Städel Museum MCP Server

A Model Context Protocol server that gives AI assistants direct access to the Städel Museum Digital Collection. It talks to the museum's public OAI-PMH interface (LIDO metadata format), so an assistant can harvest object records, read rich multilingual metadata, and display high-resolution artwork images — with an interactive browsing UI (MCP Apps) on top.

This is a personal portfolio project built to demonstrate working with the Model Context Protocol, an OAI-PMH/LIDO harvesting API, and MCP Apps interactive UIs end-to-end — not an official Städel Museum product.

Features

Because the Städel exposes its collection through OAI-PMH (a harvesting protocol, not a search engine), there is no full-text keyword search — you filter by set and/or date range, then fetch full details per object. The tool descriptions make this explicit so the assistant doesn't assume Google-style search is available.

Related MCP server: artic-mcp

Architecture

src/
├── index.ts                 # entry point (stdio or --http)
├── server-utils.ts           # transport wiring (stdio / Streamable HTTP)
├── StaedelServer.ts           # registers tools + the explorer UI resource
├── api/
│   └── StaedelApiClient.ts    # OAI-PMH client: fetch, retry, XML→JSON, error mapping
├── tools/
│   ├── ListSetsTool.ts
│   ├── SearchMuseumObjectsTool.ts
│   ├── GetObjectTool.ts        # flattens raw LIDO XML into normalized metadata
│   └── OpenStaedelExplorerTool.ts
├── ui/
│   └── explorerResource.ts     # builds the self-contained MCP App HTML for the explorer
├── types/types.ts             # Zod schemas for tool inputs/outputs
└── utils/RateLimiter.ts       # simple request-spacing limiter for the museum API

The get-museum-object tool does the interesting work: LIDO is a verbose, multilingual, deeply nested XML format (via fast-xml-parser), and GetObjectTool flattens it into a small, stable JSON shape — resolving xml:lang variants (preferring English, falling back to German), picking the highest-resolution image link, and reading each object's actual rights statement instead of assuming one blanket license.

Licensing & Attribution (important)

• Metadata harvested from the OAI-PMH interface is CC0 1.0 (public domain).

• Images are typically CC BY-SA 4.0 Städel Museum, Frankfurt am Main, though some objects carry a different rights statement (e.g. Public Domain Mark) — get-museum-object reports the actual license for each object rather than hard-coding one.

• Whenever an assistant displays or describes an image from this server, it should include the credit line reported in the license field.

Prerequisites

• Node.js v18 or later

• pnpm (via corepack)

Installation

git clone https://github.com/topoftheblock/staedel-mcp.git
cd staedel-mcp
corepack enable
pnpm install
pnpm run build

Running

Stdio (used by Claude Desktop, Claude Code, and most MCP clients):

node dist/index.js

Streamable HTTP (for testing with clients that connect over HTTP):

node dist/index.js --http
# -> http://localhost:3001/mcp  (override with PORT env var)

Add it to Claude Desktop / Claude Code

Add to your MCP client's config (e.g. claude_desktop_config.json):

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

Try it with the MCP Inspector

npx @modelcontextprotocol/inspector node dist/index.js

Development

pnpm run watch   # tsc --watch
pnpm run check   # type-check without emitting

The server is a thin, dependency-light wrapper: StaedelApiClient owns all HTTP/XML concerns, each tool class is a small, independently testable unit, and src/types/types.ts defines the Zod schemas that make tool inputs/outputs self-documenting to the model.

Configuration

Reference

• Städel Digital Collection OAI-PMH guide

• LIDO metadata schema

• Model Context Protocol

• MCP Apps