io.github.Propertyscout001/puntersedge

Australian Racing and Sports Odds MCP Server

A Model Context Protocol server for Australian horse racing, greyhound, harness and sports odds. Point Claude Desktop, Claude Code, Cursor, Windsurf or any MCP client at a live Australian odds feed and ask it what is racing next, what every bookmaker is paying, which runners are firming, and who won.

Racing data MCP · sports odds MCP server · horse racing MCP server · Australian odds for Claude · sports data for Cursor.

You: What's the next race at Randwick and who's favourite?

→ racing_next_to_go(num_races=3, country="AU")
← Wiesners Maiden Plate R1 — 11 runners, 8 bookmakers, data 38s old
  Favourite: <runner> at $2.40 (sportsbet) / $2.55 (tab) / $2.35 (neds)

15 Australian bookmakers on racing (measured 19 Oct 2026; live count at https://puntersedge.online/coverage-report.json). AU thoroughbred, greyhound and harness; NZ thoroughbred and harness; British, Irish and South African racing on smaller panels. AFL, NRL, NBA, NFL, tennis, cricket and more on sports. Free tier: 3,000 credits a month, no card.

There are two PuntersEdge MCP servers. You probably want the Python one.

Being straight about this, because the difference is large.

If you just want the odds inside your assistant, use the Python server. It is the fuller server, it is the one the documentation describes, and it will show you live prices before you sign up for anything.

This TypeScript server is a smaller, dependency-light implementation of the five core racing tools and four sports tools. It is real and it works — the install below is verified — but it is a subset.

Python server config, if you came here for that

{
  "mcpServers": {
    "puntersedge": {
      "command": "uvx",
      "args": ["puntersedge-mcp"],
      "env": { "PUNTERSEDGE_API_KEY": "your_key_here" }
    }
  }
}

Full instructions: https://puntersedge.online/developers/mcp-server?utm_source=github&utm_medium=readme

Install this server

Not published to npm. There is no npm package called puntersedge-mcp — install straight from this repo, which builds itself on install:

npx -y github:Propertyscout001/puntersedge-mcp

First run clones the repo and compiles it with tsc, so it takes a while and needs git plus a working Node toolchain. Node ≥18.

Get a free API key at https://puntersedge.online/api?utm_source=github&utm_medium=readme — 3,000 credits a month, no credit card.

Claude Desktop

Edit claude_desktop_config.json, then fully quit and reopen Claude Desktop:

{
  "mcpServers": {
    "puntersedge": {
      "command": "npx",
      "args": ["-y", "github:Propertyscout001/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 \
  -- npx -y github:Propertyscout001/puntersedge-mcp

Add -s user to make it available in every project instead of just this one. Claude Code writes project servers to .mcp.json in the project root, which you can commit and edit by hand — keep the key out of it and use an environment variable if you do.

Cursor

.cursor/mcp.json for one project, ~/.cursor/mcp.json for all of them:

{
  "mcpServers": {
    "puntersedge": {
      "command": "npx",
      "args": ["-y", "github:Propertyscout001/puntersedge-mcp"],
      "env": { "PUNTERSEDGE_API_KEY": "your_key_here" }
    }
  }
}

Windsurf

~/.codeium/windsurf/mcp_config.json — same mcpServers shape as Cursor.

VS Code / GitHub Copilot

.vscode/mcp.json in the workspace. VS Code uses a servers key rather than mcpServers:

{
  "servers": {
    "puntersedge": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "github:Propertyscout001/puntersedge-mcp"],
      "env": { "PUNTERSEDGE_API_KEY": "your_key_here" }
    }
  }
}

Anything else

Any host that accepts a stdio server takes the same three facts: a command, its arguments, and an env block carrying PUNTERSEDGE_API_KEY. Keep the key in the environment, not in a prompt — the server reads it once at start-up and never echoes it in a tool result.

Tools

Nine tools, all read-only. Costs are in API credits and are what the API bills for the underlying endpoint — they are not this server's invention. Figures below were read from https://api.puntersedge.online/openapi.json on 2026-09-15.

Racing

Sports

Account

Every response also carries the live X-Credits-* headers through to the tool result, so an assistant can read its remaining balance without spending a call to find out. A 402 means the monthly allowance is spent — a hard stop, no overage billing — and the tool result says so rather than failing quietly.

There is no horse-racing sport key. Racing lives under the racing_* tools. Passing sport_key="horse-racing" to a sports tool is an error.

What an assistant can and cannot do with this

Worth being explicit, because "betting odds" and "AI agent" in one sentence invites the wrong assumption.

It can:

• read a live odds feed — current per-bookmaker prices, the upcoming schedule, settled results, and which runners are firming or drifting
• tell you how fresh each number is, per price, not per page
• state what each call costs in credits before it makes it, and report the balance after

It cannot:

• place a bet. There is no tool here that places, stakes, cancels or settles anything. Every tool is an HTTP GET against a read-only endpoint.
• hold or use bookmaker credentials. This server has no bookmaker account, no login, no session with any bookmaker. It reads published prices. The only secret it touches is your PuntersEdge API key, read from the environment and never written to a tool result.
• change your PuntersEdge account. Key rotation, billing and webhook registration are all deliberately left out, so a chat turn cannot rotate your credentials, change what you are billed, or register a URL that then receives your data. Those endpoints exist; reach them from application code or the dashboard, on purpose.
• predict a result. Prices are what bookmakers published. Nothing here forecasts outcomes, rates a runner, or tells you what to back.

It will sometimes be wrong about the market, and the data says so. A stalled scraper does not error — it keeps serving its last value, which looks exactly like a live price. Every quote carries its own age_seconds and a stale flag for that reason. A large apparent edge on a market only one bookmaker is quoting is a data artefact, not an opportunity; check how many books stand behind a price before treating a gap as real.

What comes back

Racing responses carry the full market per runner:

{
  "race_name": "Wiesners Maiden Plate", "race_number": 1, "venue": "...",
  "category": "horse", "country": "AU", "distance_m": 1200,
  "track_condition": "...", "weather": "...", "places_paid": 3,
  "data_age_seconds": 38, "stale": false, "stale_bookmakers": [],
  "scratchings": [],
  "runners": [{
    "name": "...", "number": 1, "barrier": 4,
    "jockey": "...", "trainer": "...", "weight": 58.0, "form": "11521",
    "bookmakers": [
      { "key": "tabtouch", "win_price": 126.0, "place_price": 12.0,
        "tote_win": {"PROV": 60.6}, "last_update": "2026-08-17T02:55:3...",
        "source_url": "https://www.tabtouch.com.au/racing/..." }
    ]
  }]
}

Every response tells you its own age. data_age_seconds is race-level, each bookmaker entry carries its own last_update, and stale_bookmakers names the specific legs that have gone quiet rather than condemning the whole race. That matters when an assistant is about to state a price as fact — it can say how fresh the number is instead of implying it is live to the second.

Coverage, stated honestly

Racing: 15 Australian bookmakers (measured 19 Oct 2026) — BetRight, Sportsbet, Betr, TAB, Ladbrokes, Neds, Unibet, PointsBet, NextBet, TABtouch, Palmerbet, BetDeluxe, bet365, BetGold and BoostBet, which are 13 independent price sources (BetGold/BoostBet and Ladbrokes/Neds close at the same price on 85%+ of runners). The live count and per-book freshness are published at https://puntersedge.online/coverage-report?utm_source=github&utm_medium=readme.

Betfair and Pinnacle are not in this feed. The Betfair Exchange is ingested but its prices are withheld from customer responses pending a Betfair data licence, so it is not counted above and you will never see a Betfair price through this server. If you need exchange prices, this is not the feed for you.

Bookmaker keys in the API are not always the brand: NextBet is returned as playup, its name before rebranding. Filter on the key, not the display name.

Near the jump, coverage is broad and fairly even. In an earlier measurement, before bet365 joined the panel, over 13 days of captured snapshots — 3,339 Australian races, each within 30 minutes of jumping — every one of the fourteen sources quoted between 75% and 97% of them: TABtouch 97%, PointsBet 96%, then Sportsbet, Neds, Palmerbet, Ladbrokes, BetRight, Betr and TAB all clustered at 80–82%, Unibet 79% and NextBet 75%.

Books publish markets at different times, so a race several hours out often carries only two or three prices where more appear closer to the jump. Every response names the books that actually quoted, so you can read the real depth per race rather than trusting an average.

Sports: fewer again. On 19 Oct 2026 a quoted fixture averaged between 1 and 5 bookmakers depending on the sport, and on some competitions it is usually 1; the live per-sport figures are at https://puntersedge.online/api/coverage. That is a connector-coverage limit, not an outage. Racing is where this feed is strongest, and the tools say so rather than implying a uniform market.

Pass country=AU for racing (GB,IE for Britain and Ireland, ZA for South Africa — the code is GB, not UK). British, Irish and South African races carry smaller panels of their own (on 19 Oct 2026 a median of 4 and 5 books a race). In that earlier 13-day measurement, 87% of the other foreign and unlabelled races carried exactly one bookmaker, so there is usually nothing to compare — but "usually" is the honest word: 10% carried two books and a small tail (about 3%) carried eleven. Check bookmakers per race rather than assuming either way.

An Australian race gathers books as it approaches the jump, so its depth is a range, not a number. Measured 2026-08-17: a median of 10 bookmakers inside 30 minutes of the jump, 9 at 30–60 minutes, 5 at one to two hours, and 2 beyond that. racing_next_to_go returns imminent races, so it sits at the deep end of that range — which is why a headline average across the whole feed understates what you actually get back.

How much of the card is foreign swings with the clock rather than sitting at some headline ratio: near zero through the Australian afternoon, all of it overnight. Filter on country rather than assuming a mix.

Notes

• Not exposed: arbitrage and exchange endpoints. Some are withheld pending a Betfair licence and return 410 to customer keys, so shipping tools for them would only produce errors.
• This is data, not advice. Prices are what bookmakers published; nothing here predicts outcomes, and nothing here places a bet.
• Endpoints and parameters were enumerated from the live openapi.json, not from memory.

Links

• MCP server documentation
• API documentation
• Pricing — free tier 3,000 credits/month
• Live coverage report
• Python SDK
• Postman collection — import into Postman via Import → Link

18+ only. Gambling can be addictive — please gamble responsibly. Gambling Help: 1800 858 858, or Gambling Help Online.

MIT © PuntersEdge