Search Academic Papers from arXiv

mcp-scholarly is an MCP server for searching accurate academic articles. The project states that more scholarly vendors will be added soon. It currently offers three search tools: search-arxiv for arXiv search with no key needed, search-google-scholar for Google Scholar through the scholarly library using a free proxy pool, and search-google-web for Google web search through the SerpBase API.

The web search tool is optional. It is only registered when the SERPBASE_API_KEY environment variable is set, and SerpBase offers a free tier. The component list describes the arXiv tool as taking a required "keyword" string argument and searching arXiv for articles related to that keyword.

Use it when an assistant needs to look up academic literature by keyword, for example to check prior work on a research question, find relevant arXiv papers, or query Google Scholar. It fits well as the search capability in a research or validation workflow.

Be careful with what the results mean. arXiv returns best-effort matches for any query, including nonsense, so a non-empty result set is not by itself evidence that prior work exists. The tool description says so, because that is the text the model reads. An empty keyword comes back as an MCP tool error rather than an empty result set, which matters because a failed search that looks like "no prior work" could put a wrong novelty score into an evidence record. Also note that search-google-scholar goes through the scholarly library and a free proxy pool and is far less predictable than arXiv search, which answers in about one second according to measurements against zorp's transport. Logging goes to stderr, and nothing but JSON-RPC reaches stdout, which is what newline-delimited stdio framing requires.

The server can be added to Claude Desktop, whose configuration file is at ~/Library/Application Support/Claude/claude_desktop_config.json on macOS and %APPDATA%/Claude/claude_desktop_config.json on Windows. It gives three configurations: a development entry that runs uv against a local directory, a published entry that runs uvx with mcp-scholarly, and a Docker entry that runs the mcp/scholarly image. To install automatically for Claude Desktop through Smithery, use:

npx -y @smithery/cli install mcp-scholarly --client claude

For packaging, it describes syncing dependencies with uv, building source and wheel distributions into the dist directory, and publishing to PyPI with uv publish, using either a token or a username and password supplied through environment variables or command flags. Because MCP servers run over stdio, debugging is easiest with the MCP Inspector, launched through npm.

Using it with zorp: zorp needs a search-capable MCP tool before its validate command will run. This server satisfies that check because zorp matches on a search verb in the tool name. You can pass the server on the command line as a stdio entry, or configure it once in a .zorp/mcp.toml file with a server named scholarly, stdio transport, the uv command, a trust level of sandbox and a timeout of 60 seconds. zorp's default stdio read budget is 30 seconds, which is comfortable for arXiv, and the 60 second value is headroom for Google Scholar.

Researchers, analysts and agent builders who need scholarly search inside an MCP-capable assistant, and teams wiring literature checks into research validation flows such as zorp.