start.gg MCP Server

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

• 16 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-After support
• 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
• Compact JSON output (no pretty-printing) to keep tool results small

Requirements

• Node.js >= 22
• A start.gg API token

Getting a start.gg API token

1. Log in to start.gg
2. Open developer settings (Profile → Developer Settings)
3. 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

From npm (recommended)

# run without installing
npx startgg-mcp-server

# or install globally
npm install -g startgg-mcp-server

Requires STARTGG_TOKEN in the environment; MCP clients normally launch it for you (see the next section).

From source

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 -- npx -y startgg-mcp-server

Running from a source checkout instead:

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": "npx",
      "args": ["-y", "startgg-mcp-server"],
      "env": {
        "STARTGG_TOKEN": "YOUR_TOKEN"
      }
    }
  }
}

From a source checkout, use "command": "node" and "args": ["/path/to/startgg-mcp-server/dist/cli.js"] instead.

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

Tournament

Event

Player

Utility

Tournament/event tools accept either a numeric id, a slug, or a full start.gg URL. Scheme-less URLs (start.gg/tournament/...) and <t>/event/<e> slugs are accepted too — 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 < 0 means losers bracket; round is the human-readable name
• a score of -1 is start.gg's disqualification marker
• unstarted "preview" sets have string ids like "preview_3430499_2_0"
• state names are decoded from the integer stateRaw; both are always returned
• entrant1/entrant2 use a players array, so doubles/teams work unchanged

Games and characters

get_set_games (and get_event_sets with includeGames: true) add per-game details when start.gg has them:

{
  "orderNum": 1,
  "winnerEntrantId": 24480092,
  "stage": { "id": 484, "name": "Small Battlefield" },
  "selections": [{ "entrantId": 24481002, "character": { "id": 1846, "name": "Kazuya" } }]
}

derivedCharacters summarizes unique character names per entrant in first-appearance order across games. This is derived, not an API field.

Character and stage data only exists when the set was reported with it. In a 2026-09 sample, late rounds (Top 8 / streamed) had it on about 90% of completed sets; early pools about 5%. An empty games array means not reported, not that no games were played. Typical flow: get_event_sets (e.g. phaseIds for Top 8) → pick set ids → get_set_games. For many sets at once, get_event_sets with includeGames: true.

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

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.

Out-of-range or non-numeric values for STARTGG_RATE_LIMIT / STARTGG_TIMEOUT_MS fall back to the default and log a warning on stderr.

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
• .env files are git-ignored; use .env.example as 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 (honoring Retry-After) and transient 5xx errors with exponential backoff, at most 3 retries — GraphQL errors are never retried
• caps perPage per tool so responses stay under the 1000-object complexity limit (sets are expensive: ~26+ objects each, hence perPage <= 30)
• caps fetchAll at 5 pages per tool (output is meant for an LLM context, so it stays around 100 KB of compact JSON) and reports truncated: true when 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

MIT