ProofDesk

Inspiration

Benefit navigation is a shared-context problem. A person understands nuance, consent, and tradeoffs; an agent can search and organize quickly. Traditional browser automation forces the agent to infer controls and hidden state, which is especially risky in a workflow involving eligibility evidence. We built ProofDesk to explore a better boundary: structured assistance that stays visible, reversible, and under human control.

What it does

ProofDesk is a consent-first workspace built around an entirely synthetic benefit directory and demo case. A person and an agent can search programs, inspect evidence requirements, identify readiness gaps, stage a shortlist of up to three options, and prepare a local review packet on the same live page.

The crucial boundary is explicit: ProofDesk never submits an application or sends evidence. The final preparation tool returns submitted: false, while the interface keeps the person as the only decision-maker.

Why WebMCP

ProofDesk exposes seven narrow tools tied to the exact state the person sees:

• search_support_programs
• get_program_requirements
• review_case_readiness
• set_evidence_consent
• stage_application_plan
• prepare_review_packet
• reset_demo_case

Three tools are read-only and four make local, visible changes. Every input uses a narrow JSON schema. State-changing results appear immediately in the normal UI and can be reviewed or reversed. WebMCP removes the brittle step of guessing at DOM controls and gives the agent a precise, inspectable contract.

How we built it

The app is a responsive React 19 and TypeScript experience built with Next.js, Vinext, Vite, and Tailwind CSS. The top-level page detects document.modelContext.registerTool and registers the seven tools directly against the live application state. Read operations declare read-only behavior; write operations describe their effects in plain language.

The public build is hosted on ChatGPT Sites. Browsers without WebMCP retain the complete manual experience, so agent access is an enhancement rather than a dependency.

Safety and privacy

All programs, people, amounts, and evidence are synthetic. There is no authentication, upload, personal record, external API call, or hidden submission. The product demonstrates a review workflow, not real eligibility advice. A real deployment would require official-source verification, jurisdiction logic, accessibility review, and a documented privacy program.

Challenges

The hardest design challenge was making agent actions useful without making them opaque or overpowered. A single broad “apply” tool would have been easy to demo and wrong for the domain. We instead decomposed the flow into narrow capabilities, capped the shortlist, surfaced consent state, and made every mutation visible.

What we learned

WebMCP is strongest when it does more than accelerate clicks. It lets a site define a shared protocol for human-agent collaboration: what can be read, what can change, and where responsibility must remain with the person. That changes the product architecture, not just the automation layer.

What is next

The next step is to connect verified public program sources, add jurisdiction-aware validation, expand accessibility testing, and design a privacy-preserving evidence vault—while keeping the same consent-first contract.

Built With

• chatgpt-sites