MCP Server

Runnable example: 27_integrations_mcp_langchain_dbt.ipynb — all three integrations end-to-end, offline, executed with outputs.

Drive Portiere from any Model Context Protocol client (Claude Desktop, MCP-aware agents). Every tool runs offline — the server cannot send data off-machine.

pip install "portiere-health[mcp,polars]"
portiere mcp        # stdio server

Tools exposed

Tool	What it does
portiere_list_standards	list target standards (OMOP, FHIR, HL7 v2, OpenEHR, custom)
portiere_profile_source	profile a local CSV/Parquet/JSON — row/col counts, per-column stats
portiere_suggest_schema_mapping	source columns → target table/field, confidence, status (offline)
portiere_map_concepts	local codes → standard concepts via the knowledge layer
portiere_egress_posture	safety — report LOCAL/REMOTE per component so an agent can self-verify no-egress before acting

Wire it into Claude Desktop

Add to your MCP config (claude_desktop_config.json):

{
  "mcpServers": {
    "portiere": {
      "command": "portiere",
      "args": ["mcp"]
    }
  }
}

Now an agent can profile a hospital extract, propose an OMOP mapping, and map codes — all locally, with portiere_egress_posture available so it can prove nothing left the machine.

Programmatic

from portiere.integrations.mcp import create_mcp_server
server = create_mcp_server()
server.run()

The tools are the shared, dependency-free core in portiere.integrations.tools (also used by the LangChain adapter) — so behaviour is identical across frameworks.