Inspiration
We've all been stuck on a confusing government form, googling what "Filing Status" means. We thought — what if AI could just sit next to you on the page and help? WebMCP finally makes that possible.
What it does
FormPilot lets an AI agent help you fill out complex forms (healthcare, loans, etc). You talk naturally — "My name is Alex, I live in Austin" — and the AI types it into the right boxes. It can explain confusing fields, catch errors, and prep the form for review. But it can never see your SSN or hit Submit without you.
How we built it
Vanilla JS, Vite, no backend. We registered 6 WebMCP tools using document.modelContext.registerTool() that let the AI read form state, explain fields, update values, and validate — all client-side. Forms are JSON-driven so adding new ones takes minutes.
Challenges we ran into
Making sure the AI never sees raw sensitive data was tricky — we had to mask SSNs and income before returning them from the tool. Also, making the undo system feel instant when reverting agent changes took a lot of state management work.
Accomplishments that we're proud of
The AI genuinely can't submit without human approval. The privacy masking actually works. And we shipped two completely different forms (healthcare + business loan) with zero changes to the AI code — proving it's a universal platform.
What we learned
WebMCP is a game-changer. It turns websites from "things AI scrapes blindly" into "things AI cooperates with safely." Building the trust boundary between human and agent was the hardest and most rewarding part.
What's next for FormPilot
Adding more real-world forms (taxes, visas, DMV), multi-language support, and a form builder so anyone can make their own WebMCP-powered form without writing code.
Built With
- mcp
- vanilla-js
- vite
Log in or sign up for Devpost to join the conversation.