Biomed MCP Server

biomedmcp

BioMedMCP gives your AI agent cited access to biomedical sources, including PubMed, UniProt, ClinicalTrials.gov, ChEMBL, Open Targets and GTEx, so its answers point to records you can check.

Start with a research question

Try this with your MCP client:

Build an evidence notebook on LDL-receptor biology and cholesterol lowering. Start with a human LDLR target dossier, inspect tissue expression, then find PCSK9 genetic studies. Separate observations, intervention results and open questions. Cite the records you read and identify any unavailable sources.

Connect to https://biomedmcp.unibiointelligence.com/mcp using Streamable HTTP and sign in when prompted. For a local server, install the Python distribution or build from this checkout; see connection options.

Run the same investigation locally

Use Python 3.13+ in an activated virtual environment:

python -m pip install biomedmcp==0.2.5
UBI_PROFILE=local biomedmcp build_dossier '{"workflow":"target_dossier","focus":"LDLR","species":9606}'

The single Python distribution includes all source adapters. From a source checkout, uv sync --frozen installs the locked environment; prefix commands with uv run --frozen when using it.

These focused calls can fill gaps in the notebook:

UBI_PROFILE=local biomedmcp get_record '{"entity":"gene","id":"LDLR","species":9606,"sections":["expression"]}'
UBI_PROFILE=local biomedmcp search_literature '{"query":"PCSK9","filters":{"source":"pubmed","date_from":"2024-01-01"}}'
UBI_PROFILE=local biomedmcp get_variant_evidence '{"id":"PCSK9 R46L"}'

Check each response's source, species and identifiers before combining records; a partial or unavailable section is a gap, not a negative result. Follow its continuation and recovery links to see more than the preview.

Choose the next question

Connect a local agent

UBI_PROFILE=local biomedmcp-server --transport http --host 127.0.0.1 --port 8000

Use http://127.0.0.1:8000/mcp, or configure a stdio client using the connection guide. The call reference covers exact JSON arguments; result handling explains partial results and evidence boundaries.

Some sources need a key or installed dataset. Run biomedmcp diagnose and consult source access. Keep credentials in environment variables or client secret storage. A failed source doesn't mean there is no evidence.

The integration code uses Apache-2.0. Retrieved data retains its source terms. See notices, release notes and contributing.

Acknowledgements

We gratefully acknowledge support from Anthropic's AI for Science rare disease research grants.