Use the PuntersEdge API from ChatGPT, Claude, Cursor and Copilot

Which path fits your assistant

Each needs an API key for priced calls. Sign up at /api — the free tier is 3,000 credits a month with no credit card. The three sandbox endpoints under /v1/demo/ need no key, so an assistant can be wired up before the key arrives.

The MCP server

pip install puntersedge-mcp (or uvx puntersedge-mcp), then one config block with your key in PUNTERSEDGE_API_KEY. The server exposes ten read-only tools — next-to-go racing, best odds, results, price history, movers, sports odds, a usage check and a keyless demo — and returns the remaining credit balance with every call.

{
  "mcpServers": {
    "puntersedge": {
      "command": "puntersedge-mcp",
      "env": { "PUNTERSEDGE_API_KEY": "your-key-here" }
    }
  }
}

That block is Claude Desktop's claude_desktop_config.json and Cursor's .cursor/mcp.json as-is. Claude Code takes it as a command: claude mcp add puntersedge -e PUNTERSEDGE_API_KEY=your-key-here -- puntersedge-mcp. VS Code with Copilot reads the same command and env from a servers block in .vscode/mcp.json. The full tool table with per-tool credit costs, and the snippet for each client, are on the MCP server page.

Import the OpenAPI schema into a custom GPT or agent

Use the compact schema: the API's data endpoints (racing, archive, sports, best odds, bulk files, sandbox) as one OpenAPI 3.1 document, rebuilt from the live schema whenever it is served:

https://puntersedge.online/openapi-core.json

For a ChatGPT custom GPT: in the GPT editor add an Action, choose Import from URL and paste that address. Under Authentication choose API Key, auth type Custom, header name X-API-Key, and paste your key. Every operation description fits OpenAI's 300-character Action limit, and each carries its credit cost in x-credits-cost. The same file works for LangChain's OpenAPI toolkit, for any agent framework that turns a schema into tools, and as context for a coding assistant.

The full schema, https://api.puntersedge.online/openapi.json, adds the account, key, billing and webhook routes and keeps every long description. It is the one to generate a client from, and too large to import as an Action or read into a model's context.

Two things worth putting in the GPT's instructions, because the schema cannot say them: racing is its own endpoint family under /v1/racing/, not a sport key; and a 402 means the month's credits are spent and retrying cannot clear it, while a 429 clears in sixty seconds. The errors page has the full contract.

llms.txt — context without tools

Two plain-text files are written for language models rather than browsers. Paste either into a system prompt, or give the assistant the URL if it can fetch pages:

Both are generated when served, so the figures in them are the API's own. An assistant working from llms-full.txt can write a correct request without a second crawl; combine it with Path 2 and it can also make the call.

A worked prompt

With any of the three paths set up, this prompt produces a real call:

Through the MCP server the assistant calls racing_next_to_go(num_races=5, categories="greyhound"); through an imported schema it makes the equivalent HTTP request:

GET https://api.puntersedge.online/v1/racing/next-to-go?num_races=5&categories=greyhound&country=AU
X-API-Key: your-key-here

Either way the response is an array of races, each with its runners and every bookmaker's win price per runner, so the assistant reduces to the best price itself, and the call costs 2 credits (JSON). The remaining balance arrives in the same response — as credits_remaining from the MCP server, or the X-Credits-Remaining header over HTTP — so the last part of the prompt needs no second call. The current credit cost of every endpoint is on /api/pricing.