Open Targets MCP Server

A Model Context Protocol (MCP) server that exposes the Open Targets Platform GraphQL API as a set of tools for use with Claude Desktop and other MCP-compatible clients.

Quick Install

Option 1: Run once with uvx (no install)

uvx --from git+https://github.com/nickzren/opentargets-mcp opentargets-mcp

Option 2: Claude Desktop (MCPM)

# Install mcpm package manager
pip install mcpm

# Install the server
mcpm install opentargets

Option 3: Local install (dev or self-host)

git clone https://github.com/nickzren/opentargets-mcp
cd opentargets-mcp
pip install uv
uv sync

# Run (stdio transport by default)
uv run python -m opentargets_mcp.server

Option 4: Docker

git clone https://github.com/nickzren/opentargets-mcp
cd opentargets-mcp

# Build and run with Docker Compose
docker-compose up -d --build

Note: the default transport is http for docker deployments.

See the configuration section below for details and how to set ports and other environment variables.

Claude Desktop Manual Import (optional)

mcpm import stdio opentargets \
  --command "$(uv run which python)" \
  --args "-m opentargets_mcp.server --transport stdio"

Then restart Claude Desktop to start using the Open Targets tools.

Features

Core Capabilities

• Target Analysis: Search genes/proteins by Ensembl ID or symbol. Access expression data, genetic constraints, pathways, protein interactions, safety profiles, and mouse phenotypes
• Disease Analysis: Query diseases by ontology ID (mostly MONDO). Find associated targets, phenotypes (HPO), and research projects
• Drug Discovery: Search drugs by ChEMBL ID. Access safety data, adverse events, indications, and mechanism of action
• Evidence Mining: Explore target-disease associations with scored evidence from multiple sources
• Variant Analysis: Query genetic variants, GWAS credible sets, and pharmacogenomics data
• Study Exploration: Access GWAS studies with L2G predictions and fine-mapped loci
• Smart Search: Entity resolution with synonym handling, autocomplete, and ID mapping
• Cross-Entity Workflows: Multi-hop tools that chain disease, target, and drug evidence for prioritization
• Raw GraphQL Power Tools: Run single and batch raw GraphQL operations with structured status envelopes

Why This Server

This implementation is designed for practical Open Targets workflows:

• Curated breadth: 65 curated tools plus 3 advanced GraphQL tools (68 total), spanning target, disease, drug, evidence, variant, study, metadata, and cross-entity workflow tasks.
• Safer automation: strict ID resolution, typed parameter handling, and resilient retry behavior.
• Lower token overhead: optional fields filters on core domain tools to return only what you need.
• Flexible power mode: raw GraphQL tools are available for edge cases.

Data Sources

The Open Targets Platform integrates evidence from 22+ primary data sources:

• Genetics: Open Targets Genetics, ClinVar, UK Biobank, FinnGen, Gene2Phenotype, Orphanet, COSMIC
• Functional: CRISPR screens, DepMap, GeneBass
• Drugs: ChEMBL, FDA/EMA approvals, chemical probes
• Expression: GTEx, Human Protein Atlas, Expression Atlas
• Pathways: Reactome, Signor, IntAct
• Literature: Europe PMC text mining
• Safety: FAERS, pharmacogenomics data
• Models: Mouse (MGI, IMPC) phenotypes

Architecture

graph LR
    subgraph "Clients"
        A[Claude Desktop]
        B[Python Scripts]
        C[AI Agents]
    end
    
    subgraph "MCP Server"
        D[Open Targets<br/>MCP Server]
        E[Tool Categories<br/>Target • Disease • Drug<br/>Evidence • Search • Variant • Study]
    end
    
    subgraph "Open Targets"
        F[GraphQL API]
        G[22+ Data Sources]
    end
    
    A <-->|MCP Protocol| D
    B <-->|Direct API| D
    C <-->|Function Calls| D
    D <-->|GraphQL| F
    F <--> G
    
    E --> D
    
    style D fill:#e1f5fe
    style F fill:#fff3e0

The MCP server acts as a bridge between client applications and the Open Targets Platform. It translates tool calls into GraphQL queries and provides structured access to biomedical data from 22+ integrated sources.

Prerequisites

• Python 3.10+ with pip

Usage

Running the Server Standalone

# Using the convenience script (installs uv if missing, then syncs dependencies)
./run.sh

# Or run directly with uv (stdio transport by default)
uv run python -m opentargets_mcp.server

# Installed entrypoints
opentargets-mcp --help

# Specify transport explicitly
uv run python -m opentargets_mcp.server --transport [stdio|sse|http]

Configuration

• Environment variables: Transport/bind use MCP_TRANSPORT, FASTMCP_SERVER_HOST, and FASTMCP_SERVER_PORT (defaults: stdio, 0.0.0.0, 8000). API endpoint uses OPEN_TARGETS_API_URL (default: https://api.platform.opentargets.org/api/v4/graphql). For local-only development, prefer FASTMCP_SERVER_HOST=127.0.0.1.
• Validated settings: environment configuration is parsed with a typed settings model at startup (src/opentargets_mcp/settings.py), so invalid values fail fast.
• Name resolution: strict; unresolved names raise a clear error (use search_entities to find canonical IDs).
• Tool selection guidance: the server sends a short policy to clients to prefer curated tools, use fields to trim output, and reserve raw GraphQL for edge cases.
• Pagination guardrails: tool wrappers enforce page_index >= 0, page_size >= 1, and a global page_size <= 500.
• Command line: opentargets-mcp --transport [stdio|sse|http] --host 0.0.0.0 --port 8000 --api <url> provides flexible transport and endpoint selection.
• Verbose logging: add --verbose to elevate the global log level to DEBUG when troubleshooting.
• CLI helpers: --list-tools prints all registered tools, and --version prints the package version.
• Rate limiting: OPEN_TARGETS_RATE_LIMIT_RPS and OPEN_TARGETS_RATE_LIMIT_BURST can enable global server-side rate limiting. --rate-limiting and OPEN_TARGETS_RATE_LIMIT_ENABLED=true are also supported.

Transport Modes

The server supports multiple transport protocols powered by FastMCP:

stdio transport (default)

# For Claude Desktop (via mcpm) and local CLI tools
opentargets-mcp --transport stdio

SSE transport

# For web-based MCP clients with Server-Sent Events
opentargets-mcp --transport sse --host 0.0.0.0 --port 8000

HTTP transport

# For streamable HTTP MCP clients
opentargets-mcp --transport http --host 0.0.0.0 --port 8000

Using with MCP Clients

• Claude Desktop: Use mcpm installation (stdio) or direct server connection (sse)
• Web MCP clients: Use SSE or HTTP transports with public URL (tunnel required)
• Custom integrations: Any transport mode depending on your client implementation

Example Scripts

uv run python examples/target_validation_profile.py EGFR
uv run python examples/disease_to_drug.py "schizophrenia"
uv run python examples/drug_safety_profile.py "osimertinib"
uv run python examples/genetic_target_prioritization.py "inflammatory bowel disease"

AI Agent Example

The ReAct Agent provides an interactive terminal interface for exploring Open Targets data:

# Copy the example .env file and add your OpenAI API key
cp .env.example .env
# Then edit .env and set your OPENAI_API_KEY

# Run agent
uv run python examples/react_agent.py

The agent uses a ReAct (Reasoning and Acting) pattern to break down complex biomedical queries into steps, making it easy to explore drug targets, diseases, and their relationships.

Available Tools

The server wraps 68 operations from the Open Targets Platform: 65 curated tools plus 3 advanced GraphQL tools. Every tool returns structured JSON that mirrors the Open Targets GraphQL schema, and you can inspect the full machine-readable list with the MCP list_tools request.

Most domain tools accept either a canonical identifier (e.g., ENSG..., MONDO_..., CHEMBL...) or a human-readable name/symbol. Disease identifiers are largely MONDO since the 26.06 alignment to EFO 3.88; EFO_... IDs that were replaced no longer resolve. Colon notation (MONDO:0004979) is accepted and normalised. When a name is provided, the server automatically resolves it to the best matching Open Targets ID. Many core tools accept an optional fields list (dot-paths) to filter the response payload. search_entities also returns search.triples for compact {id, entity, name} consumption. For edge cases, prefer curated tools + fields first; use raw GraphQL only when no curated tool fits.

Quick-start shortcuts

• get_target_info – Core target identity record (Ensembl IDs, synonyms, genomic coordinates)
• get_disease_info – Disease/EFO summary with therapeutic area context
• get_drug_info – ChEMBL-backed drug profile and mechanism data
• search_entities – Unified entity search with synonym handling
• get_target_associated_diseases – High-confidence target-disease links with scores
• get_disease_associated_targets – Prioritised target list for an EFO disease
• get_target_known_drugs – Approved and investigational agents for a target
• get_target_disease_evidence – Evidence details across genetics, expression, and literature
• get_drug_repurposing_candidates – Multi-hop disease -> target -> drug candidate prioritization
• graphql_batch_query – Run one GraphQL query across many variable sets

Full catalog by category

• Target identity & biology (20 tools) — get_target_info, get_target_class, get_target_alternative_genes, get_target_associated_diseases, get_target_known_drugs, get_target_literature_occurrences, get_target_expression, get_target_pathways_and_go_terms, get_target_homologues, get_target_subcellular_locations, get_target_genetic_constraint, get_target_mouse_phenotypes, get_target_hallmarks, get_target_depmap_essentiality, get_target_interactions, get_target_safety_information, get_target_tractability, get_target_chemical_probes, get_target_tep, get_target_prioritization.
• Disease analytics (8 tools) — get_disease_info, get_disease_associated_targets, get_disease_phenotypes, get_disease_otar_projects, get_disease_known_drugs, get_disease_ontology, get_disease_literature_occurrences, get_disease_similar_entities.
• Drug profiling (10 tools) — get_drug_info, get_drug_cross_references, get_drug_linked_diseases, get_drug_linked_targets, get_drug_adverse_events, get_drug_pharmacovigilance, get_drug_warnings, get_drug_pharmacogenomics, get_drug_literature_occurrences, get_drug_similar_entities.
• Evidence synthesis (2 tools) — get_target_disease_evidence, get_target_disease_biomarkers.
• Search & discovery (4 tools) — search_entities, search_suggestions, get_similar_targets, search_facets.
• Metadata & ontology utilities (5 tools) — get_api_metadata, get_association_datasources, get_gene_ontology_terms, get_interaction_resources, map_ids.
• Workflow tools (1 tool) — get_drug_repurposing_candidates.
• Batch lookups (3 tools) — get_targets_batch, get_diseases_batch, get_drugs_batch.
• Variant interpretation (6 tools) — get_variant_info, get_variant_credible_sets, get_variant_pharmacogenomics, get_variant_evidences, get_variant_intervals, get_variant_protein_coordinates.
• Study exploration (6 tools) — get_study_info, get_studies_by_disease, get_study_credible_sets, get_credible_set_by_id, get_credible_set_colocalisation, get_credible_sets.
• Advanced GraphQL (3 tools) — graphql_schema, graphql_query, graphql_batch_query.

Response changes for Open Targets 26.06

Release 26.06 changed several upstream shapes. Tools were updated to match, so responses differ from earlier versions of this server:

• get_target_expression now returns target.baselineExpression, a paginated {count, rows} connection, instead of target.expressions. Rows are per-datasource quantitative distributions (median, min, max, q1, q3, specificity_score, distribution_score) keyed on tissueBiosample / celltypeBiosample, and no longer carry tissue / rna / protein. The tool takes page_index (default 0) and page_size (default 25); a single target can have well over a thousand rows. fields keeps its position as the first optional argument.
• Drug synonyms and tradeNames are objects rather than plain strings. Each entry is {label, source}, affecting get_drug_info, get_drug_cross_references and get_drugs_batch.
• Disease identifiers largely moved from EFO to MONDO with the alignment to EFO 3.88. Replaced EFO_... IDs no longer resolve; efo_id parameters keep their name but accept any supported disease ontology ID, in either colon or underscore notation.
• get_target_known_drugs honours page_index, and knownDrugs.count is the upstream total rather than the page length. Rows no longer carry mechanismOfAction; use get_drug_info for a drug's mechanisms.
• get_disease_known_drugs rejects cursor and free_text_query: the API exposes no paging or filtering on that field.
• get_target_subcellular_locations returns targetModifier, naming the protein form a location applies to.
• When the API rejects a query, the tool error now carries the upstream GraphQL message instead of a generic failure, and a 200 response containing GraphQL errors raises rather than returning partial data. The raw graphql_query tool still reports partial data with a warning status.

Each grouping matches the data domains described in the Open Targets docs (targets, diseases, drugs, evidence, variants, and studies). For high-volume workloads, respect the platform's throttling guidance from the Open Targets API FAQ and cache downstream where possible.

Development

# Run lint checks (same as CI/release)
uv run ruff check src tests

# Run tests
uv run pytest tests/ -v

# Inspect registered tools from CLI
uv run opentargets-mcp --list-tools