Skip to content

Bibliographic interoperability: BibTeX, RIS and CSL-JSON

The opt-in export_bibliography MCP tool serializes up to 100 caller-supplied bibliographic records into three common reference-management formats:

  • BibTeX: conventional @article/@misc entries, with TeX reserved characters escaped to prevent field/control-sequence injection.
  • RIS: one TY / TI / AU / JO / PY / DO record per work. Newlines and control characters in untrusted fields are normalized first.
  • CSL-JSON: JSON citation objects with title, issued, container-title, DOI, author names as literal values and unique IDs. This is a reference-format export, not a Zotero synchronization connector.

The tool accepts the standardized paper dictionaries already returned by search_literature or get_paper, with fields title (required), authors (list), journal, year, and doi. The caller may provide metadata_disagreements from cross-provider search.

MCP example

{
  "papers": [{
    "title": "Synthetic example: unverified sample study",
    "authors": ["Ada Example", "Alex Test"],
    "journal": "Example Journal",
    "year": 2024,
    "doi": "10.1234/example.0001"
  }],
  "output_format": "csl-json"
}

Choose bibtex (default), ris or csl-json. The response includes the encoded content, record_count, a records audit array with per-item doi_status, fields_exported, metadata_conflicts_present, and an explicit scientific_claim_verified=false / doi_resolved=false.

Scientific and data integrity limits

  • No network access and no fresh DOI or publisher authentication occurs. SYNTAX_VALID_NOT_RESOLVED means DOI format is valid, not resolved. Malformed DOI metadata is deliberately omitted and marked INVALID_DOI_OMITTED.
  • No inferred fields: missing authors, journal, year or DOI are omitted. Independently supplied DOI-less records are not merged by title; keys are unique and stable within each call, not guaranteed globally.
  • Metadata disagreements remain visible as an explicit audit flag. Conversion is not authoritative bibliographic verification.
  • No human reference manager connected: exports are importable citation representations, not evidence of success importing them into any external product. BibTeX Unicode import compatibility depends on the client's TeX engine.
  • No scientific endorsement: an accurately formatted citation cannot establish causal validity, risk of bias, study quality or scientific truth. This feature is orthogonal to HealthVer, SciFact and HealthFC metrics.

Reproducibility

python -m pytest -q tests/test_bibliography.py tests/test_mcp_protocol.py
pytest -q --cov=literature_evidence_mcp --cov-branch --cov-fail-under=100

The MCP interface is also exercised through the real Python and TypeScript client protocol tests in CI. The feature adds no third-party runtime package and performs no npm Registry publication.