WEBMCP CHALLENGE · JUDGE GUIDE10 tools · live shared state · human governed

THE PROBLEM

Doctors want to use AI.
They are not sure how to do it responsibly.

HIPAA, vendor agreements, retention, minimum-necessary data, clinical judgment, and fast-moving AI features create a wall of uncertainty. Generic articles explain pieces. They do not help a practice choose one workflow and move safely from interest to action.

WHAT THIS SITE DOES

It turns responsible AI adoption into a shared workspace. A doctor explores the visible experience while ChatGPT uses narrow page tools to inspect context, focus lessons, assess governance signals, review a public practice website, and stage a 21-day mission.

WEBMCP, IN PLAIN ENGLISH

The page gives the agent a safe vocabulary.

01Doctor chooses the goal

The person stays in control of intent and final decisions.

02ChatGPT calls a named tool

No guessing which button or field might do the job.

03The shared page changes

Doctor and agent see the same lesson, assessment, or draft.

WebMCP is an experimental open standard for websites to expose structured tools directly to compatible browser agents. It complements the human interface; it does not replace it.

THE EXACT DEMO ENVIRONMENT

Two destinations.
One connected conversation.

The URL goes in the built-in browser’s address bar. The prompt goes in the message box of the same ChatGPT Work or Codex conversation. They are not interchangeable.

Latest ChatGPT desktop app

Use a Work or Codex conversation.

GPT-5.6 Sol or Terra

Luna currently has site tools disabled.

Site tools enabled

Settings → Browser → Permissions → Enable site tools.

Eligibility note

Site tools are not currently available in Enterprise or Edu workspaces and remain rollout-dependent.

  1. 1
    Open the built-in browser from the conversation.

    Use the browser button or press Ctrl+Shift+B on Windows.

  2. 2
    Put the site URL in the browser address bar.

    Keep the live lab page open.

  3. 3
    Confirm the page’s tools.

    Select Site tools in the address bar, then Available site tools. You should see ten.

  4. 4
    Return to that conversation’s message box.

    Send the demo prompt shown here.

  5. 5
    Watch the shared page update.

    ChatGPT focuses the lesson, selects the workflow, stages the assessment, and builds the mission. The human approval gate remains untouched.

SITE URL → BROWSER ADDRESS BARhttps://lab.hipaa.technologyOpens the live clinician workspace.
DEMO PROMPT → CHATGPT MESSAGE BOX

Inspect this HIPAA.technology workspace. Focus the Minimum Necessary lesson, select the Pre-visit Brief use case, assess it with no approved tool path yet, a named human reviewer, and retention reviewed, then stage a 21-day mission for 3 clinicians at 15 minutes per day. Do not approve anything for me.

Instructs ChatGPT to use the tools on the open page.
What does not work

Opening ChatGPT in one ordinary Chrome tab and the site in another does not connect them. Pasting the URL into an unrelated chat may let ChatGPT read a public page, but it does not create a WebMCP session with that page.

THE TEN TOOLS

A deliberately bounded agent surface.

Each tool maps to a visible, reversible action. The site does not expose clinical decisions, policy acceptance, mission approval, or patient-data entry as tools.

01

Inspect the workspace

READ
inspect_adoption_workspace

Reads the current lesson, workflow, readiness signal, and mission draft without changing anything.

02

Focus a learning lesson

UI
focus_learning_lesson

Moves the shared page to one HIPAA + AI boundary so the clinician and ChatGPT discuss the same thing.

03

Select a clinical use case

UI
select_clinical_ai_use_case

Chooses one bounded workflow in the visible interface without approving it for real-world use.

04

Assess readiness

ASSESS
assess_use_case_readiness

Stages a conservative educational signal from the tool path, human reviewer, and retention review.

05

Stage a 21-day mission

DRAFT
stage_21_day_adoption_mission

Creates a synthetic-first plan with a cohort, daily time budget, and measurable proof standard.

06

Revise the mission

DRAFT
revise_adoption_mission

Changes the draft and automatically resets human approval.

07

Prepare a website review

PUBLIC WEB
prepare_practice_website_review

Validates a public practice domain and opens a safe review rubric—no portals, private pages, or PHI.

08

Deliver the website debrief

UI
deliver_practice_website_debrief

Writes strengths, opportunities, and one safe next move into the shared page.

09

Guide the clinician

UI
set_contextual_guidance

Updates the floating help box with a brief message for the section the clinician is viewing.

10

Focus a mission week

UI
focus_mission_week

Moves the visible plan to week one, two, or three without making a consequential decision.

THE HUMAN AUTHORITY BOUNDARY

ChatGPT may inspect, focus, assess, stage, and revise. A person must accept policy guardrails, approve the mission, and own every clinical or compliance decision.

WHY THIS IS MORE THAN A DEMO

Research-backed.
Built and hosted with ChatGPT.

01

ChatGPT Deep Research

Middleware Encryption for HIPAA-Compliant Use of Frontier LLMs in U.S. Healthcare informed the workflow model: map the full information path, use minimum-necessary data, govern logs and agent tools, and preserve human accountability.

02

Created with Codex

Codex designed and implemented the React experience, the ten imperative WebMCP tools, narrow JSON schemas, safety annotations, shared UI state, and explicit human-only controls.

03

Hosted with ChatGPT Sites

The production application is built and deployed with ChatGPT Sites, then served through the custom HIPAA.technology domain.

04

Verified as a working system

Chrome’s native WebMCP API discovers all ten tools. Tool execution returns structured results and visibly updates the clinician interface.

MANUAL CHROME TEST

No ChatGPT agent required.

In Chrome 149+, enable #enable-webmcp-testing and #devtools-webmcp-support, relaunch, open the live lab, then use DevTools → Application → WebMCP to inspect schemas, execute tools with parameters, and review results.