Precision Medicine MCP Platform

Dedicated to PatientOne -- a dear friend who passed from High-Grade Serous Ovarian Carcinoma in 2025.

This platform automates multi-modal data processing for clinical decision support — all results require clinician review before any clinical action.

The Problem

A patient can spend months moving between referrals, imaging, repeat evaluations, and denials without reaching a decision. Much of that motion starts at the workup — the test that would have settled the question was never ordered, or the finding that would have opened a door was never generated.

Standard oncology panels (BRCA1/2, HRD, tumor genomic) surface no immunotherapy or investigational hypotheses. Standard lipid panels and population genetic screens miss the tests most likely to change preventive management. Integrating genomics, spatial transcriptomics, imaging, and clinical data by hand is clinically impractical — the platform automates it.

What This Changes for the Patient

This platform does not triage patients or route referrals. It makes the analysis behind a decision complete on the first pass, instead of spread across a sequence of visits.

Fewer tests is also a result. A CAC score of 0 would support deferring statin therapy for PAT003. The same analysis that adds a test can remove one — the goal is the decision, not the volume of workup.

Scope of the claim. These are three synthetic patients analyzed end-to-end, not clinical outcomes. Time and cost figures are modeled and pending clinical validation. The platform's contribution is to the quality and completeness of the decision, not to the referral pathway around it.

The Results

The platform surfaces clinically actionable findings that standard workup cannot reach — 6 investigational hypotheses across 2 cancer types plus 3 preventive health evidence gaps, validated across three independent use cases:

The same 19-server architecture runs all three, with no disease-specific code changes. All tools are accessible via natural language, every AI result requires clinician APPROVE/REVISE/REJECT, and 11 servers return per-tool XAI metadata (confidence levels, evidence grades, counterfactuals). Current counts: Server Registry.

How we validate our results →

Try It

# Clone and explore
git clone https://github.com/lynnlangit/precision-medicine-mcp.git
cd precision-medicine-mcp

# Run tests for any server (DRY_RUN mode, no external deps needed)
cd servers/mcp-multiomics && uv run pytest -v

# Or use Claude Code to explore interactively
claude

Most servers default to DRY_RUN mode (mock responses, no API keys needed) for quick validation. Set *_DRY_RUN=false to use synthetic patient data for end-to-end testing. mcp-deidentify is the exception — it runs live by default, because a de-identification tool that silently returns fabricated data is a safety failure rather than a safe default.

Learn More

Video: 5-minute demo | Paper: Why MCP for Healthcare | External connectors: Setup guide

                  +--------------------------------------+
                  |           CLIENT LAYER               |
                  |  Claude Desktop / Hospital EHR       |
                  |  Adapter / Research Notebook         |
                  +----------------+-----------------+
                                   |
                         MCP (FastMCP >= 2.13)
                                   |
   +---------------------------------------------------------------+
   |                                                               |
   |  PRE-PROCESSING (Stage 0)                                     |
   |  deidentify                                                   |
   |                                                               |
   |  DATA ACQUISITION      ANALYSIS & INFERENCE      REPORTING   |
   |                                                               |
   |  mockepic              spatialtools              patient-     |
   |  epic                  multiomics                report       |
   |  geodownload           perturbation                           |
   |  mocktcga              quantum-fidelity                       |
   |  genomic-results       opentargets                            |
   |  fgbio                 neoantigen                             |
   |                        cibersortx                             |
   |                        openimagedata                          |
   |                        deepcell                               |
   |                        cell-classify                          |
   |                        cardiometabolic                        |
   +---------------------------------------------------------------+

Full detail: Architecture · Server Registry

Canonical values for all three patients: Patient Outcomes (source of truth: tests/fixtures/pat00X_canonical.py)

Apache 2.0 | Python 3.11+ | FastMCP >= 2.13 | uv for package management