chrome://flags/#enable-webmcp-testing enabled. The demo below works either way.A clinic your agent
can actually use.
Patiently exposes a working outpatient clinic — its queue, its charts, its prescriptions — as 20 WebMCP tools. A doctor runs their floor by talking. A patient does intake in their own language. Each through the agent they already have, in the tab they are already signed into.
And every action that touches a patient’s care stops and waits for a human to click.
Your agent sees the whole floor
Who is waiting, how long, who has a red flag, what the intake agents already wrote in their chart, and what their last visit was for. Five read-only tools, answered from the session the clinician is already signed into.
It can write, but nothing is signed
It drafts a SOAP note from the intake, proposes prescriptions with a rationale for each drug, and screens them against home medications and last visit’s script. Everything it produces is explicitly marked unsigned.
Then it stops and waits for you
Recording vitals, signing a prescription, calling a patient in, closing a visit — each one opens a dialog and blocks. The agent’s own execute() is parked on your click, so the write does not exist unless you approve it.
Built on WebMCP · OpenAI · FastAPI · Next.js 14 · PostgreSQL
Try it from either side
The same live queue, seen from both ends of the visit. No sign-in — open one in each tab and watch them move together.
Run the clinic floor
The live queue across five departments, pre-visit charts written before the patient walks in, and drafting that stops at your signature.
The waiting room is empty
Every patient has been seen. Open the clinician dashboard to look at the floor, or reseed the demo data to start a fresh clinic day.