MCP (AI agents)

neolife ships a first-class Model Context Protocol (MCP) server, so an AI agent can read and act on the fulfillment rail through the same guarded surface as everything else. The agent never gets a special path — the server-side rules apply identically.

What an agent can do#

• Look up orders, products, and intake status.
• Draft and submit orders (still subject to licensed-provider approval).
• Track fulfillment via the PHI-free event stream.

Safety by construction#

• Provider approval is non-negotiable. No MCP tool can submit an order that a licensed provider hasn't approved.
• Idempotent by default. The submitOrder tool derives a stable idempotency key from the order id, so an agent that retries can never double-ship.
• No PHI in keys or events. The PHI boundary holds for agents too.

Connect#

Add the server to your MCP client (for example, an agent runtime or an IDE):

{
  "mcpServers": {
    "neolife": {
      "command": "npx",
      "args": ["-y", "@neolifehealth/mcp"],
      "env": { "NEOLIFE_API_KEY": "nk_sandbox_…" }
    }
  }
}

Point it at a nk_sandbox_… key first — the agent then operates entirely on synthetic patients while you evaluate its behavior, and you switch to a live key when you're ready.

Design notes#

The MCP surface is intentionally PHI-conscious: agents act on ids and status, and any patient data is fetched over the authenticated API only when genuinely needed. See the PHI boundary for the full model.