ref-finder-mcp

An MCP server for end-to-end academic paper workflow — search across arXiv, Google Scholar, and Semantic Scholar, read paper content section by section, and auto-generate BibTeX/APA citations.

Installation

uv sync
uv run python -m ref_finder_mcp.server

MCP Tools

| Tool | Description | |------|-------------| | search_papers | Multi-source paper search (arXiv, Google Scholar, Semantic Scholar) | | get_paper_details | Retrieve detailed paper information | | generate_citation | Generate BibTeX / APA citations | | save_paper | Save a paper to the session | | list_saved_papers | List saved papers | | remove_paper | Remove a saved paper | | export_bibliography | Export all saved papers as BibTeX / APA / Markdown | | get_author_info | Look up author info via Google Scholar | | get_recommended_papers | Get similar paper recommendations via Semantic Scholar | | list_paper_sections | List sections (table of contents) of an arXiv paper | | get_paper_section | Fetch the full text of a specific section from an arXiv paper |

Search Sources

| Source | ID prefix | Notes | |--------|-----------|-------| | arXiv | arxiv: | Preprints, category filters | | Google Scholar | scholar: | Broad coverage, author search | | Semantic Scholar | s2: | Official API, citation counts, recommendations |

Duplicates are automatically merged by arXiv ID / DOI / title when searching across multiple sources.

Reading Paper Content

list_paper_sections and get_paper_section give section-level access to arXiv paper full text via ar5iv. The workflow is designed to be token-efficient: fetch the table of contents first, then pull only the sections you actually need.

• Supported: all arXiv papers that have an HTML version on ar5iv
• Math formulas are preserved as LaTeX ($...$ / $$...$$)
• Figures and tables are replaced with caption placeholders ([Figure N: ...])
• Parsed documents are cached in-memory to avoid redundant network requests

Usage Examples

Search + Generate BibTeX

"Find the ReAct paper and generate BibTeX"

→ search_papers(query="ReAct Synergizing Reasoning Acting", sources=["arxiv"])
→ generate_citation(paper_id="arxiv:2210.03629", format="bibtex")

Multi-source Search

"Search transformer papers on arXiv and Semantic Scholar"

→ search_papers(query="transformer", sources=["arxiv", "semantic_scholar"], max_results=20)
  (duplicates are automatically merged)

Paper Recommendations

"Recommend papers similar to the ReAct paper"

→ get_recommended_papers(paper_id="arxiv:2210.03629", max_results=5)

Author Info

"What's Geoffrey Hinton's h-index and top papers?"

→ get_author_info(author_name="Geoffrey Hinton")
  (returns affiliation, h-index, citation count, top publications, etc.)

Reading Paper Content

"Summarize the Methods section of the Mixtral paper"

→ list_paper_sections(paper_id="arxiv:2401.04088")
  (returns table of contents with section IDs)
→ get_paper_section(paper_id="arxiv:2401.04088", section_id="S2")
  (returns full text of the Architectural details section)

MCP Client Configuration

{
  "mcpServers": {
    "ref-finder": {
      "command": "uv",
      "args": [
        "run",
        "--directory", "/path/to/ref-finder-mcp",
        "python", "-m", "ref_finder_mcp.server"
      ]
    }
  }
}

Development

fastmcp dev src/ref_finder_mcp/server.py      # Dev server with hot reload
uv run pytest                                  # Run unit tests
uv run pytest -m integration                   # Run integration tests (network required)
uv run ruff check --fix                        # Lint

License

MIT