AuctionCopilot

Inspiration

Used-car auction screening is repetitive, high-stakes work: buyers filter hundreds of lots, re-estimate value, weigh demand and risk, and simulate bids — over and over. We wanted to see what happens when a person states their goal and an AI agent does that legwork with them, right on the page, using WebMCP.

What it does

AuctionCopilot is a single-page auction-screening dashboard where humans and agents share the same tools. The person sets a goal — budget, purpose, risk appetite — and the agent searches listings, estimates value, compares lots, recommends a strategy, and simulates bids. Every action a human can click, an agent can call through WebMCP. A live activity log shows each tool call in real time. Crucially, changing only the purpose (e.g. Family to Commute) produces a completely different recommendation — the engine actually reasons, it isn't a static list.

How we built it

It's a self-contained single index.html (vanilla JS, no build step). We register six stateful WebMCP tools via document.modelContext.registerTool (with a navigator.modelContext fallback): search_listings, get_listing, estimate_value, compare_lots, recommend_strategy, and simulate_bid. Each tool's execute reads and mutates the real page state — the same data, watchlist, and selected strategy the human UI uses — so tool calls immediately update the on-screen UI (not mocked). estimate_value is readOnlyHint:true; the state-changing tools are readOnlyHint:false. Value and deal-score use a transparent, public demo formula. A Chrome Origin Trial token exposes the tools without requiring reviewers to enable any flags.

Challenges we ran into

Getting tool registration to work reliably across Chrome versions (document vs navigator modelContext), returning results the inspector reads consistently, and making every tool mutate real UI state instead of returning mock data.

Accomplishments that we're proud of

Six WebMCP tools that all drive real UI, not mocks — including non-trivial ones like recommend_strategy and simulate_bid. Changing only the purpose visibly changes the recommendation, proving the engine reasons rather than returning a fixed list.

What we learned

WebMCP shines when the same tools serve both humans and agents, and when tool results visibly change the page — the live activity log makes the agent's multi-step reasoning tangible.

What's next for AuctionCopilot

The six-tool pattern — list, detail, compare, recommend, simulate — is not car-specific. Swap the data source and it applies to real-estate auctions, used equipment, and B2B procurement. Any marketplace with listings and a decision step can wrap itself as an agent-ready surface with the same WebMCP pattern.

Built With

• chrome
• css
• github
• html
• javascript
• model-context-protocol
• webmcp