startgg-mcp-server
Enables MCP clients to discover tournaments, inspect events, entrants, sets, standings, and streams for any game on start.gg using natural language.
README
startgg-mcp-server
A Model Context Protocol server for the start.gg GraphQL API. It lets MCP clients (Claude Code, Claude Desktop, and others) discover tournaments, inspect events, entrants, sets, standings, and streams for any game on start.gg using natural language.
What is this?
start.gg exposes a powerful but complex GraphQL API: entrants vs participants vs players, integer set states, complexity-limited pagination, epoch timestamps. This server wraps that API in a small set of MCP tools with:
- Normalized output — sets come back as
{ round, state: "COMPLETED", entrant1: { gamerTag, seed }, score, winnerEntrantId, ... }instead of raw GraphQL nesting - URL resolution — paste a start.gg URL, get tournament/event ids back
- Built-in rate limiting, retries, and caching tuned to start.gg's documented limits
The server is game-agnostic. Game-specific logic (e.g. Smash upset detection)
belongs in applications built on top — see
examples/smash-ultimate-watcher.
Features
- 15 read-only tools covering discovery, tournaments, events, players, streams, and URL resolution
- Input validation (Zod) on every tool — bad ids, oversized page sizes, and malformed URLs never reach the API
- Sliding-window rate limiter (default 75 req/60s vs start.gg's 80), retries with exponential backoff, and
Retry-Aftersupport - Short-TTL in-memory cache for metadata queries
- Typed error codes:
AUTH_ERROR,RATE_LIMITED,NOT_FOUND,INVALID_INPUT,STARTGG_GRAPHQL_ERROR,NETWORK_ERROR,INTERNAL_ERROR - GraphQL documents kept in
graphql/files, separate from code - The API token never appears in output, logs, or error messages
Requirements
- Node.js >= 22
- A start.gg API token
Getting a start.gg API token
- Log in to start.gg
- Open developer settings (Profile → Developer Settings)
- Create a personal access token and copy it
Treat the token like a password. This server reads it only from the
STARTGG_TOKEN environment variable.
Installation
git clone https://github.com/tomo789/startgg-mcp-server.git
cd startgg-mcp-server
npm install
npm run build
MCP client setup
Claude Code (CLI)
claude mcp add startgg --env STARTGG_TOKEN=YOUR_TOKEN -- node /path/to/startgg-mcp-server/dist/cli.js
Claude Desktop
Add to claude_desktop_config.json:
{
"mcpServers": {
"startgg": {
"command": "node",
"args": ["/path/to/startgg-mcp-server/dist/cli.js"],
"env": {
"STARTGG_TOKEN": "YOUR_TOKEN"
}
}
}
}
Any MCP client that supports stdio servers works the same way: run
node dist/cli.js (or the startgg-mcp-server bin once installed via npm)
with STARTGG_TOKEN set.
Available tools
Discovery
| Tool | Purpose |
|---|---|
search_videogames |
Find videogame ids by name (e.g. "Super Smash Bros. Ultimate" → 1386) |
search_tournaments |
General tournament search: name, videogame, country/state, date range, upcoming/past, open registration |
get_upcoming_tournaments |
Tournaments that haven't ended yet (includes in-progress), soonest first, with a days window |
get_tournaments_by_videogame |
Tournaments for one videogame id (upcoming / past / all) |
Tournament
| Tool | Purpose |
|---|---|
get_tournament |
Details, schedule, venue, events list, configured streams |
get_tournament_events |
Events (brackets) of a tournament, optionally filtered by videogame |
get_tournament_entrants |
Tournament-level participants (attendees); per-event seeding lives in get_event_entrants |
get_stream_queue |
Stream queue: streams (with derived Twitch URLs) and the sets assigned to each |
Event
| Tool | Purpose |
|---|---|
get_event |
Event details including phases (Pools, Top 8, ...) with phase ids |
get_event_entrants |
Entrants with seed, players, DQ flag; pagination or fetchAll |
get_event_standings |
Placements (use perPage: 8 for Top 8) |
get_event_sets |
Normalized sets; filter by state, phase, round, entrants, VOD presence |
Player
| Tool | Purpose |
|---|---|
get_player |
Player by id: gamer tag, prefix, linked user |
get_player_sets |
A player's recent sets across tournaments |
Utility
| Tool | Purpose |
|---|---|
resolve_startgg_url |
start.gg URL/slug → { type, tournamentId, eventId, slugs, names } |
Tournament/event tools accept either a numeric id, a slug, or a full
start.gg URL — you rarely need resolve_startgg_url explicitly, but it is
there when you want the ids.
Normalized set shape
{
"id": 106877974,
"round": "Grand Final",
"roundNumber": 3,
"state": "COMPLETED",
"stateRaw": 3,
"completedAt": "2026-08-24T07:19:34.000Z",
"entrant1": {
"entrantId": 24480092,
"name": "LittleMacMain",
"seed": 5,
"players": [{ "playerId": 3655189, "gamerTag": "LittleMacMain", "prefix": "" }],
"score": 2
},
"entrant2": { "...": "same shape" },
"score": { "entrant1": 2, "entrant2": 3, "displayScore": "LittleMacMain 2 - RenSuø 3" },
"winnerEntrantId": 24481002,
"phase": { "id": 1994001, "name": "Bracket" },
"vodUrl": null
}
Notes grounded in the live API:
roundNumber < 0means losers bracket;roundis the human-readable name- a score of
-1is start.gg's disqualification marker - unstarted "preview" sets have string ids like
"preview_3430499_2_0" statenames are decoded from the integerstateRaw; both are always returnedentrant1/entrant2use aplayersarray, so doubles/teams work unchanged
Examples
Things to ask an MCP client once connected:
Find upcoming Super Smash Bros. Ultimate tournaments this week.
Get the entrants and seeds for this start.gg tournament URL:
https://www.start.gg/tournament/.../event/...
Show me completed sets from Top 8 of that event.
Which streams are assigned to sets at this tournament?
What were the biggest seed upsets in this event?
A standalone example application (videogame lookup → upcoming tournaments →
sets → upset candidates by seed difference) lives in
examples/smash-ultimate-watcher.
Environment variables
| Variable | Required | Default | Purpose |
|---|---|---|---|
STARTGG_TOKEN |
yes | — | start.gg API token |
STARTGG_ENABLE_WRITES |
no | false |
Reserved. No write tools exist yet; the flag only logs a notice |
STARTGG_RATE_LIMIT |
no | 75 |
Requests per 60s window (hard-capped at 80) |
STARTGG_TIMEOUT_MS |
no | 30000 |
Per-request HTTP timeout |
STARTGG_CACHE |
no | on |
Set off to disable the in-memory cache |
The API endpoint is deliberately not configurable through the environment: the
token is only ever sent to api.start.gg. When using the client as a library
(tests, tooling), inject apiUrl/fetchFn via the StartggClient constructor.
Without STARTGG_TOKEN the server still starts and lists tools, but every
call returns a clear AUTH_ERROR explaining how to fix it.
Security
- The token is read from the environment only, sent only to
api.start.gg, and never included in tool output, logs, or error messages - All tools are read-only; no mutations are implemented
.envfiles are git-ignored; use.env.exampleas a template- User-supplied input is schema-validated before any request is built
Rate limits
start.gg allows 80 requests per 60 seconds and at most 1000 objects per request. This server:
- keeps a sliding-window budget below the request limit (default 75/60s)
- retries
429(honoringRetry-After) and transient 5xx errors with exponential backoff, at most 3 retries — GraphQL errors are never retried - caps
perPageper tool so responses stay under the 1000-object complexity limit (sets are expensive: ~26+ objects each, henceperPage <= 30) - caps
fetchAllat a fixed page budget and reportstruncated: truewhen it stops early
Development
npm run dev # run from source (tsx)
npm run build # compile to dist/
npm run typecheck # tsc --noEmit
npm run lint # eslint
npm run format # prettier
GraphQL documents live in graphql/*.graphql (one file per domain, multiple
named operations per file; requests select an operation via operationName).
Schema facts verified against the live API are recorded in
docs/startgg-api-notes.md — read it before
adding fields.
Testing
npm test # unit tests (fixtures/mocks only, no network)
STARTGG_INTEGRATION=1 STARTGG_TOKEN=... npm test # + 2 live API smoke tests
STARTGG_TOKEN=... node scripts/smoke.mjs # full stdio end-to-end smoke (~10 live requests)
Unit tests cover the URL resolver, normalizers, input validation, pagination, GraphQL/HTTP error handling, the rate limiter, and the cache.
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。