jk-mcp-wsl
MCP server that gives Claude live access to Women's Super League (England) data — teams, matches, standings, rosters, and schedule-strength analytics — via the ESPN public API.
README
jk-mcp-wsl
MCP server that gives Claude live access to Women's Super League (England) data — teams, matches, standings, rosters, and schedule-strength analytics — via the ESPN public API.
Table of Contents
- Overview
- Features
- Requirements
- Installation
- Usage
- Configuration
- Claude Code
- Claude Desktop
- Docker
- Development
- Contributing
- License
Overview
AI assistants like Claude are knowledgeable, but they have a hard cutoff date — they cannot tell you today's Women's Super League standings, last night's scores, or which teams are currently on top. This project fixes that.
It is an MCP server — a plugin that gives Claude direct access to live Women's Super League data: scores, standings, rosters, and derived schedule-strength analytics. Once installed, you can ask Claude natural-language questions about the WSL and get accurate, up-to-date answers. No subscription, no API key, and no programming required to use it.
The Women's Super League is the top flight of English women's football. In 2024 it was spun out of The FA into its own independent operating company (WSL Football Ltd), running the WSL and WSL 2 as a fully professional 14-team pyramid. This server wraps the ESPN public JSON feed, which is the cleanest freely-accessible source for the league. Rich per-player and team-season stats — historically served via api-sdp.wslfootball.com — are deferred to v2 pending discovery of the correct stats/players category enum.
Features
The v1 surface is eleven read-only, idempotent tools split across two tiers.
ESPN-backed (8)
| Tool | Description |
|---|---|
get_teams |
List all 14 clubs with IDs and abbreviations |
get_team |
Details for a specific team |
get_roster |
Team's active roster — jersey, position, age, citizenship |
get_scoreboard |
Match scores for a single day, a date range, or the current matchweek |
get_team_schedule |
Every match for a team in the current season — past + upcoming |
get_match_details |
One match's full details — score, venue, attendance, goals, cards, subs |
get_standings |
Current standings — single 14-team table ordered by points |
get_news |
Recent Women's Super League news articles |
Derived analytics (3)
Pure functions over live standings + team schedules, exposing schedule-strength context the raw table does not.
| Tool | Description |
|---|---|
get_strength_of_schedule |
Team's average opponent points-per-game across matches already played |
get_results_by_opponent_tier |
Team's W-L-T split across current top / middle / bottom standings tiers |
get_adjusted_points_per_game |
Team's raw PPG alongside an opponent-quality-adjusted PPG |
Roadmap
Deferred to v2+:
- Player leaderboards, team season stats, and per-player heatmaps via the league's own SDP tier (
api-sdp.wslfootball.com/stats/players) — thecategoryenum required to hit the endpoint has not been discovered from public traffic yet - FA Cup and Continental Cup fixtures
- Playoff / relegation bracket rendering
- Related competitions (UEFA Women's Champions League match-day fixtures for WSL clubs)
Requirements
Installation
git clone https://github.com/jedi-knights/jk-mcp-wsl.git
cd jk-mcp-wsl
uv sync
Usage
Run the server in stdio mode (the default — used by Claude Code and Claude Desktop):
uv run python -m wsl.server
Run in HTTP mode (for networked or deployed access):
MCP_TRANSPORT=streamable-http uv run python -m wsl.server
Example prompts
Standings, scores, rosters:
- Who is leading the WSL right now?
- Show me every WSL result from this past weekend.
- Who is on Arsenal's roster?
- When does Aston Villa play next?
Schedule strength:
- Which WSL side has played the toughest schedule so far?
- Show me Manchester United's record against the current top 3 clubs.
- Compare Chelsea and Manchester City on adjusted points-per-game.
Configuration
All configuration is via environment variables. None are required for local use.
| Variable | Default | Description |
|---|---|---|
MCP_TRANSPORT |
stdio |
Transport mode: stdio or streamable-http |
HOST |
0.0.0.0 |
Bind address (HTTP transport only) |
PORT |
8000 |
TCP port (HTTP transport only) |
MCP_PATH |
/mcp/wsl |
URL path (HTTP transport only) |
API_HOST |
https://site.api.espn.com |
ESPN API base URL |
LOG_LEVEL |
INFO |
DEBUG, INFO, WARNING, or ERROR |
MCP_TRACING_ENABLED |
unset | Bootstrap the OpenTelemetry SDK |
MCP_AUTH_ENABLED |
unset | Require RS256 bearer tokens on streamable-http |
MCP_AUTH_ISSUER_URL |
unset | Auth-server origin (required when auth is on) |
MCP_AUTH_RESOURCE_URL |
unset | This server's public URL for the aud claim |
Claude Code
Install from your local clone globally so the server is available in every project:
claude mcp add --scope user wsl -- uv run --directory /path/to/jk-mcp-wsl python -m wsl.server
Replace /path/to/jk-mcp-wsl with the absolute path to your clone. Verify with claude mcp list.
Drop --scope user to register only for the current project, or commit a .mcp.json to the repo root for collaborators:
{
"mcpServers": {
"wsl": {
"command": "uv",
"args": ["run", "--directory", "/path/to/jk-mcp-wsl", "python", "-m", "wsl.server"]
}
}
}
Claude Desktop
Add the following to your Claude Desktop configuration file.
Location:
- macOS:
~/Library/Application Support/Claude/claude_desktop_config.json - Windows:
%APPDATA%\Claude\claude_desktop_config.json
{
"mcpServers": {
"wsl": {
"command": "uv",
"args": [
"run",
"--directory", "/path/to/jk-mcp-wsl",
"python", "-m", "wsl.server"
]
}
}
}
If uv is not on Claude Desktop's PATH, use the absolute path (which uv will show it). Fully quit and relaunch Claude Desktop after saving — a window close is not enough.
Docker
Build the image:
docker build -t jk-mcp-wsl:latest .
Run in stdio mode (for MCP clients that spawn a subprocess):
docker run -i --rm jk-mcp-wsl:latest
Run in HTTP mode:
docker run --rm -p 8000:8000 \
-e MCP_TRANSPORT=streamable-http \
jk-mcp-wsl:latest
Development
Install
uv sync
Invoke tasks
All common workflows are invoke tasks. Run uv run inv --list to see everything.
| Task | Alias | Description |
|---|---|---|
uv run inv lint |
inv l |
Run ruff linter and format check |
uv run inv lint --fix |
inv l --fix |
Auto-fix lint violations and reformat |
uv run inv test |
inv t |
Run the full test suite |
uv run inv coverage |
inv v |
Run tests with coverage report (threshold: 90%) |
uv run inv check-complexity |
inv cc |
Check cyclomatic complexity (max 7) |
uv run inv build |
inv b |
Build wheel and sdist into dist/ |
uv run inv build-image |
inv bi |
Build the Docker image |
uv run inv clean |
inv c |
Remove build and coverage artifacts |
Project structure
src/wsl/
├── server.py # entry point, transport selection, logging setup
├── adapters/
│ ├── inbound/
│ │ ├── mcp_adapter.py # FastMCP server, health endpoints, tool registration
│ │ ├── formatters.py # domain → LLM-readable text
│ │ ├── authorization.py # inbound authz port implementations
│ │ └── tools/
│ │ ├── espn.py # 8 ESPN-backed tools
│ │ └── analytics.py # 3 schedule-strength analytics tools
│ └── outbound/
│ ├── espn_adapter.py # ESPN HTTP client
│ ├── parsers.py # ESPN JSON → domain models
│ ├── retry_adapter.py # transient-failure retry decorator
│ └── caching_adapter.py # in-process TTL cache
├── application/
│ ├── service.py # WSLService — use cases, orchestration
│ ├── _helpers.py # input validation
│ └── _analytics_helpers.py # pure math for schedule-strength tools
├── domain/
│ ├── models.py # Team, Match, Standing, etc.
│ └── exceptions.py # WSLNotFoundError, UpstreamAPIError
├── ports/
│ ├── inbound.py # Authorizer protocol
│ └── outbound.py # WSLAPIPort protocol
├── observability/ # OpenTelemetry bootstrap (opt-in)
└── security/ # JWKS token verifier
The dependency direction flows inward: adapters → ports → domain. Nothing in domain/ imports from adapters or a framework.
Contributing
- Fork the repository and clone your fork
- Create a feature branch:
git checkout -b feature/your-feature - Make your changes following the existing patterns (hexagonal architecture, TDD, conventional commits)
- Verify the full check suite passes:
uv run inv lint && uv run inv check-complexity && uv run inv coverage - Open a pull request against
main
All CI checks (lint, complexity, tests, coverage ≥ 90%) must pass before merge.
License
MIT — see LICENSE.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。