SportIQ-MCP

# sportiq-mcp

<!-- mcp-name: io.github.Ninjabeam20/sportiq-mcp -->

[![CI](https://github.com/Ninjabeam20/SportIQ-MCP/actions/workflows/test.yml/badge.svg)](https://github.com/Ninjabeam20/SportIQ-MCP/actions/workflows/test.yml)
[![PyPI](https://img.shields.io/pypi/v/sportiq-mcp.svg)](https://pypi.org/project/sportiq-mcp/)
[![Python](https://img.shields.io/pypi/pyversions/sportiq-mcp.svg)](https://pypi.org/project/sportiq-mcp/)
[![License: MIT](https://img.shields.io/badge/License-MIT-green.svg)](LICENSE)
[![MCP Registry](https://img.shields.io/badge/MCP%20Registry-listed-blue)](https://registry.modelcontextprotocol.io)

MCP server that turns any AI assistant into a sports analyst across **FIFA World Cup 2026 football, Formula 1, and IPL cricket** — 44 AI-callable tools.

![SportIQ demo — Claude calling football_simulate_bracket for World Cup 2026 title probabilities](docs/assets/demo.gif)

*SportIQ running live in Claude — Monte Carlo World Cup bracket, F1 pit strategy, and Dream11 optimisation, each a visible MCP tool call. ([1-min demo](docs/assets/SportIQ.mp4))*

<p align="center">
  <a href="https://github.com/sponsors/Ninjabeam20"><img src="https://img.shields.io/badge/%E2%9D%A4_Sponsor_SportIQ-EA4AAA?style=for-the-badge&logo=githubsponsors&logoColor=white" alt="Sponsor SportIQ"></a>
  &nbsp;
  <a href="https://sport-iq-sports-analysis.vercel.app"><img src="https://img.shields.io/badge/%F0%9F%8C%90_Website-Visit-2563EB?style=for-the-badge" alt="Website"></a>
</p>

> **Every tool is free to use** — the three flagships and everything in the INTEL columns below have no SportIQ paywall or account requirement. Live/provider-backed data still depends on the keys and quota available to the host or local operator. If SportIQ is useful to you, [sponsor the project](https://github.com/sponsors/Ninjabeam20) to support ongoing development.

## What it does

Raw-data tools are table stakes; the intelligence layer is the product. Three flagships:

- **`football_simulate_bracket`** — Monte Carlo with Poisson xG over the 48-team WC 2026 format → per-team round + title probabilities.
- **`f1_predict_pit_strategy`** — tyre-degradation model on OpenF1 telemetry → optimal stop laps + compound sequence.
- **`cricket_build_dream11_team`** — PuLP constraint solver → a valid fantasy XI under credit/role/team caps.

### Tools (44 total)

| Sport | RAW data | INTEL |
|-------|----------|-------|
| **Football** (WC 2026) | groups, fixtures, standings, squad, match stats, top scorers, odds | xg_model, match_predictor, simulate_group, **simulate_bracket**, knockout_path, form_trends, find_value_bets, build_accumulator |
| **F1** | sessions, drivers, lap_times, standings, race_results, weather | tyre_degradation, undercut_window, head_to_head_pace, weather_strategy_impact, qualifying_analysis, race_pace_compare, **predict_pit_strategy** |
| **Cricket** (IPL) | live_matches, scorecard, points_table, schedule, squad, live_odds | **build_dream11_team**, captain_recommendation, differential_picks, player_form_index, pitch_report, head_to_head, player_matchup, find_value_bets |
| **Cross-sport** | — | build_accumulator |

Plus `sportiq_health` (included in the 44; cache backend + per-adapter status and remaining API quota).

**Data sources** (per chain, with keyless fallbacks): football → [API-Football](https://www.api-football.com) → [football-data.org](https://football-data.org) → bundled `wc2026.json`. F1 → [OpenF1](https://openf1.org) → [Jolpica](https://jolpi.ca) → `fastf1`. Cricket → CricAPI + static seeds (NDTV/Cricbuzz scrapers opt-in).

## Where it works

Anywhere that speaks MCP — Claude (Desktop + web), ChatGPT, Cursor, and any MCP client. Two ways to run it:

- **Hosted (no install):** add a custom connector — works in claude.ai web & ChatGPT.
- **Local (`uvx`/Desktop config/IDEs):** install from PyPI.

## How it works

### Hosted — no install

A public instance runs on the home server behind Cloudflare. Add this as a custom connector with **No authentication**:

```
https://sportiq.utkarshgupta.org/mcp
```

Hosting history (GCP Cloud Run → home server): [`docs/wiki/findings/product-hosting-arc.md`](docs/wiki/findings/product-hosting-arc.md). Old `*.run.app` connector URLs are gone.

- **claude.ai (web):** Settings → Connectors → Add custom connector → paste URL → Save.
- **ChatGPT:** Settings → Apps & Connectors → enable **Developer mode** → Create app (MCP) → paste URL → No authentication → Connect.

All 44 tools register on the plain URL. Whether a live/provider-backed call can return current data depends on the credentials, quota, and fallbacks available to the hosted operator; the repository does not claim the public instance's current key inventory.

| Mode | What is available |
|------|-------------------|
| **Hosted** | All tools register; live/provider-backed results depend on the host's current keys, quota, and fallbacks. |
| **Local, keyless** | All tools register; bundled seeds and keyless sources work where supported, while credential-only live sources are skipped. |
| **Local, BYO keys** | The same tools can use the configured providers for fresher/live data, subject to provider quota. |

The hosted HTTP boundary rejects request bodies over 1 MiB, limits traffic to 60 requests per client and 300 total requests per minute, and permits at most two concurrent expensive model/solver calls. These counters are per process, so the home-server Compose stack runs **one** replica (always-on idle; no scale-to-zero).

### Local install

```bash
uvx sportiq-mcp                       # from PyPI
# or from source:
git clone https://github.com/Ninjabeam20/SportIQ-MCP && cd sportiq-mcp
uv sync --extra dev --extra analytics && uv run python -m sportiq.server
```

**Claude Desktop config:**

```json
{
  "mcpServers": {
    "sportiq": {
      "command": "uvx",
      "args": ["sportiq-mcp"],
      "env": {
        "CRICAPI_KEY": "your_cricapi_key",
        "APIFOOTBALL_KEY": "your_apifootball_key",
        "THEODDS_KEY": "your_theodds_key"
      }
    }
  }
}
```

The server boots and registers every tool without keys. Seed/keyless fallbacks and the intelligence layer work where their required inputs are available; provider keys add fresher/live sources and quota rather than unlocking a separate paid tool tier.

| Var | Unlocks | Free tier |
|-----|---------|-----------|
| `APIFOOTBALL_KEY` | Live football fixtures / standings / squads / scorers | 100 req/day |
| `THEODDS_KEY` | Market odds (football + cricket probability tools) | 500 req/month |
| `FOOTBALLDATA_KEY` | football-data.org fallback (token optional) | 10 req/min |
| `CRICAPI_KEY` | Live cricket scores / scorecards / schedules / squads | 100 req/day |
| `RAPIDAPI_KEY` | Paid Cricbuzz fallback (player career stats) | plan-dependent |
| `SPORTIQ_ENABLE_NDTV` / `SPORTIQ_ENABLE_CRICBUZZ` | Opt-in cricket scrapers (off by default — ToS) | — |
| `REDIS_URL` | Shared cache backend (defaults to local diskcache) | — |
| `SPORTIQ_TRANSPORT` | `stdio` (default, local) or `http` (remote / home server) | — |

> macOS arm64: the Dream11 solver needs CBC — `brew install cbc` (the binary bundled with PuLP is x86-only).

### Self-host

Set `SPORTIQ_TRANSPORT=http` and the server serves the MCP endpoint at `/mcp` (binds `0.0.0.0:$PORT`). A ready-to-build `Dockerfile` and home-server `docker-compose.yml` are included. `cloud.md` is the old Cloud Run runbook (historical). With your own keys set, the live-score and odds tools come online too.

## Support SportIQ

Every tool is free and open source — the raw-data tools, `sportiq_health`, and the full intelligence layer (the three flagships + everything in the INTEL columns). SportIQ has no paid feature gate; provider-backed data can still require operator credentials and quota.

If SportIQ saves you time, **[sponsor the project at github.com/sponsors/Ninjabeam20](https://github.com/sponsors/Ninjabeam20)** to help fund hosting and ongoing development. It's a voluntary donation — you get the same fully-unlocked server either way.

## Is it safe?

- **Open source, MIT licensed**, published on [PyPI](https://pypi.org/project/sportiq-mcp/) with signed build attestations — read the code before you connect it.
- **Read-only.** Tools only fetch and analyse public sports data — no write, delete, payment, email, or file-system tools.
- **Limited operational telemetry.** HTTP mode logs client software name/version, User-Agent, tool name, outcome, latency, selected source, and staleness. The public host (Dell) can persist `tool_call` / `mcp_request` lines to a local JSONL volume. Local stdio emits local logs but sends no telemetry to a SportIQ-hosted service.
- **Hosted abuse controls.** HTTP POST bodies are capped at 1 MiB; requests are limited to 60/client/minute and 300/process/minute; the five expensive simulation/strategy/solver tools share a concurrency limit of two.
- **Credential-aware.** A hosted operator may configure provider credentials; the repository does not claim the public instance's current key inventory. Keys are redacted from application logs and envelopes.
- Historical automated AI code-review results are documented in [`SECURITY.md`](SECURITY.md#independent-review); they are not a current third-party certification.

Every response carries a `meta.is_stale` flag + data age, so the AI tells you how fresh each answer is. Live scores refresh ~30s, F1 telemetry ~10s, standings ~10min, fixtures ~6h.

## Develop

```bash
uv sync --extra dev --extra analytics   # always both extras: dev = pytest/ruff, analytics = the dashboard's GCP libs
uv run pytest
uv run ruff check .
npx @modelcontextprotocol/inspector uv run python -m sportiq.server
```

**Analytics dashboard** (read-only local usage view — Dell JSONL / archived GCP / PyPI / GitHub). Same setup as above, then just run it:

```bash
uv run python scripts/dashboard.py     # writes dashboard.html and opens it; GITHUB_TOKEN optional (Sponsors panel)
```

> Note: the dashboard's HTML templat