DaRUS MCP

Connects Claude to DaRUS, the University of Stuttgart's research data repository. Claude can then search and review datasets for you. It is read-only: it never creates, edits or publishes anything in DaRUS.

Installation

1. Get your DaRUS API token

Log in to DaRUS, open the menu under your name and choose API Token. Copy the token. It lets Claude see your unpublished and restricted datasets; keep it private. Only reviewing published datasets? You can skip this and leave the token empty.

2a. Claude Desktop

1. Download darus-mcp.mcpb.
2. Double-click it (or drag it into Claude Desktop) and click Install.
3. Paste your API token (or leave it empty for published datasets only).

You also need git installed (macOS: run xcode-select --install; Windows: git-scm.com). Claude Desktop sets up everything else itself. To change the token later: Settings → Extensions → DaRUS → Configure.

2b. Claude Code

Install uv and git, then add your token to your shell profile (~/.zshrc or ~/.bashrc):

export DARUS_API_TOKEN="<your-darus-api-token>"

Then in Claude Code:

/plugin marketplace add SimTech-Research-Data-Management/darus-mcp
/plugin install darus-review@darus

This also installs the dataset review skill: ask Claude to review a dataset and it checks the metadata, description and files for quality, FAIR compliance and internal consistency, against these review criteria.

If the DaRUS tools don't show up, check that which uv finds uv and that DARUS_API_TOKEN is set in the shell you started Claude Code from. After an update, restart Claude Code so it reloads the tool list.

Try it

• "Review the DaRUS dataset doi:10.18419/DARUS-1234"
• "Find DaRUS datasets on molecular dynamics simulations"
• "Which files and columns does doi:10.18419/DARUS-1234 contain?"

What Claude can do with it

• search datasets and collections in DaRUS
• read dataset metadata, file lists and file contents, including tabular files
• check metadata for consistency and completeness, and write structured review notes

Answers are grounded in what is actually in DaRUS, which makes Claude useful for triaging many records before a manual expert review.

Tools

Search_Dataverse, Search_DataCite, Get_Dataset_Metadata, List_Files_in_Dataset, Get_Collection_Metadata, List_Content_of_Collection, Read_File_Content, Read_Tabular_File, Get_Tabular_File_Schema, Knowledge_Graph_Summary, Query_Knowledge_Graph, Dataverse_Metrics. All read-only.

Creating and editing draft datasets, and controlled-vocabulary lookup, are kept on the write-tools branch. They need a patched pyDataverse, so they are not part of the released server.

For developers

Built on fastmcp 2 and pydataverse[mcp] (from gdcc/pyDataverse main). Needs Python 3.12+, uv and git.

Install script (Claude Desktop without the bundle)

./install-claude.sh --token "<your-darus-api-token>"

Adds a darus-mcp entry to claude_desktop_config.json (other servers are kept), with the token in plain text. Fully quit and reopen Claude Desktop afterwards.

Run locally

API_TOKEN="<your-darus-api-token>" uv run python main.py

The server starts on http://0.0.0.0:8000 using streamable-http transport.

Build the Claude Desktop bundle

npx @anthropic-ai/mcpb pack . darus-mcp.mcpb

CI attaches it to every published GitHub release.

Configuration

All variables are optional except where noted; unset means the defaults below.

Variable	Default	Purpose
API_TOKEN	— (required for python main.py)	Dataverse API token, sent as X-Dataverse-key.
DATAVERSE_URL	https://darus.uni-stuttgart.de	Dataverse instance to talk to. May include a path (e.g. a proxy at http://127.0.0.1:9000/dv/<job_id>); a trailing / is added automatically.
MCP_TOOLS	all tools	Comma-separated allowlist of tool names (see Tools). Other tools are hidden and fail as unknown. Names must match exactly; a misspelt name is silently not exposed.

Tools always talk to DATAVERSE_URL. pyDataverse's per-call base_url argument is removed from every tool, so a caller cannot redirect a tool to another host.

The cost: Search_DataCite finds datasets on other Dataverse installations and returns a base_url for each hit. Those hits can no longer be opened through this server's tools; that takes another server instance with its own DATAVERSE_URL.

Security

• Treat your API token as a secret and never commit it.
• Prefer passing credentials via environment variables or secure secret stores.

Project layout

• main.py - MCP server bootstrap, DaRUS tool registration, and the base_url / allowlist guards
• manifest.json, .mcpbignore, icon.png - Claude Desktop bundle (.mcpb). Keep version in sync with pyproject.toml.
• install_claude.py - cross-platform Claude Desktop installer (writes the launch entry + token)
• install-claude.sh - thin shell wrapper around the installer
• test_darus.py - runnable checks (uv run python test_darus.py)
• docs/ - spec and implementation plan for the create-dataset workflow
• .claude-plugin/marketplace.json - marketplace catalog so others can install the plugin
• plugins/darus-review/ - the shareable plugin: review skill + criteria + MCP wiring
• pyproject.toml - project metadata and dependencies