sciverse-mcp-server

What it does

sciverse-mcp-server is OpenDataLab's MCP server for Sciverse, a hosted index of academic papers and patents. It runs on your machine over stdio and passes each tool call to the Sciverse API with your token.

There are six tools in the published package. search_metadata does structured search over papers, patents, authors or sources, with filters, sorting, facets and cursor paging. semantic_search pulls matching text chunks from papers so an agent can quote them. read_content returns a slice of a paper's or patent's text, 4096 code points from the start unless you pass offset and limit. list_catalog shows which fields exist and what values they take. list_paper_relations pages through a paper's citations, references or related works. get_resource fetches a figure or table and hands it back as an image block a multimodal model can look at.

The package README still lists the older search_papers tool and leaves out list_paper_relations. In the code search_papers is now a deprecated alias, offered only when SCIVERSE_ENABLE_LEGACY_TOOLS is set to true (src/server.ts, openapi.yaml). Tool names and schemas are generated from the repository's openapi.yaml at build time.

The package also installs sciverse-mcp-http, which the README skips. It serves the same tools over Streamable HTTP at /mcp on 0.0.0.0, port 8080 unless PORT is set (src/cli-http.ts).

What it needs

Get an API token at https://sciverse.space and pass it as SCIVERSE_API_TOKEN. Or run pip install sciverse && sciverse auth login once, which saves it to ~/.sciverse/credentials.json and lets you drop the env block. The server stops at start when it finds neither (src/config.ts).

Limits

• Read only All tools only read from the Sciverse index. Nothing can be uploaded, annotated or changed.
• Not covered Searches the Sciverse index only, not the open web or other paper databases.
• Not covered semantic_search covers papers only; read_content takes papers or patents, and author or source IDs cannot be read.
• Limit SCIVERSE_BASE_URL must be sciverse.space or a subdomain of it. Any other host stops the server at start (src/config.ts).
• Size cap search_metadata returns 25 results per page by default, at most 50 with page_size. semantic_search returns 10 chunks by default, up to 100 with top_k.
• Size cap read_content reads 4096 code points by default when limit is not given, up to 524288.
• Early release search_papers is deprecated and hidden unless SCIVERSE_ENABLE_LEGACY_TOOLS=true, although the README still lists it.

Read-only retrieval: it cannot upload papers, annotate or change anything on the platform. It only searches the Sciverse index, not the open web or other paper databases. SCIVERSE_BASE_URL must be sciverse.space or a *.sciverse.space subdomain; any other host stops the server at start (src/config.ts), so it cannot be pointed at a self-hosted or third-party API. The stdio program speaks stdio only; HTTP is a separate program (sciverse-mcp-http) the README does not describe.

Who it suits

Suits developers and researchers who want an agent to find papers or patents, follow citations, quote passages and look at figures, and who are fine creating a Sciverse account and token. Every query goes to the vendor's API, so it is no fit for offline work or for sources outside the Sciverse index. Setups built on the old search_papers tool should move to search_metadata or turn the legacy flag on. A team can share one instance by running the bundled HTTP program.