BioMCP

One binary. One grammar. Evidence from the biomedical sources you already trust.

What is BioMCP?

BioMCP is one CLI binary over a single command grammar that reaches ~30 trusted biomedical sources (PubMed, ClinVar, ClinicalTrials.gov, OncoKB, Reactome, and more). It is also an MCP (Model Context Protocol) server, so the same tools are available to AI agents such as Claude Code, Codex, and Claude Desktop.

BioMCP cuts through the usual biomedical data maze: one query reaches the sources that normally live behind different APIs, identifiers, and search habits. Researchers, clinicians, and agents use the same command grammar to search, focus, and pivot without rebuilding the workflow for each source. You get compact, evidence-oriented results across live public data plus local study analytics.

Features

• Search the literature: search article fans out across PubTator3 and Europe PMC, deduplicates PMID/PMCID/DOI identifiers, and can add a Semantic Scholar leg when your filters support it.
• Pivot without rework: move from a gene, variant, drug, disease, pathway, protein, or article straight into the next built-in view instead of rebuilding filters by hand.
• Choose a playbook: biomcp skill list shows shipped worked examples so you can open the matching biomcp skill <slug> workflow.
• Analyze studies locally: study commands cover local query, cohort, survival, compare, and co-occurrence workflows with native terminal, SVG, and PNG charts for downloaded cBioPortal-style datasets.
• Follow the paper trail: article citations, article references, article recommendations, and article entities turn one known paper into a broader evidence map.
• Enrich and batch: use biomcp enrich for top-level g:Profiler enrichment and biomcp batch for up to 10 focused get calls in one command.

Quick start

First useful query in under 30 seconds:

uv tool install biomcp-cli
biomcp health --apis-only
biomcp skill list
biomcp list gene
biomcp search all --gene BRAF --disease melanoma  # unified cross-entity discovery
biomcp get gene BRAF pathways hpa

Watch the introduction

A project using BioMCP won St. Jude's KIDS BioHackathon in 2025. The talk that followed is a one-hour introduction to BioMCP, recorded October 24, 2025.

Watch on YouTube · Read the write-up · Full transcript

Installation

Binary install

curl -fsSL https://biomcp.org/install.sh | bash

PyPI tool install

uv tool install biomcp-cli
# or: pip install biomcp-cli

PyPI package warning: install biomcp-cli, not biomcp. The biomcp PyPI package is unrelated to this project.

MCP Registry ownership marker: mcp-name: io.github.genomoncology/biomcp.

This installs the biomcp binary in ~/.local/bin. If that directory is not already on PATH, the installer prints one command to add it; it never edits your shell startup files.

Homebrew

brew tap genomoncology/biomcp
brew install biomcp

The separate genomoncology/homebrew-biomcp tap repository must exist before these commands can work.

Docker

docker run --rm ghcr.io/genomoncology/biomcp --version
docker run --rm ghcr.io/genomoncology/biomcp list
docker run --rm -i ghcr.io/genomoncology/biomcp serve

Use the GHCR image for quick CLI checks or stdio MCP clients without a local install.

Claude Code plugin

Install the biomcp binary first, then add the hosted plugin marketplace and install the BioMCP plugin in Claude Code:

/plugin marketplace add genomoncology/biomcp
/plugin install biomcp@biomcp

The plugin wires Claude Code to the local stdio MCP server with biomcp serve. For guided BioMCP workflows, also install the skill assets below.

Codex MCP server

Install the biomcp binary first, then register the same stdio MCP server with Codex:

codex mcp add biomcp -- biomcp serve

Claude Desktop extension (.mcpb)

Install BioMCP from the Anthropic Directory in Claude Desktop when that path is available for your environment. For local/manual setups, use the JSON MCP config below.

Install skills

Install guided investigation workflows into your agent directory:

biomcp skill install ~/.claude --force

MCP clients

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

Remote HTTP server

For shared or remote deployments:

biomcp serve-http --host 127.0.0.1 --port 8080

Remote clients connect to http://127.0.0.1:8080/mcp. Probe routes are GET /health, GET /readyz, and GET /.

Runnable demo:

uv run --script examples/streamable-http/streamable_http_client.py

See Remote HTTP Server for the newcomer guide.

From source

make install
"$HOME/.local/bin/biomcp" --version

For repo-local verification, run the standard gates directly: make lint, make test, and make spec. make test includes both Rust nextest and the Python/docs contract lane, while make release-gate adds the named full-feature proof and runs specs against the all-feature release binary. There is no supported make check command. Use make verify only for opt-in live public-upstream confidence; make release-live-smoke remains a compatibility alias.

Command grammar

search <entity> [filters]    → discovery
skill list                   → playbook catalog for how-to questions
discover <query>             → concept resolution before entity selection
get <entity> <id> [sections] → focused detail
<entity> <helper> <id>       → cross-entity pivots
enrich <GENE1,GENE2,...>     → gene-set enrichment
batch <entity> <id1,id2,...> → parallel gets
search all [slot filters]    → counts-first cross-entity orientation

Entities and sources

The tables below distinguish detail-card entities from search-only surfaces so agents do not synthesize unsupported get commands.

Gettable entities

Search-only entities

Cross-entity helpers

Pivot between related entities without rebuilding filters.

See the cross-entity pivot guide for when to use a helper versus a fresh search.

biomcp variant trials "BRAF V600E" --limit 5
biomcp variant articles "BRAF V600E"
biomcp drug adverse-events pembrolizumab
biomcp drug trials pembrolizumab
biomcp disease trials melanoma
biomcp disease drugs melanoma
biomcp disease articles "Lynch syndrome"
biomcp gene trials BRAF
biomcp gene drugs BRAF
biomcp gene articles BRCA1
biomcp gene pathways BRAF
biomcp pathway drugs R-HSA-5673001
biomcp pathway drugs hsa05200
biomcp pathway articles R-HSA-5673001
biomcp pathway trials R-HSA-5673001
biomcp protein structures P15056
biomcp article entities 22663011
biomcp article citations 22663011 --limit 3
biomcp article references 22663011 --limit 3
biomcp article recommendations 22663011 --limit 3

Gene-set enrichment

biomcp enrich BRAF,KRAS,NRAS --limit 10

Top-level biomcp enrich uses g:Profiler. Gene enrichment sections inside other entity views still reference Enrichr where that is the backing source.

Sections and progressive disclosure

Every get command supports selectable sections for focused output:

biomcp get gene BRAF                    # summary card
biomcp get gene BRAF pathways           # add pathway section
biomcp get gene BRCA1 diagnostics       # diagnostic-test pivot from GTR
biomcp get gene BRAF hpa                # protein tissue expression + localization
biomcp get gene BRAF civic interactions # multiple sections
biomcp get gene BRAF all                # standard sections; diagnostics/funding stay opt-in

biomcp get variant "BRAF V600E" clinvar population conservation
biomcp get article 22663011 tldr
biomcp get drug pembrolizumab label targets civic approvals
biomcp get drug trastuzumab regulatory --region who
biomcp get disease "Lynch syndrome" genes phenotypes variants
biomcp get disease tuberculosis diagnostics
biomcp get diagnostic GTR000006692.3 regulatory
biomcp get trial NCT02576665 eligibility locations outcomes

In JSON mode, get responses expose _meta.next_commands for the next likely follow-ups and _meta.section_sources for section-level provenance. batch ... --json returns per-entity objects with the same metadata shape.

API keys

Most commands work without credentials. Optional keys improve rate limits or unlock optional enrichments:

export NCBI_API_KEY="..."        # ClinVar, PubTator, PubMed/efetch, PMC OA, NCBI ID converter
export S2_API_KEY="..."          # Optional Semantic Scholar auth; dedicated quota at 1 req/sec
export OPENFDA_API_KEY="..."     # OpenFDA rate limits
export NCI_API_KEY="..."         # NCI CTS trial search (--source nci)
export ONCOKB_TOKEN="..."        # OncoKB variant helper
export ALPHAGENOME_API_KEY="..." # AlphaGenome variant effect prediction

search article, get article, batch article, get article ... tldr, and the explicit Semantic Scholar helpers all work without S2_API_KEY. With the key, BioMCP sends authenticated requests and uses a dedicated rate limit at 1 req/sec. Without it, BioMCP uses the shared unauthenticated pool at 1 req/2sec. search article --source supports all, pubtator, europepmc, pubmed, semanticscholar, and litsense2. The default compatible article federation uses PubTator3, Europe PMC, PubMed, and automatic Semantic Scholar; use --source semanticscholar or --source litsense2 explicitly when you want one of those sources alone. Explicit source selection also disables cross-provider row enrichment. References and recommendations can be empty for paywalled papers because of publisher elision in Semantic Scholar upstream coverage.

Configuration

Claude Desktop extension settings

The directory bundle exposes only the optional settings needed for the first reviewer-facing build:

The first directory build exposes only those three optional settings. Advanced CLI-only env vars remain documented in API Keys for the general BioMCP CLI path.

Usage Examples

Public cross-entity overview

User prompt: Give me a low-noise overview of BRAF in melanoma.

Expected tool call: biomcp search all --gene BRAF --disease melanoma --counts-only

Expected behavior: Returns a cross-entity counts summary that orients the next command instead of dumping long detail tables.

Expected output: Counts-first summary with suggested next commands for the highest-yield entity follow-ups.

Public variant evidence

User prompt: Summarize ClinVar significance and population frequency for BRAF V600E.

Expected tool call: biomcp get variant "BRAF V600E" clinvar population

Expected behavior: Retrieves the focused variant card, ClinVar section, and population-frequency data in one read-only call.

Expected output: Variant summary, ClinVar significance details, and gnomAD population frequencies.

Credentialed OncoKB example

User prompt: Show OncoKB therapy evidence for BRAF V600E.

Expected tool call: biomcp variant oncokb "BRAF V600E"

Expected behavior: Uses ONCOKB_TOKEN when configured and otherwise returns helpful guidance about the missing credential.

Expected output: Therapy and level evidence when ONCOKB_TOKEN is set, or a clear setup hint when it is not.

Credentialed DisGeNET example

User prompt: Show scored DisGeNET associations for TP53.

Expected tool call: biomcp get gene TP53 disgenet

Expected behavior: Uses DISGENET_API_KEY to retrieve the scored gene-disease association section.

Expected output: Ranked disease-association table with evidence counts and scores when DISGENET_API_KEY is configured.

Privacy Policy

BioMCP does not add telemetry, analytics, or remote log upload. Review the full privacy statement at https://biomcp.org/policies/.

Multi-worker deployment

BioMCP rate limiting is process-local. For many concurrent workers, run one shared Streamable HTTP biomcp serve-http endpoint so all workers share a single limiter budget:

biomcp serve-http --host 0.0.0.0 --port 8080 \
  --allowed-hosts biomcp.example.org

Loopback servers accept only local Host values by default. A non-loopback bind requires --allowed-hosts. The explicit --unsafe-allow-any-host escape hatch disables only this Host check; it does not add authentication, TLS, or encryption. Put remote deployments behind a trusted authenticated TLS proxy or inside a private network.

Remote clients should connect to http://<host>:8080/mcp. Lightweight process probes are available at GET /health, GET /readyz, and GET /.

Skills

BioMCP ships an embedded agent guide and a worked-example catalog. Use biomcp skill list when you need the right worked example, then use biomcp skill to read the embedded BioMCP guide or install it into your agent directory when you want local copies of the workflow references:

biomcp skill list
biomcp skill
biomcp skill install ~/.claude --force

See Skills for supported install targets, installed files, and legacy compatibility notes.

Local study analytics

study is BioMCP's local analysis family for downloaded cBioPortal-style datasets. The public entity surface handles API-backed, local-runtime, and hybrid discovery/detail; study commands work on local datasets when you need per-study query, cohort, survival, comparison, or co-occurrence workflows. Per-gene queries include mutations, CNA, expression, and structural variants/fusions from local data_sv.txt files. Mutation summaries stay mutation-only and note when fusions/SV need --type sv.

Use study download to fetch a dataset into your local study root. Set BIOMCP_STUDY_DIR when you want an explicit dataset location for reproducible scripts and demos; if it is unset, BioMCP falls back to its default study root.

export BIOMCP_STUDY_DIR="$HOME/.local/share/biomcp/studies"
biomcp study download msk_impact_2017
biomcp study query --study msk_impact_2017 --gene TP53 --type mutations --chart bar --theme dark --palette wong -o docs/blog/images/tp53-mutation-bar.svg
biomcp study query --study msk_impact_2017 --gene RET --type sv

See the CLI reference for the full study command family and dataset prerequisites.

Ops

biomcp version                            # show version and build info
biomcp health                             # inspect API connectivity plus local DDInter/EMA/cache readiness
biomcp update                             # self-update with release SHA256 checksum verification
biomcp update --check                     # check for updates without installing
biomcp uninstall                          # remove biomcp from ~/.local/bin

Support

• GitHub issues: https://github.com/genomoncology/biomcp/issues
• Troubleshooting: docs/troubleshooting.md
• Full documentation: https://biomcp.org/

Documentation

• Getting Started
• Search All Workflow
• BioASQ Benchmark
• Cross-Entity Pivot Guide
• Privacy Policy
• Source Licensing and Terms
• Data Sources
• Quick Reference
• Troubleshooting

Citation

If you use BioMCP in research, cite it via CITATION.cff. GitHub also exposes Cite this repository in the repository sidebar when that file is present.

Data Sources and Licensing

BioMCP is MIT-licensed. It performs on-demand queries against upstream providers instead of vendoring or mirroring their datasets, but upstream terms govern reuse of retrieved results.

Some providers are fully open, some BioMCP features require registration or API keys, and some queryable sources still impose notable reuse limits. The two biggest cautions are KEGG, which distinguishes academic and non-academic use, and COSMIC, which BioMCP keeps indirect-only because its licensing model is incompatible with a direct open integration.

Use Source Licensing and Terms for the per-source breakdown and API Keys for setup steps and registration links.

License

MIT