PuntersEdge MCP server

What it is

A small Python process that speaks MCP over stdio and forwards each tool call to the same REST endpoints documented on this site. It adds nothing the HTTP API does not have; what it adds for an assistant is a typed tool surface with costs in the descriptions, the credit balance returned after every call so the assistant can budget without spending a request, and a read-only design — nothing that rotates keys, touches billing or places a bet is exposed.

Install

Either install it, or let uvx fetch and run it on demand:

pip install puntersedge-mcp

# or, with uv installed, no install step at all:
uvx puntersedge-mcp

You need an API key. Sign up at /api — the free tier is 1,500 credits a month with no credit card. The demo tool works without a key, so a host can be wired up and tested before the key arrives.

Configure your client

Every MCP host takes the same three facts: a command, optional arguments and an env block carrying PUNTERSEDGE_API_KEY. The snippets use the installed puntersedge-mcp command; to run through uvx instead, set "command": "uvx" and "args": ["puntersedge-mcp"].

Claude Desktop

Edit claude_desktop_config.json (Claude → Settings → Developer → Edit Config) and restart Claude Desktop:

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

Claude Code

One command from the project directory:

claude mcp add puntersedge -e PUNTERSEDGE_API_KEY=your-key-here -- puntersedge-mcp

Cursor

Create .cursor/mcp.json in the project (or ~/.cursor/mcp.json for every project):

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

Any other MCP host

The same JSON shape works wherever a host accepts a stdio server. Keep the key in the env block or the process environment rather than in a prompt; the server reads it once at start-up.

Tools

Twenty-six tools, all read-only. Costs are in API credits and are the amounts the API bills for the underlying endpoint — a next-to-go call is 2 credits, the same as calling /v1/racing/next-to-go over HTTP. Each name links to the endpoint's reference page, which carries the full parameter list and response shape.

sports_odds is billed one credit per market type requested, which is why its row carries no single figure. Every priced response is wrapped as {"data": ..., "credits_remaining": "..."}, the second value coming from the API's X-Credits-Remaining header. A 402 means the monthly allowance is spent; it is a hard stop with no overage billing, and the tool result says so rather than failing silently.

Notes the server gives the assistant

• Racing lives under the racing_* tools, not a sport key. Categories are horse, greyhound and harness.
• Prices are decimal odds. Every racing quote carries its own age_seconds; every sports market carries a quality object. The assistant is told to trust those over assumptions.
• Settle only against status: "final" results — interim placings can change on protest.
• runner_ref in results is a stable registry identifier that persists across meetings; use it, not the name, for longitudinal joins.
• Omit country on next-to-go unless you need a filter: late-card races can be country-unresolved and a filter would drop them.

Prefer a prompt-only route?

If you cannot run a local process — a custom GPT, a hosted agent, a browser-only assistant — Use the API from ChatGPT, Claude, Cursor and Copilot covers the OpenAPI-import and llms.txt paths, and includes a worked prompt.