EHRSQL
Clinical records
Query OMOP-format health records with validated, read-only SQL.
Preconfigured for OMOP EHRs on SQLite, MySQL, and SQL Server
A standard port for clinical and biological evidence
The private data port for medical AI.
One read-only port for OMOP records and biomedical knowledge graphs. Run MedCP locally when sensitive work must stay inside your approved environment.
01Read-only by design
SQL and Cypher mutations are rejected before execution.
02Local stdio server
Choose the host, model, database endpoint, and network path.
03Clinical + biological
Connect OMOP evidence with genes, pathways, drugs, and disease biology.
MCP standardizes how clients connect, negotiate features, and discover tools. MedCP applies it to clinical and biological data.
Overview
One interface. Two key components.
MedCP brings two key components into one local interface: an OMOP EHR queried with SQL and a biomedical knowledge graph queried with Cypher. Results return to the host agent that you choose.
Knowledge graphCypher
Biomedical context
Connect diseases with genes, pathways, drugs, and other biological relationships through read-only graph queries.
Preconfigured for the SPOKE knowledge graphSystem architecture
A local bridge, with explicit data boundaries.
Follow one research question from the person, through the agent host and MedCP, to the configured databases and back.
Approved research environment
The local stdio server advertises the configured tools and rejects write-capable SQL and Cypher before execution.
View the complete clinical and biological workflow
The architecture separates the researcher, host agent, local MedCP process, and the clinical and biological databases that may be local or remote.
Privacy-focused design
Control the full data path.
Pair MedCP with a local or institution-hosted model and an approved harness to build a private research environment.
MedCP adapter
A local stdio process exposes only the configured tools and rejects write-capable SQL and Cypher.
Approved host and model
BioRouter can orchestrate local or institution-approved models. Codex CLI and Claude Code integrations are suitable only when their complete data paths are approved for the dataset.
Institutional controls
Your institution must still manage identity, database permissions, small-cell rules, logging, retention, and exports.
Compliance boundary
MedCP supplies the read-only adapter. HIPAA compliance depends on the approved end-to-end environment and is not certified by MedCP alone.
Integrations
Choose the host you already use.
Every integration launches the same MedCP core. BioRouter is the canonical benchmark harness.
Verified host
BioRouter
The canonical MedCP benchmark harness, with local orchestration and an OS-keyring path for values entered as secrets.
- Needs
- uv and an approved database path
- Data boundary
- Version 1.88.6 was verified. Values entered through its secret-setting flow can use the configured OS keyring. The full model and data path still needs institutional approval.
Setup
biorouter extension install MedCP.brxt \
--env CLINICAL_RECORDS_BACKEND=sqlite \
--env CLINICAL_RECORDS_SQLITE_PATH=/absolute/path/database.sqliteMeasured evidence
Benchmark and replication results.
Compare cohort-grounded answers, explore clinical and biological associations, and review local estimates beside published results.
Paired benchmark exampleOne real question, with and without SPOKE
GPT-5.5 · Q6.10
In sickle cell disease, which infection diagnoses occur most frequently, and what immune or hemolysis processes might explain susceptibility?
MedCP EHR-only
Ad hoc filters missed the target
The EHR-only run used diagnosis-name filters on 1,226 patients. It ranked acute chest syndrome (386) and fever (367) first, although neither is an infection diagnosis.
- Name-filtered cohort and phenotypes
- 10 EHR calls in 766 seconds
- Mean expert score: 0 of 2
MedCP EHR + SPOKE
A hierarchy-grounded answer
The fully enabled run defined 1,783 patients using standard descendants and excluded sickle cell trait. The leading infections were acute upper respiratory infection (327), pneumonia (322), viral disease (255), and urinary tract infection (239).
- 8 SPOKE calls and 10 EHR calls
- Linked HBB, splenic dysfunction, hemolysis, and oxidative stress
- Mean expert score: 2 of 2
Observed paired GPT-5.5 benchmark example. Both panels used MedCP. This compares EHR-only with EHR + SPOKE, not an unaided model with MedCP.
Clinical question benchmark
MedCP + SPOKE compared with MedCP EHR-only
MedCP + SPOKE1.265
MedCP EHR-only1.045
Mean expert score difference, 0-2+0.220
Mean elapsed time, descriptive30.1% lower than EHR-only
Graph retrievalSPOKE used in 37 of 100 questions
Score comparison: paired t-test p=0.0026; signed-rank p=0.0040. Each condition contains the same 100 questions. Scores are the mean of physician and bioinformatician ratings on the 0 to 2 scale. Results are model and deployment specific.
Study walkthroughHow MedCP links EHR comorbidity to SPOKE biology
The same disease pairs are measured in the EHR and SPOKE, then compared in an adjusted analysis.
Step 01 of 05
Define the clinical question
A researcher asks MedCP to map comorbidity among common autoimmune and neurodegenerative diseases.
Biological dimensions measured in SPOKE
- Genes
- Variants
- Symptoms
- Anatomy
- Compounds
- Pathways
- Curated disease relationships
Literature replication evidenceCompare published and local estimates
- 01SpecifyPopulation, exposure, outcome
- 02MapOMOP concepts + SPOKE context
- 03ExecuteCohort and statistical model
- 04ComparePublished and local estimates
Literature replication
Cerono et al.
Publishedn=813
2.27 (1.37-3.75)
MedCPn=714
1.94 (1.24-3.05)
The MedCP estimate points in the same direction and its interval remains above the null.
Aggregate estimates are shown for comparison. Patient-level replication data are not included in the public repository.
Data boundary
Read-only is not the same as HIPAA compliant.
MedCP does not make an AI host or model provider HIPAA compliant. Do not use PHI with Codex CLI, Claude Code, Claude Desktop, BioRouter, or another host unless your institution has approved the full database, host, model, network, logging, credential, and governance path.
A local or on-premises deployment can keep the research loop inside an institutional boundary when the complete environment is approved and governed.
Read the privacy guidance