An MCP server that exposes strict, typed Refua tools for Boltz-2 folding, affinity prediction, and BoltzGen design workflows, making drug discovery accessible to any AI assistant.
Refua MCP is a Model Context Protocol server that bridges AI assistants and the Refua drug discovery toolkit. It exposes typed, schema-validated tools that any MCP-compatible client can call to perform protein folding, affinity prediction, antibody design, ADMET profiling, and clinical simulation.
The server follows the MCP spec revision 2025-11-25 and supports
stdio, SSE, and streamable-HTTP transports with configurable security controls.
refua_fold – Fold/design with entities and constraintsrefua_affinity – Affinity-only predictionsrefua_antibody_design – Antibody CDR designrefua_validate_spec – Pre-flight validationrefua_protein_properties – Sequence-based analysisrefua_admet_profile – ADMET predictionsrefua_clinical_simulator – Trial simulationrefua_job – Background job managementAdd Refua MCP to your Claude Code configuration and let Claude perform protein folding, design antibodies, and predict ADMET properties.
Register with codex mcp add and use Refua tools from the Codex CLI
for AI-powered drug discovery workflows.
Refua MCP follows the open MCP protocol spec, so any compatible client can connect via stdio, SSE, or HTTP transports.
The server exposes rich metadata resources and recipe templates to help AI assistants discover capabilities and construct valid tool calls.
Runtime feature flags, protocol revision, transport config, and security summary
available at refua://capabilities.
Canonical recipes for fold, affinity, antibody design, protein properties, and
clinical simulation at refua://recipes/{name}.
Browse available property groups and individual property metadata through
refua://protein-properties/ resources.
Refua MCP supports three transport modes: stdio (default), SSE, and streamable-HTTP. HTTP transports include DNS rebinding protection, allowed hosts/origins configuration, and static bearer token authentication.
All major tools enforce strict JSON schemas with discriminated entity unions, and exploratory names are blocked by default to prevent costly preflight runs.
pip install refua[cuda] refua-mcpOptional ADMET and clinical extras:
pip install "refua[admet]" "refua-mcp[clinical]"python -m refua_mcp.serverOr with HTTP transport:
REFUA_MCP_TRANSPORT=streamable-http python -m refua_mcp.serverRefua MCP lets any AI assistant access the full power of the Refua toolkit through a standards-based protocol with strict type safety and security controls.