AgentBridge

Making online commerce ready for AI agents.

AgentBridge is an agent-native commerce platform that lets AI agents shop on your behalf — no screen scraping, no brittle browser automation, just natural conversation.

AgentBridge was created specifically for the OpenAI WebMCP Challenge. It demonstrates how a website can expose meaningful commerce capabilities directly to AI agents while keeping the complete experience visible and controllable for the user.

The Problem

Online shopping was designed for humans. When AI agents try to shop for you today, they rely on fragile workarounds: automating clicks, reading pixels off a screen, and breaking every time a website changes its layout. It's slow, unreliable, and not how agent-powered commerce should work.

How AgentBridge Works

Instead of forcing AI agents to pretend to be humans on websites, AgentBridge gives them a structured layer of semantic actions — things like search restaurants, compare prices, add to cart, and place order. Agents understand these actions natively, so the entire shopping experience becomes a simple conversation.

Here's what it looks like in practice:

1. You say what you want — "Find me a spicy dinner under $12."
2. The agent understands the intent — It separates the request into meal type, flavor, dietary preferences, and budget instead of relying on a literal keyword match.
3. The agent discovers options — It searches dishes across every open restaurant, ranks the strongest matches, and can build the cart for you.
4. You approve — The complete order and price breakdown remain visible before checkout or purchase.

No apps to open. No menus to scroll through. No friction.

What Makes It Different

• Voice-first interaction — Talk to the agent like you'd talk to a friend. It understands context, preferences, and budget.
• Structured commerce layer — Agents interact through clean, reliable capabilities (powered by WebMCP) instead of scraping websites.
• Intent-aware meal discovery — Requests such as "healthy lunch under $15" or "spicy dinner under $12" become structured constraints. AgentBridge searches across restaurants using meal type, price, dietary tags, and intent facets, then ranks products by relevance, popularity, restaurant quality, and price.
• More than restaurant search — The agent can discover individual dishes globally, resolve products reliably by ID, manage options and quantities, inspect totals, create checkout links, place orders, and follow order status through 12 purpose-built WebMCP tools.
• Human in the loop — The agent handles the legwork, but you always make the final call before anything is purchased.
• First-party marketplace data — A deterministic catalog keeps restaurant, menu, product, image, cart, and checkout state consistent throughout the full demonstration.

Intelligence You Can See

AgentBridge does not hide agent activity behind a chat transcript. The storefront responds alongside the conversation so users can understand what the system interpreted and what it changed:

• The search bar mirrors the agent's interpreted query, including constraints such as spicy dinner under $12.
• Intent chips expose the active meal type, keywords, dietary requirements, and budget.
• Cross-restaurant meal cards make ranked results, prices, ratings, tags, and availability easy to compare.
• Live WebMCP indicators show when a tool is running and whether it completed successfully.
• Agent-created checkout URLs restore the exact cart, selected options, quantities, delivery details, and meal thumbnails for human review.
• Resilient image fallbacks keep products recognizable even when a primary image source is unavailable.

This shared interface improves trust: the agent gets structured actions and machine-readable results, while the user gets immediate visual feedback, transparent constraints, and a clear approval point.

The Vision

We started with food ordering, but the concept applies everywhere: groceries, retail, travel, services. AgentBridge is infrastructure for a future where humans describe what they want and AI agents handle the rest — reliably, across any store and any category.

Built With

AgentBridge is built with Next.js, TypeScript, OpenAI models, and WebMCP, an emerging standard that makes the web agent-friendly by giving AI models structured access to website capabilities.

Testing WebMCP

AgentBridge implements the W3C WebMCP browser API through document.modelContext. This is different from products named "Web MCP" that expose generic browser automation, scraping, screenshots, or clicking through a backend MCP server.

Start AgentBridge, then run the end-to-end WebMCP test in another terminal:

npm run dev
npm run test:webmcp

The test reproduces "Find me a burger for one person, with a total budget of up to USD $20." It discovers all registered tools, searches burger restaurants, selects an item, verifies the final total including fees and tax, requests a checkout URL, and confirms that a fresh browser page restores the order for human review.

To test a deployment instead of localhost:

npm run test:webmcp -- https://agentbridge-delta.vercel.app

This automated test validates the website's WebMCP implementation using the included polyfill. To validate native agent integration, use a current client listed in the WebMCP implementation status, open the deployed AgentBridge page in that client's browser context, and make the same request. ChatGPT Desktop is listed as supported; the standard chatgpt.com browser session is not listed separately.

---

Created for the OpenAI WebMCP Challenge.