biomcp

Purpose

BioMCP unifies multiple biomedical data sources into a single command-line binary and one set of command syntax, helping researchers, clinicians, and AI agents quickly retrieve and integrate evidence. It is also an MCP (Model Context Protocol) server, so the same set of tools can be invoked by AI agents such as Claude Code, Codex, and Claude Desktop.

Main capabilities

• Unified access to 70 trusted sources: directly query PubMed, ClinVar, ClinicalTrials.gov, OncoKB, Reactome, and others; ten additional sources are presented indirectly through answers from other sources.
• Literature search:search article searches across PubTator3 and Europe PMC, deduplicates PMIDs/PMCIDs/DOIs, and can include Semantic Scholar data when conditions support it. Supports specifying sources, including all, pubtator, europepmc, pubmed, semanticscholar, litsense2.
• Cross-entity conversion: move directly from genes, variants, drugs, diseases, pathways, proteins, or articles into the next built-in view, with no need to manually rebuild filters.
• Local research analysis:study the command family provides local query, cohort, survival analysis, comparison, co-occurrence, and other workflows for downloaded cBioPortal-style datasets, with support for terminal, SVG, and PNG charts.
• Paper tracking:article citations、article references、article recommendations、article entities can expand a known paper into a broader evidence map.
• Enrichment and batch operations:biomcp enrich provides top-level enrichment based on g:Profiler;biomcp batch can execute up to 10 focused get calls in a single command.
• Progressive output: each get command supports selecting a section to focus the output; in JSON mode the response includes _meta.next_commands and _meta.section_sources metadata.

Usage

Installation

For PyPI installation, use biomcp-cli, not biomcp(the latter is unrelated to this project). After installation, the binary is located at ~/.local/bin; if that directory is not in PATH, the installer prints a command to add it and does not modify shell startup files. Homebrew installation depends on the existence of the separate genomoncology/homebrew-biomcp tap repository. A GHCR image can also be used for quick CLI checks or for stdio MCP clients that do not require a local installation.

As an MCP server

• Claude Code plugin: first install the biomcp binary, then add the managed plugin marketplace and install the BioMCP plugin. The plugin connects to the local stdio MCP server via biomcp serve .
• Codex MCP server: first install the binary, then register the same stdio MCP server with Codex.
• Claude Desktop extension: can be installed from the Anthropic Directory (if the environment supports it); for local manual configuration, a JSON MCP configuration can be used.
• Remote HTTP server: for shared or remote deployments, remote clients connect to http://127.0.0.1:8080/mcp, and the probe route is GET /health、GET /readyz、GET /.

Command syntax

Commands are organized by entity and source, and the table distinguishes entities that can be get from entities that can only be searched. Cross-entity helpers assist in converting between related entities. Regarding credentials: most commands require no credentials, and optional keys can improve rate limits or unlock optional enrichment. For example, without S2_API_KEY, search article and others remain available, but use the shared unauthenticated pool, rate-limited to 1 request/2 seconds; with a key, a dedicated rate limit of 1 request/second is used.

Use cases

• Researchers who need to quickly search for evidence across multiple biomedical databases.
• Clinicians and developers who want to let AI agents invoke biomedical tools via natural language.
• Research teams conducting local cBioPortal-style dataset analysis (cohorts, survival, comparison, co-occurrence).
• Developers building biomedical AI workflows who need MCP integration.

Limitations

• BioMCP itself is MIT-licensed, but the terms of use of upstream data sources constrain data reuse, and some sources restrict commercial or clinical use. Check each source's terms and obtain any required permissions before use. KEGG distinguishes between academic and non-academic use; COSMIC is provided only indirectly due to an incompatible licensing model.
• Some features require registration or an API key, for example OncoKB requires ONCOKB_TOKEN and DisGeNET requires DISGENET_API_KEY.
• References and recommendations may return empty for paywalled papers due to omissions by publishers covered upstream by Semantic Scholar.
• Rate limiting is process-local; multi-concurrent-worker scenarios require running a shared Streamable HTTP biomcp serve-http endpoint.
• Non-loopback binding requires --allowed-hosts；--unsafe-allow-any-host only disables the Host check; it does not add authentication, TLS, or encryption. Remote deployments should be placed behind a trusted authenticated TLS proxy or within a private network.
• BioMCP does not add telemetry, analytics, or remote log uploads; see the privacy statement at https://biomcp.org/policies/。

Other information

• Repository address:https://github.com/genomoncology/biomcp
• Issue feedback:https://github.com/genomoncology/biomcp/issues
• Full documentation:https://biomcp.org/
• The remote HTTP server has a newcomer guide; repository-local verification runs make lint、make test、make spec, with no supported make check command;make verify is only used for optional live public upstream confidence testing.
• The project won St. Jude's KIDS BioHackathon in 2025 with BioMCP.
• MCP Registry ownership marker:mcp-name: io.github.genomoncology/biomcp.
• If you use BioMCP in your research, you can cite it via CITATION.cff.

The above information is compiled from public sources; for specific feature details, installation steps, and configuration, please refer to the official documentation.