Corti Medical Assistant

A clinical chat assistant built on the Corti Agentic Framework: type a clinical question, the agent routes it to the right specialist expert, and streams back an evidence-backed answer with inline activity indicators.

Wired-up experts

Expert	What it does
pubmed-expert	Biomedical literature search and abstracts
coding-expert	ICD-10 / CPT diagnosis and procedure codes
clinical-trials-expert	Active and past clinical trials
medical-calculator-expert	Clinical calculators and scores
web-search-expert	Current information from the web

Setup

1. Get credentials from the Corti Console (Developer Quickstart → API credentials).

2. Fill in .env (already scaffolded, gitignored):

   CORTI_TENANT_NAME=...
   CORTI_CLIENT_ID=...
   CORTI_CLIENT_SECRET=...
   CORTI_ENVIRONMENT=eu

3. Install and run:

   npm install
   npm run dev

Open http://localhost:3000 (desktop, ≥1280px). Use the "Try sample" buttons to exercise each expert. Multi-expert answers can take 30–90s — watch the inline "calling …-expert" status indicators and the elapsed counter; Cancel aborts an in-flight request.

Architecture

• lib/agent.ts — server-only CortiClient + singleton ephemeral agent (created once on boot) and an A2AClient subclass that injects Corti auth headers into every outbound request, including the agent-card fetch.
• app/api/chat/route.ts — POST endpoint: buildParams(messages) → sendMessageStream → toUIMessageStream, wrapped in a 180s idle-timeout so a hung expert can't hold the request open forever.
• app/page.tsx — chat UI on useChat from @ai-sdk/react: streamed markdown answers, orchestrator status cues, per-turn credit usage (4 decimals), and context-preserving multi-turn memory (the adapter propagates contextId; "New chat" starts a fresh context).

Dependency versions are pinned exactly — ai@6.0.0, @corti/ai-sdk-adapter@0.1.1, and @a2a-js/sdk@0.3.0 are version-sensitive; don't upgrade without testing.