DataCite Librarian MCP (mcp-test)-test

Local [Model Context Protocol](https://modelcontextprotocol.io) server for the DataCite community: repository QA, funder compliance, CSV index analytics, search/facets, and exports over DataCite monthly and public datafiles you host on disk.

Built with FastMCP · uv · Python 3.12+ · MIT

This repository does not ship production datafiles. You must download them yourself and keep them outside git (see below). CI runs only against a small mock corpus.

Strict requirements

1. You must obtain datafiles from DataCite, not from this repo.
2. Never commit real part_*.jsonl.gz, YYYY-MM.csv.gz, TAR archives, or monthly/public extracts. They are gitignored under data/local/ and at the repo root.
3. Set `DATACITE_DATA_DIR` to your extracted data directory. If the variable is set but the path does not exist, the MCP errors (it does not silently use mock data).
4. Mock data is for demos/tests only (or when DATACITE_DATA_DIR is unset / DATACITE_USE_MOCK=1). It is not a substitute for real monthly/public files.
5. Aggregate tools scan a limited number of records by default (DATACITE_MAX_RECORDS, default 10000). Always inspect truncated, scan_limit, and (for indexes) coverage_pct so you do not treat a sample as the full corpus.
6. Metadata vs index: a YYYY-MM.csv.gz index can list hundreds of thousands of DOIs; you need the matching part_*.jsonl.gz files for full QA/search. Use coverage_report to measure the gap.
7. Respect DataCite access terms: the public annual file is openly documented; the monthly file is for DataCite Members and Consortium Organizations (authenticated S3 access). See official docs linked below.

Obtain DataCite datafiles (official documentation)

Follow only DataCite’s documentation and portals. Do not rely on third-party mirrors unless you trust them and accept their terms.

High-level download steps (summary only — details are on Support)

Public annual file

1. Open datafiles.datacite.org and locate the latest public release (e.g. public-2025).
2. Download the TAR (or equivalent) per the Public Data File page.
3. Extract locally to a directory you control (recommended: this repo’s data/local/, which is gitignored).
4. Confirm you see something like dois/updated_YYYY-MM/part_*.jsonl.gz and/or monthly YYYY-MM.csv.gz indexes inside the extract.

Monthly file (members)

1. Confirm your organization is a DataCite Member or Consortium participant.
2. Follow Monthly Data File for temporary AWS credentials and S3 access.
3. Sync/extract to data/local/ (or another path outside git).
4. Point DATACITE_DATA_DIR at that root.

After download — required layout for this MCP

Preferred (matches DataCite releases):

data/local/                    # or any path you pass as DATACITE_DATA_DIR
  STATUS.json                  # optional
  MANIFEST.json                # optional
  dois/
    updated_2026-06/
      2026-06.csv.gz           # index: doi, state, client_id, updated
      part_0000.jsonl.gz       # full metadata (~10k records/part typical)
      part_0001.jsonl.gz
      …

Also supported (experimental/flat):

data/local/
  2026-06.csv.gz
  part_0000.jsonl.gz

See data/local/README.md and docs/DATAFILE_SCHEMA.md.

Quick start (development)

Prerequisites

• Python 3.12+
• uv

Install

git clone <this-repo-url> mcp-test
cd mcp-test
uv sync --all-groups

Run MCP (stdio — for Cursor / Claude Desktop / other MCP hosts)

# Demo/mock only (no real datafiles)
uv run datacite-librarian-mcp

# Production/local datafiles (STRICT: path must exist)
export DATACITE_DATA_DIR="/absolute/path/to/data/local"
export DATACITE_MAX_RECORDS=20000   # optional; raise for fuller scans
uv run datacite-librarian-mcp

Example MCP host config (mcp.json pattern):

{
  "mcpServers": {
    "datacite-librarian": {
      "command": "uv",
      "args": [
        "run",
        "--directory",
        "/absolute/path/to/mcp-test",
        "datacite-librarian-mcp"
      ],
      "env": {
        "DATACITE_DATA_DIR": "/absolute/path/to/mcp-test/data/local",
        "DATACITE_MAX_RECORDS": "20000"
      }
    }
  }
}

Run MCP (HTTP, local testing)

export DATACITE_DATA_DIR="/absolute/path/to/data/local"
uv run python -c "from datacite_librarian_mcp.server import mcp; mcp.run(transport='http', host='127.0.0.1', port=8765)"

Natural-language REPL (maps questions → tools; not a full LLM)

export DATACITE_DATA_DIR="/absolute/path/to/data/local"
uv run datacite-librarian-chat
# or: uv run python scripts/interactive_client.py

Examples: how many DOIs?, how many funders?, repository health for zenodo, funder compliance for European Commission.

Who this is for

Call `community_guide` from any MCP client for persona-oriented workflows.

Tools (summary)

Discovery: community_guide, server_info, corpus_status, corpus_inventory, diff_partitions_summary

Metadata QA / compliance (needs part_*.jsonl.gz): repository_health, funder_compliance, search_dois, get_doi, check_doi_qa, list_clients, list_funders

Analytics: facets, top_subjects

CSV index only (no JSONL required): index_summary, index_client, coverage_report

Exports (writes under exports/ or DATACITE_EXPORT_DIR): export_health_issues, export_funder_issues, export_search_results

Ops: regenerate_mock_data

Configuration

Development & CI

uv sync --all-groups
uv run pytest
uv run ruff check src tests

GitHub Actions (.github/workflows/ci.yml) runs ruff + pytest on Python 3.12 and 3.13 with DATACITE_USE_MOCK=1 only—no real datafiles in CI.

Project docs:

• docs/DATAFILE_SCHEMA.md — file/record schema notes
• docs/DEVELOPMENT.md — contributor workflow
• data/local/README.md — where to put downloads

Design principles

1. Local-first — organizations keep datafiles; this project never distributes bulk DOI corpora.
2. Stream-read — gzip JSONL/CSV line-by-line; suitable for large files without a database.
3. Light dependencies — fastmcp + pydantic (+ stdlib).
4. Honest limits — truncated, scan_limit, coverage_pct on tool outputs.
5. Index without metadata — CSV tools help before all part_*.jsonl.gz are downloaded.

License

MIT — see LICENSE.

DataCite bulk metadata licensing and access are governed by DataCite (public file documentation typically describes CC0 for metadata; confirm on Support). Member monthly access may be restricted. This software does not redistribute production datafiles.

Links

• Data files portal
• Public data file docs
• Monthly data file docs
• DataCite Metadata Schema
• MCP specification
• FastMCP