MCP server for Claude and ChatGPT

The Model Context Protocol is a standard way for an AI assistant to call an external service. Point Claude or ChatGPT at our MCP server and racing data becomes something it can look up mid-conversation, rather than something you have to fetch and paste in.

There is nothing to install and nothing to run. The server is hosted at https://greyhoundapi.com/mcp and authenticates with the same API key you already use. The overview page covers the pitch; this guide is the detail.

Why bother

• No integration to write. The 23 tools describe themselves, so the assistant works out which one answers your question and how to fill in its arguments.
• Names resolve themselves. Ask about a dog by name and the assistant calls find_dog first to get an id, then the tool that needs it. You never see the two-step.
• It is your key. Same tier scope, same rate limits, same quota. Nothing is bypassed and nothing is billed differently.
• Good for exploring. Working out whether the data suits you takes minutes in a chat window rather than an afternoon of writing throwaway scripts.

Connect it in Claude

Settings → Connectors → Add custom connector.

1. URL: https://greyhoundapi.com/mcp
2. Authentication: None. This is the step that catches people — Claude also offers two OAuth options and either will fail to connect. This server uses an API key header, not OAuth, and Claude normally detects that and marks None as Detected.
3. Under Additional request headers, add x-api-key with your key as the value.
4. Transport: leave it. Streamable HTTP is set from the URL and is correct.

Other clients work the same way: that address, and your key in an x-api-key header. Authorization: Bearer <key> is accepted too, for clients that only offer bearer tokens.

What the assistant actually does

Two calls, both JSON-RPC 2.0 over HTTP POST. On connecting it asks what exists:

curl -X POST https://greyhoundapi.com/mcp \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

Then, when your question needs data, it calls one:

{
  "jsonrpc": "2.0", "id": 2, "method": "tools/call",
  "params": {
    "name": "find_races",
    "arguments": { "date_from": "2026-08-24", "date_to": "2026-08-24", "region": "GB" }
  }
}

The server forwards that to the public REST API with your key, so every gate that applies to a normal request applies here too. The answer comes back as text the assistant reads, prefixed with anything worth knowing — how many results were truncated, how fresh the data is, or why a list came back empty.

The tools

23 of the 62 endpoints are exposed, chosen because they match how people ask questions. A model given every endpoint picks badly, and most clients show the list in a picker. Anything not here is still reachable through the REST API.

7 of them work on a free sandbox key, within its data window. The rest return a plain-English message telling the assistant which plan is needed, rather than a raw 403 for it to misreport.

Limits

An assistant answering one question can fire several tool calls in a few seconds, which trips a per-minute limit a human never would. When that happens the server says so in words — that it is a pace limit and not an outage — so the model waits rather than telling you the API is broken. Your rate limits are otherwise unchanged.

What this is not

It returns factual race data: results, times, sectionals, form and published prices. There are no predictions, ratings or selections in it, and the server tells the assistant so on connecting. A model can still reason out loud about what it fetched — that is the model doing it, not the data, and it is not betting advice.