OpenAI's WebMCP Challenge Lets Websites Talk Directly to AI Agents

OpenAI just kicked off a 10-day sprint to see what the web looks like when pages talk to agents directly instead of getting scraped, clicked, and screenshotted. The WebMCP Challenge is a hackathon co-sponsored by Chromium, Cloudflare, Shopify, Vercel, Render, and Netlify, and it lands alongside a quiet but significant product change: WebMCP is now supported inside the ChatGPT desktop app's built-in browser and on ChatGPT Sites.

Ship a live site that exposes structured tools to an agent, and if it lands in the top 10, you walk away with cash, hardware, and credits from every sponsor on the list.

What WebMCP actually does

WebMCP is a browser-side sibling of the Model Context Protocol. As a W3C browser standard, it lets a website expose its own features to AI agents as structured, callable tools through the document.modelContext API. Instead of an agent taking a screenshot of your page and guessing where to click, the website hands it a list of things it can do and the exact parameters each one needs.

Traditional MCP requires a backend server in Python or Node.js, separate authentication, and server-to-server communication. WebMCP runs entirely in the browser tab. Tools execute in the page's JavaScript context, share the user's session, and the browser enforces permissions. No backend required.

The current specification covers tools only: named functions with descriptions and typed input schemas. There are no equivalents to MCP's resources, prompts, or sampling primitives, which keeps the surface area small enough to actually ship in a browser.

Who is behind the standard

The proposal comes from the W3C Web Machine Learning Community Group, authored by engineers from Microsoft and Google, and was first announced on February 10, 2026. It is being developed as an open web standard rather than a Chrome-exclusive feature, and it is explicitly model-agnostic: it works with any AI agent, whether powered by Gemini, Claude, ChatGPT, or something open-source.

Chrome shipped it first. Chrome 146 includes a DevTrial for WebMCP, hidden behind the "Experimental Web Platform Features" flag. Firefox, Safari, and Edge are participating in the W3C working group but have not shipped implementations yet.

WebMCP arrives inside ChatGPT

The bigger news for anyone building on the OpenAI stack is that WebMCP is now a first-class citizen in ChatGPT itself. When you visit a compatible website in the ChatGPT desktop app's browser, ChatGPT or Codex can automatically use the site's registered tools to complete your task. Codex can also generate a WebMCP-enabled app and deploy it directly to ChatGPT Sites.

The team seeded a showcase to give participants a feel for the format:

• 3D Modeling: build and refine models while the agent guides each change
• Collaborative Writing: an agent that can leave comments in a shared doc under its own identity
• Crossword Builder: turn a topic into a personalized puzzle and refine clues together
• Wandernote: travel notes that become a live itinerary
• Duckboard: query and combine data with DuckDB-Wasm inside the browser

Prizes, sponsors, and a tight clock

Top 10 submissions each get $3,000 in cash from OpenAI, a year of ChatGPT Pro, a Codex Micro keyboard, and swag. Sponsors stack on top: Vercel is offering $300 per month in Vercel credits and $50 per month in Gateway credits for twelve months per winner, and Shopify is adding $250 in limited-edition Supply gear per winning submission. Google Chrome throws in Google AI Ultra subscriptions.

Registration and submissions opened today, the deadline is September 3, and winners are announced September 23. Judging includes Sarah Drasner from Chrome, Ilya Grigorik from Shopify, Andrew Galloni from Cloudflare, Jude Gao from Vercel's Next.js core team, Sean Roberts from Netlify, Alex Nahas (creator of MCP-B), and Justin Rushing, who leads OpenAI's browser agent work.

Why this matters right now

Agentic browsing has been stuck in a bad equilibrium. Every major AI vendor is building an agent that operates a browser like a human, and every site owner is either indifferent or actively hostile to that traffic. WebMCP is the first serious attempt to make the site an active participant. Sites that implement it give browsing agents a fast, reliable, token-efficient interface to work with, so an agent can call bookFlight(destination, date) directly instead of interpreting a page visually. Fewer failed interactions, lower compute costs on the agent side.

OpenAI has been consolidating its client surface, folding ChatGPT Atlas, the ChatGPT app, and Codex into a single desktop "superapp" as part of an effort to simplify its product lineup and respond to growing competition from Anthropic. A browser-native tool protocol that ships inside that superapp gives OpenAI a way to compete with Chrome's Gemini integration on Chrome's own turf, using a standard Google helped write.

Where to start if you want to enter

A few things worth trying if you are considering an entry:

1. Wrap an existing internal tool. You do not have to start from scratch, and adding WebMCP to a site you already run is explicitly allowed.
2. Build a collaborative surface where the agent and user share the same document rather than a chat sidecar. The Margin Editor and Wandernote demos are the template.
3. Expose read tools, not just write tools. Query builders, data explorers, and search UIs benefit as much from structured tool calls as checkout flows do.

You can test locally in ChatGPT's in-app browser out of the box, or in Chrome via chrome://flags/#enable-webmcp-testing. Submissions go through Devpost, and the opening livestream runs today at 3pm PT.