Most AI assistants, asked who wins Arsenal vs Chelsea, will improvise. They'll reason from vibes, training-data memories, and the occasional hallucinated injury list. The answer sounds confident and is worth exactly nothing.

We run Footalytics, a football prediction service that publishes probabilities before kickoff and grades every one of them after the final whistle. As of this week your AI assistant can pull those numbers directly. This post is the 5-minute setup, three ways.

The numbers you're wiring in

Quick honesty check on the model, from our public ledger (n=189 matches, Oct 17 to Nov 3, 2026):

  • 59.8% of match outcome picks hit (baseline: always-pick-home is 48.7%, random is 33.3%)
  • mean Brier score 0.577
  • every settled prediction is graded in public — hits and misses — with a full CSV

Now three ways to give your agent access.

Route A: MCP connector (no key files)

Footalytics runs a hosted MCP server at https://footalytics.farvor.com/mcp.

In Claude (web/desktop): Settings → Connectors → Add connector, paste the URL, and sign in with OAuth — one click, no tokens to manage. In any other MCP-capable client (Cursor, Codex, custom agents), add the same URL as a remote Streamable HTTP server.

The server exposes five tools — predictions, fixtures, teams, results, and the accuracy ledger — with the same allowance as the API key.

Route B: the agent skill (SKILL.md)

For agents that speak the Agent Skills standard — Claude Code, Codex, Cursor and friends — grab the skill:

git clone https://gitlab.com/footalytics/footalytics-skill.git

Point your client at SKILL.md (Claude Code: /plugin marketplace add lewis-v/footalytics-skill, or drop the folder into your skills directory). The frontmatter teaches the agent when to reach for football predictions; the body teaches it how to call the API, which fields to show, and how to present the graded record.

--- name: footalytics description: Football match predictions and football data from Footalytics — win/draw/loss probabilities, confidence tier, over/under 2.5 and BTTS, plus upcoming fixtures, team ratings and graded past results... ---

Route C: plain REST

Prefer no magic? Create a free key at footalytics.farvor.com/api-keys — new accounts get 5 free API calls, no card — and call it:

curl -H "Authorization: Bearer $FOOTALYTICS_API_KEY" "https://footalytics.farvor.com/api/v1/predictions?home=Arsenal&away=Chelsea"

Every response carries x-credits-remaining, so the agent knows when to warn you. The skill also handles the fun errors: E_AMBIGUOUS_MATCH (three teams called "River Plate"), E_PREDICTION_UNAVAILABLE (predictions publish up to 48h before kickoff), and rate limits with a Retry-After.

What comes back

A trimmed prediction response looks like this:

{ "prediction": { "outcome": { "home": 0.267, "draw": 0.268, "away": 0.465, "pick": "away" }, "confidence": "medium", "top_scores": [{ "score": "1-1", "p": 0.128 }], "over_2_5": 0.515, "btts": 0.553 }, "links": { "share": "https://footalytics.farvor.com/p/5vx6s98xpaqu" } }

And for settled matches, the grading rides along: our_prediction.hit_1x2, hit_over_2_5, hit_btts, and a Brier score per match.

Ask it something

With any of the three routes connected, try:

"Who wins Arsenal vs Chelsea this weekend? How confident is the model?"

The attribution link in every response goes to the public record, so you can check the model's homework yourself. That's the whole pitch: not smarter-sounding answers — auditable ones.

Links: accuracy ledger · API docs · skill repo · official MCP Registry entry (search com.farvor/footalytics)