BioMCP is an open source (MIT License) toolkit that empowers AI assistants and agents with specialized biomedical knowledge. Built following the Model Context Protocol (MCP), it connects AI systems to authoritative biomedical data sources, enabling them to answer questions about clinical trials, scientific literature, and genomic variants with precision and depth.
BioMCP is certified by MCP Review. This certification ensures that BioMCP follows best practices for Model Context Protocol implementation and provides reliable biomedical data access.
While Large Language Models have broad general knowledge, they often lack specialized domain-specific information or access to up-to-date resources. BioMCP bridges this gap for biomedicine by:
- Providing structured access to clinical trials, biomedical literature, and genomic variants
- Enabling natural language queries to specialized databases without requiring knowledge of their specific syntax
- Supporting biomedical research workflows through a consistent interface
- Functioning as an MCP server for AI assistants and agents
BioMCP integrates with multiple biomedical data sources:
- PubTator3/PubMed - Peer-reviewed biomedical literature with entity annotations
- bioRxiv/medRxiv - Preprint servers for biology and health sciences
- Europe PMC - Open science platform including preprints
- ClinicalTrials.gov - Clinical trial registry and results database
- MyVariant.info - Consolidated genetic variant annotation
- TCGA/GDC - The Cancer Genome Atlas for cancer variant data
- 1000 Genomes - Population frequency data via Ensembl
BioMCP provides two essential tools for biomedical research:
IMPORTANT: Always start with domain="thinking" for systematic analysis of biomedical queries.
# Start analysis with sequential thinking
search(
domain="thinking",
thought="Breaking down the query about BRAF mutations in melanoma...",
thoughtNumber=1,
totalThoughts=3,
nextThoughtNeeded=True
)The sequential thinking domain helps:
- Break down complex biomedical problems systematically
- Plan multi-step research approaches
- Track reasoning progress
- Revise and branch thinking paths as needed
The search tool supports three modes:
Use domain="thinking" with thought parameters for systematic analysis before searching.
Use the query parameter with structured field syntax for powerful cross-domain searches:
# Simple natural language
search(query="BRAF melanoma")
# Field-specific search
search(query="gene:BRAF AND trials.condition:melanoma")
# Complex queries
search(query="gene:BRAF AND variants.significance:pathogenic AND articles.date:>2023")
# Get searchable fields schema
search(get_schema=True)
# Explain how a query is parsed
search(query="gene:BRAF", explain_query=True)Supported Fields:
- Cross-domain:
gene:,variant:,disease: - Trials:
trials.condition:,trials.phase:,trials.status:,trials.intervention: - Articles:
articles.author:,articles.journal:,articles.date: - Variants:
variants.significance:,variants.rsid:,variants.frequency:
Use the domain parameter with specific filters:
# Search articles
search(domain="article", genes=["BRAF"], diseases=["melanoma"])
# Search trials
search(domain="trial", conditions=["lung cancer"], phase="3")
# Search variants
search(domain="variant", gene="TP53", significance="pathogenic")Retrieve full details for a single article, trial, or variant:
# Fetch article details
fetch(domain="article", id="34567890")
# Fetch trial with all sections
fetch(domain="trial", id="NCT04280705", detail="all")
# Fetch variant details
fetch(domain="variant", id="rs113488022")Domain-specific options:
- Articles:
detail="full"retrieves full text if available - Trials:
detailcan be "protocol", "locations", "outcomes", "references", or "all" - Variants: Always returns full details
-
Install
uvif you don't have it (recommended):# MacOS brew install uv # Windows/Linux pip install uv
-
Configure Claude Desktop:
- Open Claude Desktop settings
- Navigate to Developer section
- Click "Edit Config" and add:
{ "mcpServers": { "biomcp": { "command": "uv", "args": ["run", "--with", "biomcp-python", "biomcp", "run"] } } }- Restart Claude Desktop and start chatting about biomedical topics!
# Using pip
pip install biomcp-python
# Using uv (recommended for faster installation)
uv pip install biomcp-python
# Run directly without installation
uv run --with biomcp-python biomcp trial search --condition "lung cancer"BioMCP provides a comprehensive CLI for direct database interaction:
# Get help
biomcp --help
# Run the MCP server
biomcp run
# Article search examples
biomcp article search --gene BRAF --disease Melanoma # Includes preprints by default
biomcp article search --gene BRAF --no-preprints # Exclude preprints
biomcp article get 21717063 --full
# Clinical trial examples
biomcp trial search --condition "Lung Cancer" --phase PHASE3
biomcp trial get NCT04280705 Protocol
# Variant examples with external annotations
biomcp variant search --gene TP53 --significance pathogenic
biomcp variant get rs113488022 # Includes TCGA and 1000 Genomes data by default
biomcp variant get rs113488022 --no-external # Core annotations onlyTest your BioMCP setup with the MCP Inspector:
npx @modelcontextprotocol/inspector uv run --with biomcp-python biomcp runThis opens a web interface where you can explore and test all available tools.
OncoMCP extends BioMCP with GenomOncology's enterprise-grade precision oncology platform (POP), providing:
- HIPAA-Compliant Deployment: Secure on-premise options
- Real-Time Trial Matching: Up-to-date status and arm-level matching
- Healthcare Integration: Seamless EHR and data warehouse connectivity
- Curated Knowledge Base: 15,000+ trials and FDA approvals
- Sophisticated Patient Matching: Using integrated clinical and molecular profiles
- Advanced NLP: Structured extraction from unstructured text
- Comprehensive Biomarker Processing: Mutation and rule processing
Learn more: GenomOncology
For comprehensive documentation, visit https://biomcp.org
Looking to see BioMCP in action?
Check out the companion repository: 👉 biomcp-examples
It contains real prompts, AI-generated research briefs, and evaluation runs across different models. Use it to explore capabilities, compare outputs, or benchmark your own setup.
Have a cool example of your own? We’d love for you to contribute! Just fork the repo and submit a PR with your experiment.
This project is licensed under the MIT License.