Model Context Protocol

Footdigest MCP

Connect Footdigest to Claude, or any MCP client, and ask football questions in plain language. Instead of guessing from memory, the model calls our engine and answers with calibrated, sourced numbers, every one timestamped and traceable back to us.

The server speaks JSON-RPC 2.0 over Streamable HTTP. It is read-only: it returns numbers and never changes anything. Fourteen tools are live today, backed by the same backtested Dixon-Coles + Elo engine that powers the site.

Quickstart

Using Claude Desktop, you are connected in three steps: open Settings → Developer → Edit Config, add the Footdigest server, and restart.

{
  "mcpServers": {
    "footdigest": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://footdigest.com/mcp"]
    }
  }
}

Then ask “What are France's chances against Spain?” and watch it call the tool.

Connect a client

Footdigest works with any MCP-compatible client. Pick yours.

{
  "mcpServers": {
    "footdigest": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://footdigest.com/mcp"]
    }
  }
}

$ claude mcp add --transport http footdigest https://footdigest.com/mcp

Name  Footdigest
URL   https://footdigest.com/mcp
Auth  None

{
  "mcpServers": {
    "footdigest": { "httpUrl": "https://footdigest.com/mcp" }
  }
}

{
  "mcpServers": {
    "footdigest": { "url": "https://footdigest.com/mcp" }
  }
}

$ curl -s https://footdigest.com/mcp \
    -H 'content-type: application/json' \
    -H 'accept: application/json' \
    -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

MCP support is spreading fast. Footdigest also works from VS Code (Copilot agent mode), Windsurf, Zed, Cline, and any other client that speaks Streamable HTTP — just point it at the endpoint. Steps for ChatGPT and Gemini vary by plan and version.

Coverage

The tools answer about the competitions Footdigest models, and that set keeps growing. For the current live list, open the competition menu on footdigest.com: whatever is there is what the tools can answer about. get_match_probabilities also works for any two teams in our ratings history, even outside a live tournament.

Ask about a team or competition Footdigest doesn't cover and the tool returns an explicit “unknown” rather than inventing a number, the guarantee behind every answer.

Example prompts

Once connected, try asking your assistant. Click a prompt to copy it.

Tools

Call a tool with a JSON-RPC tools/call request. The shape is identical for every tool; only name and arguments change. Your client builds this for you, the raw form is here for direct HTTP use.

{
  "jsonrpc": "2.0", "id": 1, "method": "tools/call",
  "params": {
    "name": "get_match_probabilities",
    "arguments": { "home_team": "France", "away_team": "Spain" }
  }
}

All fourteen tools

Here is the complete set. Every one returns the same provenance envelope, rejects inputs the data doesn't cover, and takes a competition slug (e.g. world-cup-2026) or team names where noted. Three of them get a full walkthrough below.

The authoritative list is always tools/list on the endpoint itself; each tool carries its own JSON Schema, so your client discovers the exact arguments automatically.

get_match_probabilities

Calibrated home / draw / away probabilities and expected goals for a matchup. Ask by fixture match_id, or by team names for any pairing.

Provide match_id, or both home_team and away_team. Teams the ratings dataset doesn't know are rejected rather than scored at a coin-flip.

{
  "data": {
    "home_team": "France", "away_team": "Spain",
    "probabilities": { "home_win": 0.3108, "draw": 0.2874, "away_win": 0.4018 },
    "expected_goals": { "home": 1.2027, "away": 1.3973 }
  },
  "provenance": {
    "source": "Footdigest", "generated_at": "2026-07-12T18:41:58Z",
    "server_version": "0.1.0", "url": "https://footdigest.com"
  }
}

simulate

Run the seeded Monte Carlo over a competition's group stage and get each team's advancement probabilities. Pass a seed and the run is fully reproducible.

{
  "competition": "World Cup 2026", "trials": 10000, "seed": 42,
  "note": "Reproducible: the same competition, trials, and seed return the same probabilities.",
  "teams": [
    { "team": "Spain", "advance": 0.94, "best_third": 0.03, "win_group": 0.71 }
    /* one row per team, sorted by advance */
  ]
}

get_model_card

The receipts. The live backtest behind every number: accuracy versus a naive baseline, the calibration table, the methodology, and an explicit list of what we don't measure. Takes no parameters.

{
  "model": "Dixon-Coles scoreline model over World-Football-Elo ratings",
  "scoring": {
    "metric": "RPS (ranked probability score; lower is better)",
    "engine_rps": 0.1723, "baseline_rps": 0.2273,
    "improvement_vs_baseline": 0.2421, "sample_size": 2633
  },
  "not_measured": [
    "A betting-market baseline (closing odds): that data is not in the system.",
    "Expected goals (xG): not currently in the dataset."
  ]
}

Read the full method, with interactive figures, in Distilled.

Responses & provenance

Every tool result is wrapped in the same envelope: the data payload plus a provenance block. An agent that uses our numbers can always cite where they came from, and can never quietly launder them into an anonymous claim.

Results also carry an MCP-standard content array (the same payload as text), so clients that don't read structuredContent still get the answer.

Errors

Two kinds of failure, both following JSON-RPC conventions.

Protocol errors (unknown method or tool, invalid arguments) return a JSON-RPC error with code -32601 or -32602:

{ "jsonrpc": "2.0", "id": 1,
  "error": { "code": -32602, "message": "missing required argument: competition" } }

Tool errors (e.g. a team we don't cover) come back as a normal result flagged isError, so the assistant can read and relay them:

{ "content": [{ "type": "text", "text": "Tool error: unknown_team \"Atlantis\"" }],
  "isError": true }

Verifying an answer

Because an assistant relays our numbers in its own words, here is how to be sure a football answer actually came from Footdigest and wasn't invented:

• Check the provenance. A real result carries the provenance block, with source: "Footdigest", a timestamp, and a link.
• Ask it to cite. “Did you use the Footdigest tool, and what did it return?” A genuine call shows the numbers and the link; a guess can't.
• Match the site. The url in the response points to the Footdigest page for that answer — open it and the numbers line up.
• Audit the model. Ask for get_model_card to see how calibrated the number you are trusting actually is.

Access & limits

The endpoint is open and requires no key. It is read-only. Heavy calls (large simulate runs) are cached, and trials is capped at 50,000 per call. A per-client rate limit is in place, so behave reasonably and you'll never meet it.

Privacy

We don't store the questions you ask through the tools, and we don't build a profile from them. Like any web service, the server keeps short-lived operational logs (request timing, errors) for reliability and abuse prevention, not your content.

Responsible use

These are model probabilities, not tips. A calibrated forecast is designed to be wrong a fair share of the time, that is what calibration means. Nothing here is betting advice, and the numbers should be read as estimates with the uncertainty the model card makes explicit.

Troubleshooting

The server doesn't show up

Fully quit and reopen the client after editing its config, a reload isn't always enough. Confirm the JSON is valid (no trailing commas).

npx / mcp-remote errors

Claude Desktop's bridge needs Node.js installed and on your PATH. Claude Code and Cursor connect to the URL directly and don't need it.

“Unknown team” or “unknown competition”

We reject inputs the dataset doesn't cover rather than inventing a number. Check spelling, and use a competition slug like world-cup-2026 for simulate (see Coverage).