After-visit clarity from live FHIR context.
Featherless is a Prompt Opinion Marketplace healthcare agent that turns a synthetic FHIR visit into a plain-language patient packet, draft care-team closure resources, and an auditable MCP/A2A trace. It is built as one reusable MCP server plus one External A2A orchestrator on Cloudflare Workers.
| Surface | Status |
|---|---|
| Prompt Opinion BYO Agent | Published and used as the chat entry point |
| External A2A Agent | Published: https://app.promptopinion.ai/marketplace/agent/019e12e2-d251-7de3-bf49-16376ff51e73 |
| MCP Server | Published: https://app.promptopinion.ai/marketplace/mcp/019e12cd-76bd-708f-b230-e48da20ad8bc |
| Primary Marketplace listing | https://app.promptopinion.ai/marketplace/agent/019e13c6-7704-75aa-be80-8ef528a3cb7f |
| Live MCP endpoint | https://featherless-mcp.inf3ctious007.workers.dev/mcp |
| Live A2A AgentCard | https://featherless-orchestrator.inf3ctious007.workers.dev/.well-known/agent-card.json |
| Detail | Value |
|---|---|
| Authors | Chadwick Jones, Alessandro Frau |
| Hackathon | Agents Assemble: The Healthcare AI Endgame Challenge |
| Data posture | Synthetic FHIR only, no PHI, no token storage |
Open the primary Prompt Opinion Marketplace listing, start the Featherless Visit Assistant, enable FHIR Context, choose a synthetic patient, and turn Show Tool Calls on.
Primary listing:
https://app.promptopinion.ai/marketplace/agent/019e13c6-7704-75aa-be80-8ef528a3cb7f
Try these judge-style prompts:
Create an after-visit summary for this patient in plain English. Include what changed today, medication instructions, warning signs, follow-up steps, and where the information came from.
Can you make the same after-visit summary for this patient in Spanish? Keep it simple, and include medicine instructions, warning signs, follow-up steps, and sources.
Can you show me a visual dashboard for this patient with charts and safety highlights? If the dashboard opens as a link, put that link first, then briefly explain what I should look at.
Expected result: the BYO agent consults the external featherless A2A agent, the orchestrator calls the live MCP Worker, and the response shows readable patient guidance, citations or grounding details, and an auditable tool trace. For the dashboard prompt, Featherless returns the MCP-UI ui:// resource plus a short-lived hosted /render/<token> URL that can be opened directly from the response.
For the current patient in Prompt Opinion, Featherless:
- reads the FHIR visit context,
- packs the relevant encounter, problems, medications, orders, vitals, and labs,
- generates a grade-6 patient packet through Workers AI,
- validates readability and grounding,
- prepares draft FHIR R4 closure resources for care-team review,
- returns traceable MCP/A2A evidence.
The patient-facing packet is intended to answer a simple question: "What do I do next, and when should I ask for help?"
Prompt Opinion BYO Agent
| consults external Featherless A2A agent with FHIR context
v
Featherless Orchestrator Worker
| forwards request-scoped FHIR metadata as SHARP-style headers
v
Featherless MCP Worker
| clinical_pack_visit_context
| clinical_generate_patient_packet
| clinical_prepare_care_team_closure
v
Prompt Opinion chat response + tool trace
The orchestrator uses a Cloudflare service binding named FEATHERLESS_MCP to call the deployed MCP Worker. FHIR access tokens are passed per request and are not stored, logged, or returned.
Patients often leave visits with instructions they cannot fully remember or act on. The project is grounded in plain-language and patient-education guidance: patients can forget 40-80% of clinical information from encounters [CIT-003], patient education should be understandable and actionable [CIT-002], and clear communication should prioritize concrete actions over jargon [CIT-001, CIT-005].
Featherless focuses on the handoff moment after a clinician has made a plan, but before the patient and care team have something clear enough to use.
| Tool | Purpose |
|---|---|
clinical_pack_visit_context |
Reads patient, encounter, problems, medications, orders, vitals, labs, and visit summary into a typed payload. |
clinical_generate_patient_packet |
Uses Workers AI to generate patient-facing packet markdown; reports Flesch-Kincaid and INFLESZ readability; validates grounding. |
clinical_prepare_care_team_closure |
Emits draft FHIR Task, CommunicationRequest, and DocumentReference resources for care-team review. |
Tool source lives under src/tools. The A2A orchestrator lives in orchestrator/src/index.ts.
| Concern | Featherless posture |
|---|---|
| PHI | Synthetic/de-identified data only. No real PHI is used in the demo. |
| FHIR token handling | Tokens are request-scoped and never persisted. |
| Patient messaging | Nothing is auto-sent. The send artifact is a draft CommunicationRequest. |
| Write-back | Disabled by default; requires caller write_back: true and WRITE_BACK=1. |
| LLM provider | Cloudflare Workers AI only; no external LLM API key. |
| Grounding | Citation-or-cut: unsupported clinical claims are omitted or fail validation. |
The MCP server also exposes visualization tools that return MCP-UI ui:// resources. The dashboard HTML loads Chart.js in the browser; the Worker emits HTML and does not execute chart code.
Prompt Opinion currently exposes the MCP-UI payload as tool output rather than natively rendering the iframe. Featherless also emits a short-lived hosted /render/<token> URL for that same dashboard, so the demo can open the live visualization directly from the Prompt Opinion response.
Featherless uses two synthetic data paths:
- Prompt Opinion synthetic patients for the main Marketplace/A2A/MCP proof.
- Public synthetic HAPI data for the optional MCP-UI showcase clip.
The original hero patient is Mrs. María García, a synthetic cardiology follow-up patient with a new PRN furosemide instruction, labs due soon, an echo to schedule, and care-team follow-up tasks. See HERO_PATIENT.md.
Install and verify:
npm install
npm run typecheck
npm testLoad the local synthetic hero bundle:
docker run -p 8080:8080 hapiproject/hapi:latest
npx tsx scripts/load-hero.tsRun the Workers locally:
npm run dev
npm run dev:orchestratorFor a fresh workspace, register:
MCP endpoint:
https://featherless-mcp.inf3ctious007.workers.dev/mcp
A2A AgentCard:
https://featherless-orchestrator.inf3ctious007.workers.dev/.well-known/agent-card.json
Then enable Prompt Opinion FHIR context and consult the external featherless agent from the BYO agent. Full steps are in docs/po-registration.md.
Recorded final engineering gate:
- Cloudflare MCP Worker deployed.
- Cloudflare A2A orchestrator deployed.
- Prompt Opinion MCP, External A2A Agent, and BYO Agent published.
- Latest recorded local test suite: 70/70 passing after the hosted render bridge.
- Prompt Opinion dashboard prompt returns a hosted MCP-UI render URL for the live visualization.
Submission helpers:
docs/publish-readiness.md- Marketplace and submission checklist.CITATIONS.md- closed-world evidence list.DECISIONS.md- key architecture decisions.
MIT. See LICENSE.