Get the model's numbers into your code in minutes

Quickstart

Try it now — no signup

Run this. It returns the five most-likely tournament winners straight from the live model — no account, no API key. Add ?format=csv for the tabular variant.

curl "https://onthepitch.now/api/v1/sample/"

Want the full picture — all 48 teams, every fixture, players, squads? Sign in, generate an API key (10 free calls to start), and pass it as a Bearer token on any endpoint:

curl -H "Authorization: Bearer otp_live_…" \
  "https://onthepitch.now/api/v1/forecast/"

Examples

Three things you can do today

Set OTP_KEY to the API key you generate here (export OTP_KEY=otp_live_…) — one token, no cookie juggling, works the same from a shell, a notebook, or a scheduled job. (Signed in already? A Pro session cookie works too, but the key is the friendlier path.)

Full forecast as CSV

curl -H "Authorization: Bearer $OTP_KEY" \
  "https://onthepitch.now/api/v1/forecast/?format=csv" \
  -o forecast.csv

WC2022 historical slice

curl -H "Authorization: Bearer $OTP_KEY" \
  "https://onthepitch.now/api/v1/majors/?tournament=FIFA%20World%20Cup&year=2022" \
  -o wc2022.json

Bulk snapshot as a single zip

curl -H "Authorization: Bearer $OTP_KEY" \
  "https://onthepitch.now/api/v1/bulk/snapshot/" \
  -o onthepitch-snapshot-v1.zip

Build on the data

Pull the snapshot, fit your own model

The bulk zip ships everything the public site renders from — forecast, fixtures, predicted squads, head-to-heads, intl majors history. Three starter recipes for getting it into a notebook and running your own analysis against it.

1 · Load the snapshot into pandas

Pull the zip, extract it once, then read each file as a DataFrame. The forecast envelope nests one row per team — flatten with pd.json_normalize.

import io, json, os, zipfile, requests, pandas as pd

z = zipfile.ZipFile(io.BytesIO(requests.get(
    "https://onthepitch.now/api/v1/bulk/snapshot/",
    headers={"Authorization": f"Bearer {os.environ['OTP_KEY']}"},
).content))

forecast = pd.json_normalize(
    json.loads(z.read("data.json"))["teams"],
)
fixtures = pd.json_normalize(
    json.loads(z.read("fixtures.json"))["fixtures"],
)
squads = pd.json_normalize(
    json.loads(z.read("predicted_squads.json"))["squads"],
    record_path="players",
    meta=["team_id"],
)

2 · Fit a baseline win-probability model

Roll your own H/D/A model from the predicted squads + the team forecast. Composite ratings + a squad-strength sum is a reasonable starting feature set; compare against the published probabilities to see where your model agrees and disagrees.

from sklearn.linear_model import LogisticRegression

# Aggregate predicted-XI composite to a team strength score.
xi = squads.query("slot == 'xi'")
strength = xi.groupby("team_id")["composite"].sum().rename("strength")

# Join strength onto fixtures and fit a logistic on prior tournament
# results (intl_majors.json has W/D/L from previous WCs + qualifiers).
intl = pd.json_normalize(
    json.loads(z.read("intl_majors.json"))["rows"],
)
features = intl.merge(strength, left_on="team_id", right_index=True)
X = features[["strength", "fifa_rank"]]
y = (features["finish"] >= 9).astype(int)  # 1 = quarter-final or better

clf = LogisticRegression().fit(X, y)
print(clf.coef_)

3 · Calibrate your model against the published forecast

Once your model produces team-level tournament-win probabilities, compare against ours via Brier score and a reliability plot — the same metric we use in the methodology docs.

from sklearn.metrics import brier_score_loss

merged = forecast.merge(
    pd.Series(my_probs, name="my_p_win"),
    left_on="team_id", right_index=True,
)
print("OnThePitch Brier:", brier_score_loss(merged["won"], merged["p_win"]))
print("Yours        Brier:", brier_score_loss(merged["won"], merged["my_p_win"]))

The Pass licence is research-use — fit, backtest, publish your findings. Attribution back to the methodology doc is appreciated but not required.

Agents

Connect Claude, Cursor, or any MCP client

A hosted Model Context Protocol server at /api/mcp exposes all nine endpoints as MCP tools, so you can ask questions against the live model right inside your assistant — no install, no glue code. Authenticate with an API key — any signed-in account can generate one, and your first 10 tool calls are free.

Add it to Claude Code

claude mcp add --transport http onthepitch \
  https://onthepitch.now/api/mcp \
  --header "Authorization: Bearer otp_live_…"

Then prompt naturally — “which team is most likely to win the group?”, “compare the projected XIs for France and England”, “how did Brazil do at WC2022?” — and the assistant calls the matching tool (get_forecast, get_predicted_squads, get_majors) and answers from the live data.