ArXiv MCP Server

ArXiv MCP Server bridges MCP clients and arXiv so AI assistants can search papers, download full text, read locally stored content, and run research workflows.

Install and Launch

The documented simple stdio launch is:

{
  "mcpServers": {
    "arxiv": {
      "command": "uvx",
      "args": ["arxiv-mcp-server"]
    }
  }
}

Manual installation is documented with uv tool install arxiv-mcp-server, or uv tool install 'arxiv-mcp-server[pdf]' for older papers that require PDF fallback. A custom storage directory can be supplied with --storage-path when launching via uv tool run arxiv-mcp-server.

HTTP mode is available with environment variables, for example TRANSPORT=http HOST=127.0.0.1 PORT=8080 arxiv-mcp-server --storage-path /path/to/papers, and serves Streamable HTTP at /mcp.

Capabilities

Core tools include search_papers, download_paper, list_papers, read_paper, and get_abstract. Additional tools include semantic_search, reindex, citation_graph, watch_topic, and check_alerts. The server also provides research prompts including deep paper analysis, summarization, paper comparison, and literature review workflows.

Configuration

Runtime configuration is through --storage-path and environment variables: MAX_RESULTS, BATCH_SIZE, REQUEST_TIMEOUT, TRANSPORT, HOST, PORT, ALLOWED_HOSTS, and ALLOWED_ORIGINS. Optional APP_NAME and APP_VERSION are source-level Settings fields but are not normally needed in launch config.

Limitations and Security

Requires Python 3.11 or newer and network access to arXiv. Paper content is untrusted external input and may contain prompt injection attempts; the README and SECURITY.md recommend sandboxing and caution in multi-tool/agentic workflows. Semantic search and advanced research features require optional [pro] dependencies, and PDF fallback requires the [pdf] extra. Streamable HTTP enables DNS rebinding protection and should be kept bound to localhost unless upstream authentication and network controls are in place.

Wardn Hub Source Review Notes

• Wardn import was called first but returned 404 with no usable imported serverJson; no imported transport fields were available to preserve.
• The repository includes server.json metadata with PyPI package arxiv-mcp-server version 0.5.0; placeholder package arguments from that file were normalized into optional package argument metadata.
• Package identifier and version are split: PyPI identifier arxiv-mcp-server, version 0.5.0.
• Default launch args are only ["arxiv-mcp-server"]; optional --storage-path, local checkout --directory, and HTTP env settings are not forced into the default stdio launch.

Upstream version evidence

Reviewed upstream metadata identifies this server as version 0.5.0 in pyproject.toml, PyPI metadata, and upstream server.json. The repository metadata uses https://github.com/blazickjp/arxiv-mcp-server.