Duet.

Human + agent co-operate one live commerce workflow — the agent works through document.modelContext tools, not by guessing at pixels. The approve step stays human.

Catalog

Compare options

Side-by-side price / stock / lead time — the output of the agent's compare_options tool.

Cart

Shared state — agent and human see the same lines.

Approvals

propose_order creates a pending proposal; you approve it. That loop is the pitch.

Inventory register

Live stock + lead times — the output of check_inventory.

Item	Stock	Lead time	Status

Agent console

Call any registered tool directly — identical code path to a real agent. Works in every browser; with WebMCP enabled, the same seven tools are on document.modelContext for ChatGPT / Chrome.

Arguments are JSON. Empty object = no args.

{ }

Reliability telemetry

The challenge asks for “faster, more accurately, more reliably.” This is the number behind that claim: five scripted shop tasks replayed through the UI path vs. the WebMCP tool path.

Duet — OpenAI Devpost WebMCP Challenge entry. Vanilla TypeScript + Vite, no framework.
Test with real WebMCP: Chrome → chrome://flags/#enable-webmcp-testing (or ChatGPT's in-app browser), then ask the agent to “compare the two solenoid valves and propose an order for the cheaper one.”
MIT licensed — see LICENSE. Reusable core: src/webmcp-kit.ts.