On August 25, OpenAI added WebMCP support to the built-in browser of the ChatGPT desktop app. WebMCP is a mechanism that lets a website declare its own capabilities, such as booking a seat or buying a product, in a form AI agents can read. Instead of guessing its way through a page, an agent now calls the entry points the site has prepared for it.

The site declares how it works, rather than being read off the screen

Until now, an AI operating a browser had to look at the rendered page, guess which button did what, and assemble a sequence of steps from there. Because it imitates human eyes and hands, a layout change or an ambiguously labeled element can break the whole task.

WebMCP inverts that assumption. Site authors publish a structured list of the operations their site supports, and the agent picks from that list. OpenAI describes WebMCP as an experimental open standard, with the specification being developed in the open on GitHub. With less room for guesswork, tasks should run faster and finish more reliably.

The built-in browser in ChatGPT desktop now supports the standard out of the box. Open a compatible site and ChatGPT or Codex can use the site's own functions as needed to carry out the request. Trying it requires updating to the latest version of the desktop app. Google Chrome can also run WebMCP once it is enabled through an experimental flag or an origin trial.

Building compatible sites from ChatGPT as well

At the same time, ChatGPT Sites, the feature for building websites inside ChatGPT, can now produce WebMCP-enabled sites. Describing the functionality you want in plain language is enough to create and publish a compatible site, without following the implementation details yourself.

OpenAI has also published examples built by its own team. They include a tool for assembling 3D models while talking to an agent, a collaborative writing tool where the agent can leave comments under its own identity, a planner that turns travel notes into an itinerary, and a data explorer that aggregates and visualizes data inside the browser using DuckDB-Wasm. All of them assume that a person and an agent are working on the same screen at once.

The 3,000 USD contest closes on September 4

OpenAI has also launched the WebMCP Challenge, a contest for WebMCP-enabled apps. Registration and submissions opened at noon Pacific Time on August 25, and the deadline is 1 p.m. Pacific Time on September 3, which corresponds to 5 a.m. Japan time on September 4. Winners are scheduled to be announced on September 23, though the date may shift depending on the volume of submissions.

The top 10 entries each receive 3,000 USD (about 480,000 yen) in cash from OpenAI, one year of ChatGPT Pro, one Codex Micro keyboard, and OpenAI swag. Additional prizes are provided by Shopify, Google Chrome, Netlify, Cloudflare, Vercel, and Render.

※1 USD = 159 JPY (as of August 26, 2026)

Judging looks at usefulness, originality, execution, whether WebMCP is used thoughtfully, and the quality of the experience when people and agents work together. The judging panel includes engineers from Google's Chrome team, Cloudflare, Shopify, Netlify and Vercel, along with the creator of MCP-B, the project WebMCP grew out of, and OpenAI's browser agent lead. Entries can be brand-new apps or WebMCP support added to an existing one.

How far the agent-ready web will spread

A standard means little if few sites adopt it. Running the contest and publishing working examples at the same time reads as an effort to get builders writing code first. Now that both ChatGPT and Chrome can act as the receiving end, sites have more reason to add support.

At the same time, declaring functionality to the outside world raises a new question: how much an agent should be allowed to do. Irreversible operations such as payments and reservations will need a clearer answer once implementations spread.

Summary

On August 25, OpenAI brought WebMCP support to the ChatGPT desktop app's built-in browser and to ChatGPT Sites. Because sites publish their operations in a structured form for agents to call directly, tasks should be more stable than automation that relies on reading the screen. The accompanying WebMCP Challenge closes at 5 a.m. Japan time on September 4, with 3,000 USD and other prizes going to each of the top 10 entries.