Paperwork

A gnarly form your agent fills in while you watch — and correct.

Nobody's brain fails at forms; their patience does. A renters-insurance application asks forty questions, a third of which only appear if you answered an earlier one a certain way, in its formats and its order. You already know all of it — the form just makes you retype it its way.

Paperwork flips the order. You tell the agent what happened in plain language; it drafts answers onto the real form, in front of you, and you accept or correct each one with your own keyboard. Confused by a question? The agent can jump to it and explain what the insurer is actually asking for.

Built for the WebMCP Challenge.

---

What makes it a WebMCP app, not a chatbot

Nothing the agent writes is an answer yet. Agent writes land in a pending layer — highlighted in the form, with accept and edit controls — and count toward the application only once the human accepts them. The form tracks each field's provenance (empty, human, pending, accepted, uncertain), so an agent can never overwrite something you typed and then quietly present it back to you as your own answer.

The agent can say "I'm not sure." Any draft can carry a note, which marks the field uncertain and shows your reason beside it. An agent that guesses at your deductible and flags the guess is far more useful than one that guesses silently — and the form makes that difference visible rather than leaving it in chat.

There is no submit tool. Structurally. The agent gets review_submission, which is read-only. The three attestations — the fraud acknowledgement, the animal attestation, the final accuracy confirmation — are marked human-only and cannot be touched by any tool at all. Signing is a legal act; it stays with the person whose name is on the form.

The agent and the human share one code path. Every tool routes through the same setField() the inputs commit through, via one invoke() chokepoint. Conditional fields spawn, repeaters grow, and validation re-runs identically whichever side acted.

The tools

Tool	What it does	Read-only
get_form	The whole application as JSON: six steps, every field's id, type, options, value, provenance, whether it's required or human-only, and why it is hidden if it is	✅
focus_field	Jumps the form to the step holding a field, scrolls it into view and rings it — so you're both looking at the same question	✅
explain_field	Why the insurer asks a question, what leaving it blank costs the application, and what a typical answer looks like	✅
validate_form	Checks the form against its own rules and returns three buckets: errors, missing, and worth-a-second-look. Whole form or one step	✅
review_submission	Reads the application back in prose before signing: what's complete, what's awaiting review, what was flagged uncertain	✅
fill_field	Drafts one answer into the pending layer. Commits nothing. Add a note to flag it uncertain
fill_section	Drafts several answers into one step in a single call, with per-field notes
add_repeater_item	Adds an item to a repeating list — prior claims, or scheduled high-value items — and can draft its fields at the same time

Running it

WebMCP needs an origin-isolated document, so file:// will not work — serve it over HTTP:

npx serve .

Or, with no Node installed:

python3 -m http.server 8000

Then, to let a real agent drive it:

1. Chrome 149 or newer → chrome://flags/#enable-webmcp-testing → Enabled → relaunch.
2. Open the served URL.
3. Install the Model Context Tool Inspector extension, or open the page in ChatGPT's in-app browser, and describe your situation in plain language.

For a public deployment, join the WebMCP origin trial so visitors get WebMCP without touching a flag.

No agent handy? The Tool console tab at the bottom of the page runs every tool by hand with JSON arguments. The status pill tells you which mode you are in. Nothing about the app breaks when WebMCP is absent — it degrades to an ordinary form.

Things to say to the agent

• "I rent the top floor of a 1974 duplex, moved in two years ago, about $30k of stuff. Fill in what you can and flag anything you're guessing at."
• "What does 'construction class' actually want here?" — it jumps to the field and explains it.
• "I've had one claim — kitchen fire, March 2023, about $4,000. Add it."
• "What's still missing before I can submit?"
• "Read the whole thing back to me before I sign."
• Then try "submit it for me" — it will tell you it can't, and why.

What's in here

index.html    the entire app — no build step, no runtime dependencies
vercel.json   static deploy config (no build)
LICENSE       MIT

One file, vanilla JS, zero runtime dependencies. The form is a six-step Marlow Mutual renters application with conditional fields, two repeating lists, and cross-field validation. State is in memory by design; a reload gives you a blank form.

Developed against a headless-Chromium Playwright suite that drives the app through the same tool entry points an agent uses.

Demo video script (~2:30)

1. 0:00 — Show the blank six-step form next to Priya's brief: a plain-language paragraph — where she's moving, her partner, the dog, the bike, one prior claim. "This is what a person actually knows. Not what the form asks for, in the form's own order."
2. 0:15 — Ask the agent to fill the Applicant step from the brief. Fields light up purple as drafts land — fill_field, fill_section calls in the activity log. "Nothing is answered yet. It's drafted, waiting for her."
3. 0:40 — Point at "Time at current address" — a select field, options under_2/2_plus — correctly set from "moved in two years ago" in the brief. "It didn't paste text into a dropdown. It matched what she said to what the form actually offers."
4. 0:55 — Ask the agent to check the fraud acknowledgement for her. It refuses, live: "fraud_ack is a human-only attestation — Priya must answer it herself." "That's not an instruction telling it not to. There's no tool that can touch it."
5. 1:15 — Click the ✓ on one drafted field — the highlight clears, it becomes her own answer. "I'm not retyping anything it got right. I'm just saying yes."
6. 1:35 — Ask "what does construction class actually want here?" — explain_field jumps to it and explains in plain language, next to the field itself.
7. 1:55 — Give the agent the claims detail from the brief — one prior claim, water damage, March 2024, about $1,800 — and watch add_repeater_item spawn a numbered claim entry and fill it.
8. 2:10 — Ask "what's left before I can submit?" — review_submission reads the application back in prose. Point out the Submit button itself, disabled, and that there is no submit tool in the list at all — only a human hand can press it.
9. 2:25 — Close on the form, mostly filled, three attestations still waiting on her.

License

MIT — see LICENSE.