in the first equals precipitation_sum in the second.

This server Open-Meteo

06

Limits, stated plainly.

The verdicts come from planning thresholds, not safety limits. They are shown in every answer and can be overridden per call. None of them is a safety clearance.
The forecast runs sixteen days. Dates past that come back in out_of_range with no data, and the tool tells the model in as many words not to substitute an opinion about the weather.
Public holiday data is third party and covers states, not regional show days. Every answer links the official state page.
The geocoder cannot resolve Australian postcodes at all. A search for 2150 returns nothing, so the tools ask for a place name.

Connect it while you read this.

Six tools.

Two calls I made, rather than bugs I fixed.

What broke.

Every local test passed, and every deployed call returned 500

A silently ignored parameter, surfacing as a confident wrong answer

A stateless server hanging on a request it should have refused

Check it without trusting this page.

Limits, stated plainly.