PaperBridge

Inspiration## Inspiration

Government-style forms — like a PAN application — are dense, jargon-heavy, and intimidating, especially for first-time applicants unfamiliar with formal English. I've filled out this exact kind of form for my own PAN card update, and for my brother's, from a remote village in Assam where digital form-filling isn't second nature for most people. I wanted to see whether an AI agent could handle the confusing parts directly, through the website's own tools, instead of a person having to decode every field alone.

What it does

PaperBridge lets a person describe their PAN application details once, in plain language, to an AI agent. The agent then calls the site's own WebMCP tools to fill out the form field by field — full name, father's name, date of birth, address — live, on screen. Once everything is filled, the site checks completeness and asks the human for a single Yes/No approval before anything is "submitted." Nothing happens without that explicit confirmation. The site also keeps a searchable history of past applications, useful if one device is used to help multiple people — like a shared community device.

How we built it

Built entirely on an Android phone using Termux, GitHub's web editor, and Vercel for deployment — no laptop involved. The frontend is vanilla HTML/CSS/JavaScript with no framework and no backend; state is held in-memory and mirrored to localStorage. Three tools are registered via document.modelContext.registerTool(): start_form, fill_field, and validate_form, chained together through a form_id returned by the first call. The agent performs all the reasoning — deciding field values from natural language — while the website only exposes structured capabilities, validates input, and renders the result.

Challenges we ran into

The biggest challenge was verification. WebMCP's live agent support in ChatGPT is currently limited to the desktop app, and I had no desktop access at any point while building this. To verify the tool logic was correct despite this, I wrote an automated test harness (Node.js + jsdom) that simulates the exact sequence of tool calls an agent would make — registration, chaining via form_id, field updates, and validation — and confirmed all three tools execute correctly and update the UI as expected. I also added a "Simulate Agent Call" button directly in the app, which manually triggers the same tool functions with sample data, so the flow can be demonstrated visually even without a live agent connection.

Accomplishments that we're proud of

Getting a fully working, correctly-chained three-tool WebMCP flow built and verified entirely from a phone, with no desktop access at any point — including independently proving the tool logic works through automated testing rather than just assuming it does.

What we learned

That verification doesn't have to mean live end-to-end testing in the exact target environment — when that environment isn't accessible, testing the underlying logic directly and being transparent about what wasn't (and couldn't be) verified is an honest, credible alternative.

What's next for PaperBridge

Testing live agent execution once desktop access is available, extending support to the minor-to-major PAN correction flow (the same three tools, a different application_type), and exploring what other government-style forms could benefit from the same conversational approach.

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for PaperBridge

Built With

• css
• github
• html
• javascript
• localstorage
• vercel
• webmcp