OpenAI’s 10-Day WebMCP Challenge Pushes Websites to Expose Tools for AI Agents

OpenAI has opened a 10-day WebMCP Challenge around an experimental web standard that lets a site define structured actions an AI agent can use directly. The contest, run with Chromium, Cloudflare, Shopify, Vercel, Render and Netlify, aims to produce applications built for people and agents working together.

A more explicit contract between a site and an agent

WebMCP’s proposed shift is simple: a website can publish the functions it wants an agent to perform, rather than leaving the agent to infer a path through the site’s interface. OpenAI says that arrangement can make task completion faster, more accurate and more reliable, though those are company-stated expected benefits rather than independently measured results.

That is a different model from asking an agent to treat a page primarily as a visual workspace. It gives the site operator a role in defining the available actions, while the challenge asks developers to show that this shared human-agent design makes an application meaningfully better.

A browser-native tool surface, with a narrow first scope

The current specification is limited to callable tools: named functions with descriptions and typed input schemas. It does not yet include the resources, prompts or sampling primitives found in the broader Model Context Protocol, keeping the proposed browser surface narrower for now.

Those tools run in the webpage’s JavaScript context and share the user’s session, rather than requiring a separate backend server. That design makes WebMCP a potential interface at the point where an agent is already using a site, not a separate service connection outside the browser tab.

WebMCP is supported in the ChatGPT desktop app’s built-in browser and in ChatGPT Sites deployments. The standard is described as model-agnostic and was proposed through the W3C Web Machine Learning Community Group by engineers from Microsoft and Google, so OpenAI is adopting a proposal it does not solely control.

Adoption remains the constraint

The standard is still experimental, and browser availability is uneven. Chrome 146 has a developer trial behind its Experimental Web Platform Features flag; Firefox, Safari and Edge were participating in the W3C work but had not shipped implementations. That leaves the challenge as a test of developer interest and product design, not evidence of a settled web-wide interface.

The participation roster nevertheless puts browser, hosting, commerce and developer-platform companies around the same experiment. Challenge judges include representatives from Google Chrome, Cloudflare, Vercel, Shopify, Netlify and OpenAI, while Shopify, Google Chrome, Netlify, Cloudflare, Vercel and Render are offering additional prizes.

A short contest with a longer standards question

Registration and submissions opened August 25 at noon Pacific time and close September 3 at 1 p.m. OpenAI plans to announce winners September 23, although it says that date may move depending on submission volume.

The top 10 submissions will each receive $3,000, a year of ChatGPT Pro, a Codex Micro keyboard and OpenAI merchandise. OpenAI Developers described the total cash pool as $35,000, creating an incentive not only to prototype tools but to put them into web applications that agents and people can use together.

The useful result will not be a single winning app. It will be whether developers find enough real tasks worth exposing as structured tools to justify supporting a still-experimental browser standard across a wider set of sites and browsers.