WEXSPACE ProofDesk

A governed WebMCP evidence workbench where humans and browser agents share one auditable workflow while final release remains human-controlled.

Comment

Inspiration

Browser agents can act quickly, but high-consequence work becomes hard to trust when intent, evidence, verification, and release authority are hidden inside an automation trace. WEXSPACE ProofDesk makes that collaboration visible and accountable.

What it does

ProofDesk gives a human and a WebMCP browser agent one shared evidence workspace. The human defines the real deliverable contract and acceptance criteria. Five narrow WebMCP tools let an agent inspect the workspace, create a bounded work item, add criterion-linked HTTPS evidence, run a deterministic evidence-coverage gate, and prepare a passing package for human review. The visible UI reads the same domain state and audit ledger.

The workflow ends at AWAITING_HUMAN_REVIEW. No WebMCP tool can approve or release work.

Why WebMCP is the right fit

WebMCP makes the website itself an explicit, typed agent interface. The agent does not scrape button labels or depend on brittle coordinates, while the human keeps the full visible product experience. Tool names, input schemas, side effects, and authority boundaries are declared by the site.

What people and agents can do together

Before ProofDesk, a human must reconstruct what was requested, which claims have evidence, whether the acceptance criteria passed, and whether release occurred. ProofDesk gives both sides one deliverable contract, one evidence set, one deterministic score, one audit history, and one human-controlled release boundary.

How we built it

  • Five imperative WebMCP tools registered with document.modelContext.registerTool
  • Shared TypeScript domain functions used by both the React UI and WebMCP adapter
  • Deterministic criterion-coverage scoring
  • Versioned browser-local persistence with localStorage
  • Auditable state transitions
  • Human-only approval and release controls
  • Production deployment on ChatGPT Sites
  • Unit, contract, rendering, and public judge-path tests

Challenges and accomplishments

The hardest part was proving that WebMCP actions and the visible human interface mutate the same real workflow state while preventing an agent from crossing the final release boundary. The public judge path discovered all five tools, added three evidence items, produced a 100% deterministic gate score, and reached AWAITING_HUMAN_REVIEW with releasePerformed: false.

What we learned

WebMCP is most valuable when it is treated as a governed product interface rather than an invisible automation shortcut. Narrow tools, shared state, explicit side effects, deterministic checks, and a clear human boundary make browser-agent work easier to inspect and trust.

Prior work and challenge-new work

The WEXSPACE name and high-level governance ideas predate this challenge. ProofDesk itself—including its product scope, React/Vinext application, shared domain engine, five WebMCP tools, deterministic gate, persistence, audit behavior, tests, deployment, documentation, and demo evidence—was created during the WebMCP Challenge period. Earlier WEXSPACE/IEW engineering work is unrelated prior work and is not presented as newly created here. The app started from the OpenAI Sites Vinext starter; standard framework plumbing is scaffold dependency, not an original product claim.

Judge test path

  1. Open the live application in ChatGPT's in-app browser or Chrome 149+ with WebMCP testing enabled.
  2. Discover the five wexspace.* tools.
  3. Create a package with 2–6 acceptance criteria.
  4. Add medium- or high-confidence HTTPS evidence for every criterion.
  5. Run the deterministic gate.
  6. Prepare human review.
  7. Confirm that both the visible UI and the tool response show 100% coverage, AWAITING_HUMAN_REVIEW, and releasePerformed: false.
  8. Confirm that no agent release tool exists.

Live app: https://wexspace-proofdesk.moatazalqobati20.chatgpt.site/

Source: https://github.com/moatazfarea/wexspace-iew-agentic-fleet/tree/webmcp-challenge-2026

Security and limitations

ProofDesk requests no credentials and stores workspace state only in the current browser profile. Evidence URLs must use HTTPS. Evidence confidence is declared rather than independently fact-checked. This release supports one active package per browser profile, and native discovery requires a WebMCP-capable browser. ProofDesk is a governance and evidence workflow, not a professional certification authority.

Built With

Share this project:

Updates

Submission history