PubMed Custom Connector Server

A Model Context Protocol (MCP) server that brings the public PubMed biomedical literature database into Microsoft 365 Copilot as a custom federated connector.

Users can ask Copilot natural-language questions ("find recent trials on CRISPR for sickle cell disease") and Copilot calls this server in real time to search and read PubMed articles.

What's in the box

Path	Purpose
src/server.py	MCP server exposing the read-only search and fetch tools
src/pubmed.py	Async client for the NCBI E-utilities API (with rate-limit handling)
src/auth.py	Microsoft Entra ID bearer-token validation (ASGI middleware)
src/config.py	Environment-driven configuration
src/main.py	ASGI entrypoint (Uvicorn)
Dockerfile	Container image
azure.yaml, infra/	azd + Bicep for Azure Container Apps
scripts/smoke_test.py	End-to-end MCP client test
docs/tenant-setup-guide.md	Step-by-step tenant configuration + validation guide

Architecture

Microsoft 365 Copilot
   │  (1) user prompt
   ▼
Copilot orchestrator ──(2) bearer token from Microsoft enterprise token store──┐
   │                                                                            │
   │  (3) MCP search / fetch over HTTPS                                         │
   ▼                                                                            │
Azure Container Apps  ──(4) validate Entra JWT (issuer, audience, client app)──┘
   │
   │  (5) NCBI E-utilities (esearch / esummary / efetch)
   ▼
PubMed (public)

MCP tools

• search(query, max_results) → { "results": [ { id, title, url, snippet } ] } where id is the PubMed PMID.
• fetch(id) → { id, title, text, url, metadata } where text is the article abstract and metadata includes journal, authors, publication date, and DOI.

Both tools are read-only.

Quick start (local)

python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -r requirements.txt

# Run without auth for local testing
$env:AUTH_REQUIRED = "false"
$env:NCBI_EMAIL    = "[email protected]"   # recommended by NCBI
python -m src.main

# In another terminal: end-to-end test
python scripts\smoke_test.py

Then point MCP Inspector at http://localhost:8000/mcp (transport: Streamable HTTP) to explore the tools interactively:

npx @modelcontextprotocol/inspector

Configuration

All configuration is via environment variables (see .env.example):

Variable	Required	Description
AUTH_REQUIRED	no (default true)	Enforce Entra bearer-token validation.
ENTRA_TENANT_ID	when auth on	Your Microsoft Entra tenant ID.
ENTRA_AUDIENCE	when auth on	Accepted token audience = the Application ID URI from the Teams Developer Portal SSO registration. Comma-separated list allowed.
ENTRA_ALLOWED_CLIENT_IDS	no	Allowed calling client app IDs. Defaults to the Microsoft enterprise token store ab3be6b7-f5df-413d-ac2d-abf1e3fd9c0b.
NCBI_API_KEY	no	NCBI API key for higher PubMed rate limits (10/sec vs 3/sec).
NCBI_EMAIL	no	Contact email reported to NCBI (recommended).
MCP_PATH	no (default /mcp)	Path of the MCP streamable-HTTP endpoint.
ALLOWED_HOSTS	no (default *)	Hosts allowed by the MCP transport-security (DNS-rebinding) check. * disables the check (recommended behind Azure ingress, where Entra auth still applies). Set explicit hostnames to enable strict checking.
ALLOWED_ORIGINS	no (default *)	Origins allowed by the DNS-rebinding check.
PORT	no (default 8000)	Listen port.

Deploy to Azure

azd auth login
azd env new pubmed-connector
azd env set AUTH_REQUIRED false   # first deploy: get the URL, then lock down
azd up

Full deployment, Entra SSO setup, M365 admin center registration, and validation steps are in docs/tenant-setup-guide.md.

Security notes

• The connector exposes only public PubMed data, but the transport is still authenticated: every /mcp request must carry a valid Entra token whose audience matches your Application ID URI and whose calling app is the Microsoft enterprise token store.
• /health is intentionally unauthenticated for container liveness probes.
• The server is read-only and performs no writes to any system.