Literature Evidence MCP¶
Literature Evidence MCP is an open-source scientific literature and evidence provenance server for the Model Context Protocol. It searches multiple bibliographic providers, resolves and verifies citation metadata, reads full text where a PMCID is supplied, anchors supplied quotes and annotations to paragraphs with hashes, and keeps a local hash-chained evidence ledger.
It is a bibliographic, traceability, and evidence-audit toolset. It is not an automated peer reviewer, meta-analysis engine, certified PRISMA system, or scientific truth oracle. It never labels a paper scientifically proven from an abstract or citation alone, and its methodology checks are screening aids, not formal risk-of-bias judgments. Search results are non-exhaustive, and absence of a retraction notice is not a clearance certificate.
The current release is v0.5.1. It requires Python 3.11 or later and is published on PyPI:
python -m pip install literature-evidence-mcp
literature-evidence-mcp
To work on the source instead, install the checkout with
python -m pip install -e '.[dev]'.
Documentation¶
Evaluation and benchmarks¶
- Evidence audit and benchmark protocol
- Independent labeled science benchmarks
- SciFact full-corpus retrieval
- Two public-dataset exploratory evaluations
- Dual-track head-to-head benchmark
- Candidate datasets for independent evaluation
- Optional BM25 candidate retrieval
Evidence assessment¶
- Experimental offline NLI claim-evidence triage
- Exploratory NLI challenger comparison
- HealthVer three-class validation
- HealthFC syntax-only stress test
Validation protocols and scope¶
- Prospective cross-disciplinary validation protocol
- Updated general-science blind validation protocol
- General-science source inventory
- Scientific scope and residual limitations
Comparisons and features¶
- Competitor audit
- Live competitor smoke test
- Four-provider discovery and citation traversal
- Bibliographic export formats
- Optional native components
Scope note¶
Results in these pages are exploratory unless a page explicitly says otherwise. Public benchmark labels were not prospectively blinded, provider metadata can be incomplete or stale, and no independent expert-labeled superiority is claimed.