SofaScore MCP

SofaScore MCP

An MCP server that answers football questions with interactive widgets, centered on the 2026 FIFA World Cup using SofaScore API data.

Category
Visit Server

README

SofaScore ChatGPT App βš½πŸ†

A ChatGPT App (OpenAI Apps SDK / MCP-UI) that answers football questions with rich, interactive widgets β€” centered on the 2026 FIFA World Cup (48 teams, hosted by the USA, Canada & Mexico) while still handling general football.

It's an MCP server (Streamable HTTP) whose tools each render a React widget inside ChatGPT via the openai/outputTemplate + text/html+skybridge pattern. Data comes from SofaScore's public JSON API.

World Cup hub

<p> <img src="docs/match-detail.png" width="32%" alt="Match detail" /> <img src="docs/knockout-bracket.png" width="32%" alt="Knockout bracket" /> <img src="docs/standings-light.png" width="32%" alt="Standings (light theme)" /> </p>

What it does

Tool What it answers Widget
world_cup Anything about the World Cup β€” groups, knockout bracket, live/upcoming matches. Views: overview Β· groups Β· knockout Β· matches. Tabbed World Cup hub with a knockout bracket
list_matches "What football is on today?", live scores, a date's fixtures, one competition's matches. Scoreboard grouped by tournament, live pulse
get_match Full detail of one match: score, goal/card timeline, team stats, momentum. Match page with momentum chart & stat bars
get_standings A competition's league table (by name or SofaScore id). Standings table (multi-group aware)
search_football Find a team / player / competition / manager by name. Result list that drills into the other tools

Widgets are theme-aware (light/dark follow window.openai.theme), use SofaScore crest/flag images with an initials fallback, and are interactive β€” match rows and search results call back into other tools via window.openai.callTool.

Architecture

src/
  server.ts            Express + StreamableHTTPServerTransport (POST /mcp, /health)
  mcp.ts               McpServer factory: server instructions + resources + tools
  resources.ts         registers each built widget as a ui:// text/html+skybridge resource
  sofascore/           data layer: SofaScore client (browser headers, TTL cache,
                       optional proxy, server-side image fetch) + ESPN fallback
                       source, both behind one FootballApi interface (provider.ts,
                       fallback.ts, espn.ts); raw→structuredContent mappers, WC logic
api/index.ts           Vercel serverless entry (re-exports the Express app)
  shared/shapes.ts     structuredContent types shared by server + widgets
  tools/               list_matches Β· get_match Β· get_standings Β· world_cup Β· search_football
widgets/               React sources (built per-widget to one self-contained HTML)
  shared/              window.openai bridge + hooks, theme, reusable components
scripts/               build-widgets Β· smoke (wiring) Β· screenshots (visual)
test/fixtures/         demo SofaScore responses for offline runs

Each widget is built by vite + vite-plugin-singlefile into a single dist/widgets/<name>/index.html (JS + CSS inlined), which is served verbatim as the skybridge resource β€” no separate asset host needed.

Images (why they route through this server)

Team/tournament/player crests can't be loaded straight from api.sofascore.com inside the ChatGPT Apps sandbox: the iframe's CSP (img-src) blocks the host, and SofaScore itself 403s the referer-less cross-origin request. So the server exposes an image proxy and the widgets point at it instead:

GET /img/team/:id          β†’ https://api.sofascore.com/api/v1/team/:id/image
GET /img/tournament/:id     β†’ …/unique-tournament/:id/image
GET /img/player/:id         β†’ …/player/:id/image
GET /img/flag/:alpha2       β†’ …/img/flags/:alpha2.png

Only these allow-listed shapes are proxied. resources.ts injects window.__SOFA_BASE__ (from PUBLIC_BASE_URL) into each widget's <head>, and widgets/shared/img.ts builds ${__SOFA_BASE__}/img/... URLs from it. If PUBLIC_BASE_URL is unset (e.g. local MCP Inspector) the widgets fall back to loading images directly from SofaScore.

Run it

npm install
npm run dev          # builds widgets, then starts the server on :3000 (POST /mcp)

Point an MCP client at http://localhost:3000/mcp, or run with SOFA_MOCK=1 to serve the bundled fixtures instead of the live API:

SOFA_MOCK=1 npm run dev

Environment variables

Variable Purpose
PORT HTTP port (default 3000).
SOFA_MOCK 1 serves the bundled fixtures instead of the live API β€” no network needed.
PUBLIC_BASE_URL Public origin this server is reachable at (e.g. https://sofa.example.com). Needed for the image proxy so crests load inside the ChatGPT sandbox. On Vercel it defaults to the deployment URL. Unset elsewhere β†’ widgets load images directly.
SOFA_PROXY Outbound HTTP(S) forward proxy for all SofaScore fetches (data + images). Point at a residential / unblocked proxy to defeat the Varnish/IP block. Falls back to HTTPS_PROXY.
ESPN_ONLY 1 skips SofaScore entirely and serves everything from the ESPN fallback. Handy on hosts where SofaScore is always blocked (avoids the per-request probe + cooldown).
# Live, behind a proxy, with images proxied through this origin:
PUBLIC_BASE_URL=https://sofa.example.com \
SOFA_PROXY=http://user:pass@proxy-host:port \
npm run dev

Add it to ChatGPT (Developer Mode)

  1. Expose the server over HTTPS (e.g. ngrok http 3000).
  2. In ChatGPT β†’ Settings β†’ Connectors β†’ Advanced β†’ Developer mode, add a new connector pointing at https://<your-tunnel>/mcp.
  3. Ask: "Show me the World Cup group standings" or "What football is live right now?" β€” the matching widget renders inline.

Deploy to Vercel

The repo is Vercel-ready: api/index.ts re-exports the Express app as a single serverless function and vercel.json rewrites every path to it.

vercel deploy        # first run links the project; prod: `vercel --prod`
  • Build β€” vercel.json's buildCommand runs npm run build, bundling the five widget HTML files into dist/widgets/; includeFiles ships them with the function (the resource loader also resolves them from process.cwd()).
  • Base URL β€” PUBLIC_BASE_URL defaults to the deployment's VERCEL_URL, so the image proxy works with no extra config. Set it explicitly for a custom domain.
  • Data β€” ESPN fallback means live data works immediately. To keep SofaScore primary, add SOFA_PROXY in the Vercel project's Environment Variables; to drop it, set ESPN_ONLY=1.
  • Connector URL β€” point ChatGPT at https://<deployment>/mcp.

The function's maxDuration is 30s to cover ESPN's multi-league fan-out; the landing page at / is served statically from public/.

Verify

All verification runs offline (no live network needed):

npm run typecheck        # server + widgets typecheck
npm run build:widgets    # bundle the 5 self-contained widget HTML files
npm run verify           # build + wiring smoke test + Playwright screenshots
  • scripts/smoke.ts boots the server with fixtures, connects an in-process MCP client, and asserts every tool's openai/outputTemplate resolves to a registered ui:// resource with non-empty structuredContent.
  • scripts/screenshots.ts renders each built widget with the real server output (Chromium, window.openai set exactly like the host) into .artifacts/shots/.

Note on data & this environment

The primary source is the live SofaScore API (api.sofascore.com), used with browser-like headers. Its edge (Varnish) returns 403 to datacenter IPs and flagged TLS fingerprints regardless of headers β€” so a cloud host, CI runner, or this sandbox often can't reach it. This is an egress-IP block, not a missing header.

So the server falls back to ESPN's public soccer API automatically. Every data source sits behind one FootballApi interface; FallbackApi tries SofaScore first (honouring SOFA_PROXY) and, on any 403 / timeout / empty result, transparently serves the same request from ESPN β€” mapped into the exact same shapes, so tools and widgets never know the difference. ESPN is reachable from blocked networks, covers the FIFA World Cup + the top-5 European leagues, and needs no API key. Result: live data flows out of the box, no proxy required. Team crests come from a.espncdn.com via the same image proxy.

Ways to control the source:

  1. Default β€” SofaScore β†’ ESPN fallback. Just works anywhere.
  2. ESPN_ONLY=1 β€” skip SofaScore entirely (best on always-blocked hosts).
  3. SOFA_PROXY=<url> β€” give SofaScore a residential / unblocked egress so it stays primary (see below).
  4. SOFA_MOCK=1 β€” bundled fixtures, fully offline (dev / tests / demos).

The World Cup season resolves dynamically (not hardcoded).

Getting a proxy endpoint

SOFA_PROXY takes a standard http(s)://[user:pass@]host:port URL. What you need is an IP SofaScore hasn't blocked β€” in practice a residential or mobile proxy (datacenter proxies are usually blocked too). Options:

  • Managed residential proxy providers β€” Bright Data, Oxylabs, Smartproxy, IPRoyal, Soax, etc. Sign up, create an endpoint, and they give you a http://user:pass@gateway:port URL. Paste it into SOFA_PROXY. Cheapest to start; billed per GB (crest images are tiny, JSON is tiny β€” usage is low).
  • Your own home/residential IP β€” run the server on a machine at home, or tunnel through one (e.g. a small box on a home connection running ngrok / tailscale + a forward proxy like tinyproxy or squid). Free if you already have the hardware.
  • A cheap VPS on a residential-friendly ASP β€” some VPS/edge providers hand out IP ranges SofaScore doesn't block; try a curl from the box first (below).

Test any candidate before wiring it in:

curl -x "http://user:pass@proxy-host:port" \
  -H "User-Agent: Mozilla/5.0" -H "Referer: https://www.sofascore.com/" \
  -o /dev/null -w "%{http_code}\n" \
  https://api.sofascore.com/api/v1/sport/football/events/live

200 β†’ set that URL as SOFA_PROXY. 403 β†’ that IP is blocked too; try another.

Not affiliated with SofaScore; uses their public endpoints for informational use.

Recommended Servers

playwright-mcp

playwright-mcp

A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.

Official
Featured
TypeScript
Magic Component Platform (MCP)

Magic Component Platform (MCP)

An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.

Official
Featured
Local
TypeScript
Audiense Insights MCP Server

Audiense Insights MCP Server

Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.

Official
Featured
Local
TypeScript
VeyraX MCP

VeyraX MCP

Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.

Official
Featured
Local
graphlit-mcp-server

graphlit-mcp-server

The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.

Official
Featured
TypeScript
Kagi MCP Server

Kagi MCP Server

An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.

Official
Featured
Python
Neon Database

Neon Database

MCP server for interacting with Neon Management API and databases

Official
Featured
E2B

E2B

Using MCP to run code via e2b.

Official
Featured
Exa Search

Exa Search

A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.

Official
Featured
Qdrant Server

Qdrant Server

This repository is an example of how to create a MCP server for Qdrant, a vector search engine.

Official
Featured