The pivots are what make it more than a federated search: from a variant to the trials that mention it, from a drug to its adverse events, from a paper to what cites it — without rebuilding the filters each time you cross a source boundary.
A single binary with one grammar that reaches the biomedical databases that normally each have their own API, identifiers and search habits. The same tools are available to an agent over MCP as to a person at a shell.
- A grammar rather than a tool per source: `search <entity>` for discovery, `get <entity> <id> [sections]` for a detail card, `<entity> <helper> <id>` for cross-entity pivots, `enrich` for gene-set enrichment, `batch` for parallel gets, and `search all` for counts-first orientation across entities
- Detail-card entities and the providers behind them: gene (MyGene.info, UniProt, Reactome, STRING, GTEx, Human Protein Atlas, DGIdb, ClinGen), variant (MyVariant.info, ClinVar, gnomAD v4, CIViC, OncoKB, cBioPortal, AlphaGenome), article, trial (ClinicalTrials.gov API v2 and NCI CTS), drug, disease, pathway, protein, diagnostic, adverse-event and pgx
- Search-only surfaces for gwas via the GWAS Catalog and phenotype via Monarch Initiative HPO semantic similarity
- Literature search that fans out across PubTator3 and Europe PMC, deduplicating PMID, PMCID and DOI identifiers, with an optional Semantic Scholar leg
- Cross-entity pivots that avoid rebuilding filters — variant to trials or articles, drug to adverse events or trials, disease to trials, drugs or articles, gene to trials, drugs, articles or pathways, protein to structures, and article to citations, references, recommendations or entities
- A paper trail from one known paper: citations, references, recommendations and extracted entities
- Local study analytics for downloaded cBioPortal-style datasets — query, cohort, survival, comparison and co-occurrence, with terminal, SVG and PNG charts
- `biomcp skill list` for shipped worked examples you can open as a workflow
Install the CLI with `uv tool install biomcp-cli`, a one-line binary installer, Homebrew, Docker, or the PyPI package; the MCP server is the same package started with `serve`. Claude Code has a plugin, Claude Desktop an extension, and a remote HTTP server is available. Most commands work with no credentials at all. Optional keys raise limits or unlock enrichments: `NCBI_API_KEY` for PubTator, PubMed and PMC, `S2_API_KEY` for authenticated Semantic Scholar with a dedicated 1 req/sec quota instead of the shared 1 req/2sec pool, `OPENFDA_API_KEY`, `NCI_API_KEY` for NCI trial search, `ONCOKB_TOKEN` and `ALPHAGENOME_API_KEY`. References and recommendations can come back empty for paywalled papers because of publisher elision upstream.
One command — uv tool install biomcp-cli