worldcup-mcp
# worldcup-mcp
A local **MCP server (stdio)** exposing live **2026 FIFA World Cup** group standings and a
Round-of-32 bracket, so a Cowork HTML artifact can call it on page load and render current data.
- TypeScript (strict), Node ≥ 20, `@modelcontextprotocol/sdk` over stdio.
- Provider abstraction (`StandingsProvider`) + 60s cache + `stale` last-good fallback.
- Pure-logic layer (third-place ranking, best-third allocation, confidence) is fully unit-tested.
> **Status:** the server is complete and its logic/handshake tests pass offline. The **live data
> gate is unmet until you add a free API token** — see [`DATA_SOURCE.md`](./DATA_SOURCE.md).
> Without a token the tools return an all-undecided bracket marked `stale` (no faked standings).
---
## 1. Install
```bash
cd worldcup-mcp
npm install
cp .env.example .env # then edit .env (see below)
```
## 2. Environment
| Var | Purpose |
| --- | --- |
| `PROVIDER` | `football-data` (default, recommended) or `thesportsdb` |
| `FOOTBALL_DATA_TOKEN` | free token from <https://www.football-data.org/client/register> |
| `THESPORTSDB_KEY` | `3` (free, **capped at ~5 rows**) or a Premium key |
Get the free football-data token (instant email), put it in `.env`:
```
PROVIDER=football-data
FOOTBALL_DATA_TOKEN=xxxxxxxxxxxxxxxxxxxxxxxx
```
## 3. Scripts
```bash
npm run dev # run the server from source (tsx) over stdio
npm run build # tsc -> dist/
npm start # node dist/server.js
npm test # vitest — pure-logic unit tests (no network)
npm run verify # LIVE smoke test: one real fetch, prints 12 tables + 16 R32 slots, asserts
npm run handshake # builds, launches server over stdio, lists+calls both tools, validates shapes
```
## 4. Tools & output contracts
### `get_group_standings`
Returns an object keyed by group letter `A`–`L`; each value is an array ordered rank 1→4 of
`{ name, code, played, won, drawn, lost, goalsFor, goalsAgainst, goalDifference, points, rank,
confidence }`. `confidence` ∈ `confirmed | likely | open` (see classifier below). The object also
carries `stale` and `fetchedAt` alongside the letter keys.
### `get_round_of_32`
Returns `{ matches: [...16], stale, fetchedAt }`. Each match:
`{ match, posA, posB, slotA, slotB }`, where a slot is either
`{ posCode, decided:true, team:{name,code,confidence} }` or
`{ posCode, decided:false, eligibleGroups:[...] }`.
**Bracket positions are fixed** (from the official draw); only the teams change. Position map is in
[`src/logic/bracket.ts`](./src/logic/bracket.ts).
**Resolution rules**
- *Rank (1x/2x) slot:* `decided:true` with the current holder of that exact rank once the group has
begun, carrying its `confidence`; `decided:false` with `eligibleGroups:[group]` before any game.
- *Best-third slot:* stays `decided:false` (with the 5-group pool) until the group stage is
**complete**; then the top-8 thirds are ranked by **points → goal difference → goals scored** and
allocated to slots honoring each slot's `eligibleGroups`.
### `get_knockout_results`
Returns `{ rounds, stale, fetchedAt }` where `rounds` is `{ r16:[…8], qf:[…4], sf:[…2],
third:[…1], final:[…1] }`. Each match: `{ match, posA, posB, slotA, slotB, status, score,
winnerCode }`:
- `slotA`/`slotB` mirror the `get_round_of_32` shape: `{ posCode, decided:true,
team:{name,code,confidence} }` once the feeding match is finished, else `{ posCode,
decided:false, feeder }` (a label like `"R32 M12"` / `"R16 #89"`).
- `status` ∈ `scheduled | live | finished`; `score` is `{ a, b, duration?, penalties? }` or `null`;
`winnerCode` is set **only** when `status === "finished"`.
- **Bracket wiring is ours** (the official 2026 map: R32 ties M1..M16 feed forward as consecutive
pairs → R16 → QF → SF → final); **results are overlaid from the live source** by matching teams.
Match numbers use the official 89–104. See [`src/logic/knockout.ts`](./src/logic/knockout.ts).
- Real winners propagate: a slot fills as soon as its feeding match is finished, otherwise it stays
a placeholder. If the provider can't supply knockout data, every node is a placeholder + `stale`.
### `confidence` classifier
`confirmed` = mathematically locked at the exact rank (every other team is clearly above or clearly
below given remaining fixtures; reachable-point ties count as not-yet-separated, so it never
over-claims). `likely` = currently in a qualifying slot (rank 1–2) but not locked. `open` = otherwise.
Once every team in a group has played all 3 games, all four are `confirmed`.
## 5. Cowork / Claude Desktop integration
Register this server under the slug **`worldcup`** so the tools resolve to the names the artifact
expects:
- `mcp__worldcup__get_group_standings`
- `mcp__worldcup__get_round_of_32`
- `mcp__worldcup__get_knockout_results`
These are already the defaults in [`artifact/wc2026_bracket.html`](./artifact/wc2026_bracket.html)
(`TOOL_STANDINGS` / `TOOL_RO32` / `TOOL_KO`). If you register under a different key, update those
constants. The artifact renders the live R32 + a projected bracket, and overlays **real** knockout
results from `get_knockout_results` once matches are played (green "result" boxes with scores;
muted boxes remain projections).
MCP config block (Claude Desktop `claude_desktop_config.json` → `mcpServers`, or the Cowork MCP
config). Use the **built** entry for stability:
```json
{
"mcpServers": {
"worldcup": {
"command": "node",
"args": ["C:\\Users\\U\\OneDrive\\worldcup-mcp\\dist\\server.js"],
"env": {
"PROVIDER": "football-data",
"FOOTBALL_DATA_TOKEN": "your-free-token"
}
}
}
}
```
(Run `npm run build` first so `dist/server.js` exists. For a no-build dev setup, use
`"command": "npx", "args": ["tsx", "C:\\Users\\U\\OneDrive\\worldcup-mcp\\src\\server.ts"]`.)
**Freshness for the artifact's "Live / stale" banner:** football-data.org refreshes tables within
**~1–5 minutes** of a match finishing (it is not a per-second live feed). Expect "Live" right after
a refresh and "stale" only if the provider call fails (then last-good is served).
## 6. Manual FIFA cross-check
To eyeball that live numbers are real and current, compare one group against the official source:
```bash
# Group A table straight from the provider (needs your token):
curl -s -H "X-Auth-Token: $FOOTBALL_DATA_TOKEN" \
"https://api.football-data.org/v4/competitions/WC/standings" \
| npx --yes node-jq '.standings[] | select(.group=="GROUP_A") | .table[] | {pos:.position, team:.team.name, pld:.playedGames, pts:.points, gd:.goalDifference}'
```
Then open the official FIFA standings page for the 2026 World Cup
(<https://www.fifa.com/fifaplus/en/tournaments/mens/worldcup/canadamexicousa2026>) and confirm the
played/points/GD for Group A match. `npm run verify` performs the structural half of this check
automatically (groups, team counts, codes vs. the official draw).
## 7. Known limitations
- **Best-third allocation method.** FIFA publishes a fixed lookup table keyed by *which* 8 of 12
groups produce a qualifying third (C(12,8)=495 rows). Rather than embed that table verbatim, this
server solves the same constraint it encodes — a deterministic eligibility-honoring bipartite
matching (see [`src/logic/thirdPlace.ts`](./src/logic/thirdPlace.ts)). Every group always lands in
an **eligible** slot; for some of the 495 combinations the exact group→slot pairing may differ
from FIFA's printed table while staying eligibility-correct. Swap in the literal table here if you
need byte-for-byte parity with FIFA's published bracket.
- **Free data sources are credential-gated or capped** — see `DATA_SOURCE.md`. Nothing is hardcoded;
absent a credential the bracket is all-undecided + `stale`.
## 8. Layout
```
src/
draw.ts official group draw + name→code mapping (validation only)
types.ts shared types
cache.ts 60s TTL + last-good stale fallback
service.ts provider selection, enrich(), the two tool entry points
server.ts MCP stdio server (registers the two tools)
verify.ts live smoke test
env.ts local .env loader
providers/
types.ts StandingsProvider interface (the swap seam)
footballData.ts primary provider
theSportsDb.ts secondary provider (capped on free tier)
logic/
confidence.ts confidence classifier
thirdPlace.ts third-place ranking + best-third allocation
bracket.ts fixed R32 position map + resolution
test/ vitest unit tests (confidence, thirdPlace, bracket)
scripts/handshake.ts MCP stdio handshake test
```
TDQS
Scored across 2 tools
The two tools have clearly distinct purposes: one for group standings and one for the round of 32 bracket. There is no overlap, so an agent can easily differentiate them.
Both tools follow the consistent 'get_<noun>' naming pattern (get_group_standings, get_round_of_32), providing a predictable and uniform interface.
With only 2 tools, the server feels under-scoped for a comprehensive World Cup data service. While the tools are relevant, additional tools for match details or team info would be expected.
The server covers only group standings and the round of 32 bracket, missing critical features like match results, team information, player stats, and later knockout stages. Significant gaps exist for a full World Cup data server.