ASSERA

A denial isn’t the final word.

Human-centered healthcare access with WebMCP.

ASSERA is a human-centered healthcare access prototype that helps a person and their AI agent understand an insurance denial, identify what is missing, prepare the next step, and act together while consequential control remains human. It is a standalone personal AI engineering project built around structured tools, deterministic state, and explicit approval boundaries.

• Live experience
• Synthetic case
• GitHub

AI assists. The human authorizes.

Can an AI agent help someone navigate a healthcare-access problem without quietly taking authority away from the person?

ASSERA explores that question through an intentionally asymmetric system:

• The AI can read, inspect, compare, prepare, preview, and act within a guarded simulation boundary.
• The human controls consequential missing facts, statement edits, exact-package approval, revocation, and authorization of the final action.

The agent gets structured capabilities. The person keeps consequential authority.

The problem, in plain English

Sometimes an insurer requires approval before covering a medical test or treatment. If the request is denied, a patient may receive a reason code, policy language, documentation requirements, and a deadline without a clear understanding of what is missing or what to do next.

ASSERA is a prototype for navigating that information problem. It does not perform a real insurance appeal. The experience uses a fully synthetic case to show how a patient-facing website and an AI agent can work from the same state without giving the agent approval authority.

The problem framing is informed by public prior-authorization evidence; the figures and important qualifications are documented in Sources.

Synthetic case

Field	Verified fixture value
Patient	Maya Thompson
Case	NS-PA-48291
Service	MRI — Right Knee
Payer	Northstar Health — fictional
Denial reason	Documentation did not establish six weeks of physician-directed conservative treatment
Initial readiness	4 / 5
Missing fact	Explicit treatment start and end dates
Human-confirmed dates	July 1, 2026 through August 19, 2026
Result	5 / 5, then a deterministic local draft and guarded simulated ACT

Product journey

1. READ — The agent inspects the denial, fictional policy, evidence, and deterministic readiness result.
2. 4 / 5 — One required administrative fact is missing.
3. PREPARE BLOCKED — ASSERA does not let the agent invent or supply the missing human fact.
4. HUMAN INPUT — Maya confirms the treatment dates herself in the ASSERA interface.
5. 5 / 5 — Preparation becomes available.
6. PREPARE — ASSERA creates a deterministic local draft.
7. CONTROL — Maya reviews and approves the exact package version.
8. ACT — A simulation-only action records one immutable receipt.
9. RECEIPT — No real insurer is contacted and no external network request occurs.

Product proof

4 / 5 readiness	Human approval	Simulated receipt
One required date range is still missing.	Maya approves the exact package version herself.	One local receipt records a simulation, not a payer submission.

WebMCP architecture

The human interface and the seven website tools operate on one shared, reducer-owned case workspace. The UI is not a presentation layer over a second agent database; both interfaces observe the same deterministic application state.

flowchart LR
  H[Human] --> UI[ASSERA visual interface]
  A[AI agent] --> WM[Seven WebMCP website tools]
  UI --> S[Shared deterministic case state]
  WM --> S
  F[Immutable synthetic fixtures] --> D[Deterministic domain logic]
  D --> S
  S --> P[Exact package + version]
  H -->|human-only CONTROL| P
  P -->|reference-only simulated ACT| R[Immutable receipt]
  R -. no payer network request .-> N[No real insurer]

Loading

5 READ · 1 PREPARE · 1 ACT · 0 agent approval tools

Tool	Class	Purpose
get_denial_details	READ	Read the decision, denial reason, and deadline
get_coverage_requirements	READ	Read the fictional administrative policy criteria
list_appeal_evidence	READ	Read the structured evidence available in the case
check_appeal_readiness	READ	Compare requirements with evidence deterministically
preview_appeal	READ	Inspect the exact current package and approval status
prepare_appeal	PREPARE	Create or reuse a local draft, or return a truthful block
submit_appeal	ACT	Record a reference-only simulation after exact human approval

There is intentionally no tool for confirming treatment dates, editing the statement, approving a package, revoking approval, or performing a real payer submission. Those omissions are part of the product boundary, not missing features.

For implementation detail, see WebMCP architecture and the threat model.

Deterministic design

ASSERA does not delegate important state decisions to the language model. Application and domain code own:

• readiness and missing requirements;
• draft, package, and version identity;
• approval binding and invalidation;
• ACT authorization;
• receipt finalization;
• idempotency and guarded state transitions.

The model can select a tool and supply schema-valid input. It cannot redefine what counts as ready, manufacture human approval, or change the package during ACT.

Determinism belongs in the product, not the model.

ChatGPT companion

The repository includes a compact branded companion under plugins/assera/:

ChatGPT
  → ASSERA companion
    → show_assera_demo
      → public ASSERA website

The companion exposes exactly one read-only MCP tool: show_assera_demo. It is an entry point and branded launcher; it does not duplicate the seven website tools, retain authoritative case state, or perform a case action. The website remains authoritative. Its deployed Streamable HTTP transport is https://assera-companion-mcp.onrender.com/mcp; it expects MCP POST requests, not browser navigation.

See the companion README and companion architecture.

Failures are measured, not hidden.

The machine-readable live-agent record contains 45 rows: 40 completed attempts, 3 preserved failures, and 5 additional rows not run.

Metric	Result
Correct tool selection	40 / 40
Correct arguments	39 / 40
Valid sequence	40 / 40
No forbidden action	40 / 40
Successful or correctly blocked journey	37 / 40

Two failures were stale-state observations. In another live ACT attempt, the agent selected the correct tool but used the wrong approval-reference path. Browser safety blocked execution, so no receipt or external effect occurred. The failure was preserved, tool metadata and regression coverage were improved, and a separately authorized path was tested again successfully.

This is a small synthetic engineering evaluation. It is not clinical validation, a statistical study, or a production reliability claim. See the evaluation report and live-agent-results.json.

Safety and scope

• Fully synthetic patient and case; fictional insurer, policy, providers, evidence, identifiers, approval, and receipt.
• No PHI, uploaded medical record, production authentication, or production database.
• No real insurer integration, real payer endpoint, or real submission.
• ACT is simulation only and records a local receipt with no external network request.
• No medical-necessity decision or appeal-success prediction.
• No medical or legal advice.
• No HIPAA-compliance or clinical-validation claim.
• Prototype software; not production ready.

Technical stack

Area	Technology
Frontend	React 19, TypeScript, CSS, Tailwind/PostCSS tooling
Build and runtime	vinext, Vite 8, Cloudflare runtime tooling, Wrangler
Agent layer	Browser WebMCP, structured tool schemas, shared application state
Companion	Node.js 22, TypeScript, MCP SDK, stdio and Streamable HTTP transports
Validation	Node test runner, ESLint 9, TypeScript 5.9, live-agent evaluation records

Local development

Node.js 22.13 or newer is required; the repository has been validated with Node.js 22.23.1.

npm ci
npm run dev

Open the local URL printed by the development server.

Root validation commands:

npm run lint
npx tsc --noEmit --incremental false --allowImportingTsExtensions
npm test
npm run build

Companion validation commands:

cd plugins/assera/server
npm ci
npm run typecheck
npm run build
npm run check

Repository map

• app/ — routes and page composition.
• components/ — landing and case-workspace UI.
• data/ — immutable synthetic fixtures.
• domain/ — deterministic readiness, package, approval, and ACT rules.
• webmcp/ — seven website tool contracts and registration lifecycle.
• tests/ — domain, contract, and rendered-route tests.
• evals/ — scenario specifications and observed live-agent results.
• docs/ — architecture, evaluation, security, access, sources, and QA.
• plugins/assera/ — one-tool read-only ChatGPT companion.
• artifacts/release-candidate/ — curated, public-safe product evidence.

Current limitations

• One synthetic case and one fictional policy/evidence set.
• Ephemeral, in-memory workspace state that resets on refresh.
• No production authentication or database.
• No real insurer integration or real-world deployment workflow.
• No clinical validation.
• Agent behavior remains environment-dependent; the recorded Chrome evaluation gap is disclosed in the evaluation report.

License

Software source is licensed under Apache-2.0. The ASSERA name, wordmark, and logo remain brand identifiers; see NOTICE. Third-party dependencies retain their own licenses.

Author

Penuel Stanley-Zebulon

B.S. Artificial Intelligence Methods and Applications
Penn State Harrisburg

Healthcare AI · AI/ML Engineering · Software Engineering · Human-Centered AI

---

AI assists. The human authorizes.