Nest — Hunt flats with your agent

Inspiration

Flat hunting is a messy human workflow: filters, maps, shortlists, viewing slots, and offer letters all fight for attention. Agents already talk about housing well, but they usually fail when they have to click through a UI built for a mouse. WebMCP flips that: the site can expose the real actions (search, pin, compare, hold a viewing, draft and submit an offer) as structured tools on the same live page the person is looking at.

Nest is that idea made concrete — a London rental board where you and your agent share one session, one map, and one shortlist.

What it does

Nest is an agent-native flat hunt:

Browse seeded London listings on a live map
Pin and compare homes side by side
Propose and hold viewing slots (agent proposes; you confirm commits)
Draft and submit an offer letter from the shortlist
Inspect board state and registered WebMCP tools on an in-app Agent tools panel

You steer in plain English in ChatGPT; Nest’s site tools update the UI in real time. Anything consequential (holding a viewing, submitting an offer) is a non-read-only tool so the browser can require confirmation.

How we built it

Vite + React + TypeScript for a fast, deployable SPA
Zustand store shared by the human UI and every WebMCP execute handler (one source of truth)
Leaflet map with OSM tiles for the spatial board
WebMCP via document.modelContext.registerTool (with navigator.modelContext fallback), feature-detected, registered on the top-level page
Seeded demo listings (no paid APIs) so judges can try the full path offline-friendly
Built with Codex

Core tools include search_listings, get_listing, pin_listing / unpin_listing, compare_shortlist, propose_viewings, hold_viewing, cancel_viewing, draft_offer, submit_offer, get_board_state, and set_preferences.

Challenges

Same-state collaboration — agent tools and UI clicks must never drift. We solved this by routing both through one store.
Read vs commit — WebMCP’s value is trust. Marking search/compare read-only and viewings/offers as side-effecting was deliberate product design, not just annotations.
Demo realism without backends — landlords, holds, and offers are simulated with structured JSON so the agent can reason, while the product still feels complete.
Judges’ path — a clear happy path (search → pin → compare → hold viewings → draft → submit) plus a tools panel so WebMCP leverage is obvious in under three minutes.

What we learned

WebMCP isn’t “MCP in the browser” as a slogan — it’s best when the agent and the human must see the same thing: a map marker lighting up, a shortlist changing, an offer panel filling. Housing is a perfect stress test because the workflow is multi-step, spatial, and full of confirm-worthy actions.

Try the demo path

Open Nest in ChatGPT’s in-app browser (or Chrome with WebMCP testing enabled)
Ask: “Find 2-beds under £2200 in Zone 2 that allow pets, pin the best three, compare them, hold two viewings, and draft an offer for my favorite.”
Watch the board update; confirm when the browser asks on commits