stardew-mcp

An MCP server that exposes the Stardew Valley Wiki as queryable tools.

It parses a MediaWiki XML dump of the wiki (shipped as stardewvalleywiki-dump-20260628.zip) into structured JSON at build time, then serves it through a small set of MCP tools: keyword search, optional semantic search, typed listings, and per-page/per-section retrieval.

What it gives an LLM

• search_pages(query, limit?, type?) — BM25 keyword search over page titles, intros, infobox fields, and section headings.

• vector_search(query, limit?, type?) — semantic search over sentence embeddings (all-MiniLM-L6-v2). Falls back to keyword search if embeddings aren't built.

• get_page(title, include_raw?) — full structured page: entity type, parsed infobox fields (rendered to text), intro, all sections, categories.

• get_page_section(title, section) — a single section by heading.

• list_types() — entity types (from infoboxes) with counts.

• list_by_type(type, limit?) — all page titles of one type.

• search_by_type(type, query, limit?) — keyword search scoped to one type.

Entity types are derived from the wiki's infobox templates and include: villager, fish, item (plain {{Infobox}}, e.g. crops/forage), seed, weapon, clothing, furniture, mousehat, mineral, monster, location, artifact, building, tool, animal, tree.

Prerequisites

• Node.js 22+

• The unzip binary on PATH (used to extract the dump zip)

• The dump zip at the repo root: stardewvalleywiki-dump-20260628.zip

Build

npm install
npm run build          # parse XML -> data/*.json, then compile TS

This produces:

• data/pages.json — every page, parsed (infobox fields, sections, categories, excerpt).

• data/index.json — title + type indices for fast lookup.

• dist/ — compiled server.

(Optional) Build embeddings for vector search

npm run build:embed     # downloads ~23MB model once, embeds 2005 pages

This adds data/embeddings.json. vector_search works without it (it falls back to keyword search), but the tool is more useful with embeddings. Requires the optional dependency @huggingface/transformers, which is installed by default via optionalDependencies.

Run

npm start               # node dist/server/index.js  (stdio transport)
npm run dev             # tsx src/server/index.ts    (no build step needed)

Configure an MCP client

Add the server to your MCP client config (e.g. .opencode/opencode.json or Claude Desktop's claude_desktop_config.json):

{
  "mcpServers": {
    "stardew": {
      "command": "node",
      "args": ["/absolute/path/to/stardew-mcp/dist/server/index.js"]
    }
  }
}

The server loads data/pages.json at startup; build before first run.

How the data is parsed

src/build/wikitext.ts is a focused MediaWiki parser (not a full one):

• Finds the first top-level {{Infobox ...}} template with balanced-brace scanning, and parses its key = value fields.

• Renders fields to plaintext: {{Price|N}} → "N gold", {{Name|X}} → "X", {{NPC|name|role}} → "name (role)", {{Season|Spring|13}} → "Spring 13", links [[A|b]] → "b", and strips refs/tables/file links.

• Splits sections on =-headings and strips the infobox from the intro.

• Extracts [[Category:...]].

Search (src/search/search.ts) is BM25 over tokenized excerpts with title boosting, plus optional cosine search over L2-normalized embeddings.

Scripts

Layout

src/
  build/
    wikitext.ts     # MediaWiki parser (infobox, sections, render)
    build.ts        # XML dump -> data/pages.json + data/index.json
    embed.ts        # pages -> data/embeddings.json (transformers.js)
  search/
    search.ts       # BM25 + optional vector search, data loader
  server/
    index.ts        # MCP server + tool handlers
tests/
  parse.test.ts     # parser unit tests
  e2e.test.ts       # end-to-end server tests via MCP client SDK
data/               # generated (gitignored)