Developer API Beta

👋 What you get

Dublin Bay buoy readings + marine forecasts, on tap. Same dataset, two surfaces:

• MCP endpoint — for Claude Code, Cursor, or any LLM agent. 8 curated marine-data verbs; the server holds the key for you. Jump to MCP →
• PostgREST API — for apps, scripts, dashboards. REST + RPCs with the anon JWT. Details below.

Backed by 27k+ hourly buoy readings since 2023-01-01, 400k+ forecast rows across ECMWF IFS and EWAM (refreshed every 30 min), and 5 RPCs for the heavy queries. No sign-up for read access.

🤖 Use it as an MCP server

Point your MCP-capable client at https://mcp.dublinbaybuoy.com/ (streamable-HTTP transport) and you get 8 named marine-data tools. No signup, no token, no config surface — the server holds the anon key for you.

Claude Code / Claude Desktop / Cursor / Windsurf

Add to ~/.claude.json (or a project-scoped .mcp.json — same shape for Cursor, Windsurf, and other MCP-capable IDEs):

{
  "mcpServers": {
    "dublin-bay": {
      "type": "http",
      "url": "https://mcp.dublinbaybuoy.com/"
    }
  }
}

Or one-liner on the Claude Code CLI:

claude mcp add --transport http dublin-bay https://mcp.dublinbaybuoy.com/

Restart the client — tools appear as mcp__dublin-bay__*.

Python MCP SDK

from mcp.client.streamable_http import streamablehttp_client
from mcp import ClientSession

async with streamablehttp_client("https://mcp.dublinbaybuoy.com/") as (r, w, _):
    async with ClientSession(r, w) as s:
        await s.initialize()
        tools = await s.list_tools()
        result = await s.call_tool("get_current_conditions", {})
        print(result)

Tools exposed

Forecast variables: wind_speed_10m, wind_gusts_10m, wind_direction_10m, wave_height, wave_period, wave_direction, sea_surface_temperature. Models: ecmwf_ifs025, ewam.

Try these prompts

• “What's the wind doing in Dublin Bay right now?” → get_current_conditions
• “How accurate is ECMWF's 72-hour wind forecast for Dublin Bay over the last month?” → get_forecast_skill
• “What direction does the wind usually blow from here in winter?” → get_wind_rose

Why curated verbs, not raw SQL?

LLMs burn tokens and get creative with SQL. Named verbs like get_wind_rose(days=30) are cheaper, safer, and self-documenting. The server holds the anon key, so clients never rotate tokens. Read-only by design — RLS on every table, SECURITY DEFINER on every RPC, no write paths exposed.

Self-host it

The source is ~200 lines of FastMCP + httpx, trivially portable to any Supabase project:

cd dublin-bay-bot/mcp
uv pip install --system -r requirements.txt
SUPABASE_URL=https://api.dublinbaybuoy.com \
SUPABASE_ANON_KEY=<anon key> \
PORT=8080 \
python server.py

🔑 Base URL & key

All endpoints are served from:

https://api.dublinbaybuoy.com/rest/v1/

Pass the public anon key on every request. It’s safe to embed in client code — RLS limits it to SELECT on public tables:

apikey: eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJyb2xlIjoiYW5vbiIsImlzcyI6InN1cGFiYXNlIiwiaWF0IjoxNzc2NDU3MzAxLCJleHAiOjE5MzQxMzczMDF9.4ee1fobnsvjyRO14xKTi3rNhBv-UTeQRlQHMIp5Oqmw

🗂️ Tables

Anonymous SELECT is allowed on these tables. Use standard PostgREST query syntax — select, order, limit, range filters like timestamp=gte.2026-01-01.

Units

• Wind & gusts — knots (buoy and forecast agree)
• Direction — degrees true (meteorological: from)
• Waves — metres / seconds
• Temperature — °C
• All timestamps — UTC

🧮 RPCs

Server-side functions for the heavy stuff. Call with POST /rest/v1/rpc/<name> and a JSON body of arguments.

rpc_summary(days int)

Min/mean/max for every variable over the window. One-call dashboard feed.

rpc_wind_rose(days int default null)

Wind-direction histogram in 16 buckets of 22.5°, zero-padded. Suitable for polar charts. Omit days for the full archive.

rpc_readings_downsampled(hours int)

Bucketed averages for long time-series charts. Hourly buckets for up to 30 days (hours ≤ 720); daily buckets beyond that.

rpc_forecast_drift(target_hours int)

Each model’s single best forecast for a target lead-time (e.g. “what does each model say about the weather 72 h from now?”).

rpc_forecast_drift_snapshots(variable_name, model_name, requested_leads int[])

Multiple forecast runs at progressively older leads, each with its full hourly curve. Powers the Mini App’s drift visualisation — plot how the forecast for the coming week has evolved.

rpc_forecast_skill(variable_name, model_name default null, window_days default 30) Beta

Joins forecast_log against readings at target_at_utc and returns MAE, RMSE and bias per 6-hour lead bucket. Direction variables use short-arc circular error. See the spec section below.

📈 Forecast-skill endpoint (spec)

Answers: how well does ECMWF predict Dublin Bay wind 24 h out? 72 h? A week?

Request

POST /rest/v1/rpc/rpc_forecast_skill
{
  "variable_name":    "wind_speed_10m",  // or wind_direction_10m, wave_height, ...
  "model_name":       "ecmwf_ifs025",    // optional — null = both models
  "window_days":      30,                // lookback, default 30
  "lead_bucket_hours": 6                 // bucket width, default 6
}

Response

[
  {"model":"ecmwf_ifs025","variable":"wind_speed_10m","lead_bucket":0,  "n":168,"mae":1.02,"bias":-0.11,"rmse":1.38},
  {"model":"ecmwf_ifs025","variable":"wind_speed_10m","lead_bucket":24, "n":167,"mae":1.84,"bias":-0.22,"rmse":2.41},
  {"model":"ecmwf_ifs025","variable":"wind_speed_10m","lead_bucket":72, "n":165,"mae":3.12,"bias":+0.04,"rmse":3.95},
  ...
]

How it’s computed

• Every forecast row in the window whose target_at_utc matches an observed readings.timestamp becomes a (forecast, observed) pair.
• Pairs are grouped by lead_bucket = floor(lead_hours / lead_bucket_hours) * lead_bucket_hours.
• MAE = mean of |forecast − observed|. For direction variables, short-arc distance: min(|f−o|, 360−|f−o|).
• Bias = mean signed error. For direction, signed short-arc (NULL is returned instead when circular-bias is meaningless).
• RMSE = √(mean of err²) using the same error definition as MAE.
• Anon-readable, stable, 30-day rolling lookback default. Full spec in supabase/migrations/20260423140000_rpc_forecast_skill.sql.

Variable ↔ observation mapping

⚖️ Fair use & attribution

• Soft limit: ~60 requests/min on the anonymous key. Burst is fine; sustained load gets noticed.
• Please credit: “Data from Dublin Bay Buoy (dublinbaybuoy.com), sourced from Irish Lights MetOcean and Open-Meteo”.
• Data is provided as-is. Not for navigation.
• Questions, quota requests, or issues — [email protected].