Biomcp

What genomoncology/biomcp does

BioMCP combines a command-line interface and MCP server around one grammar for biomedical discovery and analysis. The server gives agents access to live public data from 72 biomedical sources, including PubMed, PubTator3, Europe PMC, ClinicalTrials.gov, MyVariant.info, ClinVar, Reactome, OncoKB, UniProt, and many others. Its coverage includes genes, variants, articles, authors, trials, diagnostics, drugs, diseases, pathways, proteins, adverse events, pharmacogenomics, and cell lines.

The central operations are discovery with search, concept resolution with discover, focused retrieval with get, and cross-entity navigation through helper commands. The same interface supports literature evidence, clinical-trial lookup, molecular and disease research, drug information, pathway exploration, and related biomedical pivots.

How it works

Install the biomcp executable and run biomcp serve to expose the local stdio MCP server. An MCP client sends requests using the BioMCP command model, while the server gathers results from the relevant upstream providers. Literature searches can combine PubTator3, Europe PMC, and PubMed, with optional Semantic Scholar support when filters allow it. Results are deduplicated across common article identifiers such as PMID, PMCID, and DOI.

The command model also supports search all for cross-entity orientation, batch for up to 10 focused get calls, and enrich for g:Profiler gene-set enrichment. Article helpers can retrieve citations, references, recommendations, and recognized entities. Local study workflows provide querying, cohort construction, survival analysis, comparison, co-occurrence analysis, and terminal, SVG, or PNG charts for downloaded cBioPortal-style datasets.

Setup and configuration

Install with the BioMCP installer, PyPI, Homebrew, Docker, or from source. For PyPI, the package name is biomcp-cli; the unrelated biomcp package should not be used. A typical local MCP configuration is:

{
  "mcpServers": {
    "biomcp": {
      "command": "biomcp",
      "args": ["serve"]
    }
  }
}

Claude Code can use the provided plugin, and Codex can register the same executable with codex mcp add biomcp -- biomcp serve. Docker can run the stdio server without a local binary. For shared deployments, biomcp serve-http --host 127.0.0.1 --port 8080 serves MCP at /mcp; health routes include /health, /readyz, and /.

Tools and capabilities

• Search biomedical entities with filters such as genes, diseases, drugs, and trial status.
• Retrieve sections for genes, variants, articles, trials, pathways, proteins, and other supported entities.
• Pivot between related records instead of reconstructing filters manually.
• Trace article citations, references, recommendations, and extracted entities.
• Run gene-set enrichment and parallel focused retrievals.
• Follow shipped workflows with biomcp skill list, skill render, and skill install.
• Analyze downloaded study datasets with local cohort, survival, comparison, and co-occurrence commands.

Limitations and notes

Returned data remains subject to the terms of the upstream sources. Some providers restrict commercial or clinical use, so downstream users must review the applicable source terms and obtain required licenses before relying on results. Several capabilities use local bundles or exports, while some upstream features are optional; for example, Semantic Scholar authentication is supported through S2_API_KEY.

BioMCP reaches multiple independent biomedical systems, so result scope and available sections depend on the entity, filters, and providers involved. The project distinguishes detail-card entities from search-only surfaces to avoid treating every searchable concept as retrievable with get.