dspace-mcp 0.2.0

Read-only MCP server for DSpace 7+ repositories

dspace-mcp

A read-only MCP server that lets an AI assistant talk to any DSpace 7+ repository.

Ask your repository questions in plain language — "how many theses did we publish in 2025?", "which authors appear most often in this collection?", "summarise the PDF attached to this item" — and the assistant answers by querying the DSpace REST API directly.

Read-only by construction, not by promise

The server holds no credentials and never issues a request other than GET. It cannot deposit, edit or delete anything, and it cannot read what the public cannot already see — embargoed items, workflow submissions and restricted collections stay invisible.

That is a property of the code, not of the model's behaviour: there is nothing to prompt-inject your way into. A test in the suite asserts that no other HTTP method ever leaves the process.

Install

Nothing to install if you have uv:

uvx dspace-mcp --base-url https://demo.dspace.org/server

Or with pip:

pip install dspace-mcp

Configure your MCP client

Claude Code:

claude mcp add dspace -- uvx dspace-mcp --base-url https://demo.dspace.org/server

Claude Desktop / any client using mcp.json:

{
  "mcpServers": {
    "dspace": {
      "command": "uvx",
      "args": ["dspace-mcp", "--base-url", "https://demo.dspace.org/server"]
    }
  }
}

Point --base-url at your own repository's REST API — usually your DSpace URL with /server appended. If you leave /server off, the server detects it and corrects the URL for you.

To connect to several repositories, add several entries under different names. One process serves exactly one repository, which keeps the assistant from mixing them up.

Tools

Tool	What it does
search_items	Search items; filter by year range, author, collection; sort by relevance, date or title. limit=0 returns just the count.
get_item	Fetch one item by UUID, Handle or DOI — whichever identifier you happen to have.
list_communities	Walk the community tree (up to 3 levels).
list_collections	List collections, of one community or of the whole repository.
list_bitstreams	List an item's files with sizes, MIME types, checksums and download URLs.
get_bitstream_text	Extract the text of a bitstream so the assistant can read or summarise it — PDF, Word (.docx/.doc), OpenDocument (.odt/.ods/.odp) and Office XML (.pptx/.xlsx).
list_facet_values	Count values of a facet (authors, subjects, years) — the repository does the counting, so no records are downloaded.
get_item_statistics	View count of an item.
get_repository_info	Name, version, item counts, and which search filters, sort fields and facets this instance actually supports.

Two things worth knowing

Ask get_repository_info first. DSpace installations differ: the set of search filters and facets is configured per instance, so a filter that exists on one repository returns an error on another. This tool reports what the instance in front of you supports.

Counting is free. search_items with limit=0, and list_facet_values, answer "how many" questions with a single request and a handful of tokens, instead of downloading records and counting them.

Configuration

Variable	Default	Meaning
DSPACE_BASE_URL	(required)	REST API root, e.g. https://demo.dspace.org/server
DSPACE_TIMEOUT	15	seconds per HTTP request
DSPACE_MAX_RESULTS	50	hard ceiling on how many records any tool may return
DSPACE_EXTRACT_MAX_MB	20	refuse to download bitstreams larger than this for text extraction

Every variable has a matching flag: --base-url, --timeout, --max-results, --extract-max-mb.

DSPACE_PDF_MAX_MB / --pdf-max-mb still work as backward-compatible aliases for DSPACE_EXTRACT_MAX_MB / --extract-max-mb from before text extraction covered more than PDF.

Compatibility

Nothing here branches on a version number. The server asks each instance which search filters, sort fields and facets it supports and works from that answer, which is what actually varies between installations — two sites on the same DSpace version can differ more than two versions of the same site.

The offline test suite runs against recorded responses from a live DSpace 10.1-SNAPSHOT (demo.dspace.org), covering every endpoint the server touches, plus root and search responses from a vanilla 7.6.5, a DSpace-CRIS 8.2 and an 11.0-SNAPSHOT for comparing response shape. The contract tests (pytest -m live) run against one instance at a time — demo.dspace.org by default, or point DSPACE_TEST_URL at your own.

While the tool was being designed, its assumptions about the REST API were checked by hand against live instances of 7.2.1, 7.5, 7.6.5, 8.2, 8.4, 9.2, 10.1 and 11.0-SNAPSHOT. That was a survey of the API, not a test run of this code. If it misbehaves on a version or a configuration I could not test, a bug report is welcome.

Development

git clone https://github.com/mpasternak/dspace-mcp
cd dspace-mcp
uv sync --dev
uv run pytest              # unit tests, offline
uv run pytest -m live      # contract tests against demo.dspace.org
uv run ruff check .

Contributions are welcome. The design and its rationale — including why several plausible-sounding assumptions about the DSpace API turned out to be wrong — live in docs/superpowers/specs/.

License

MIT — see LICENSE.