ChatGPT's site tools let websites hand AI agents a proper API: no scraping required
OpenAI's Site Tools feature lets ChatGPT use structured tools websites expose via WebMCP, working inside your live signed-in session.
On 25 August 2026, OpenAI shipped a feature called Site Tools to the ChatGPT desktop app’s built-in browser. On the surface it sounds modest: a small arrow appears in the address bar when a website supports it, and ChatGPT can use the tools that site has made available. Under the hood, it represents a meaningful shift in how AI agents interact with the web.
What’s actually happening here
Until now, AI agents working inside a browser have had two options: simulate what a human does (click buttons, fill forms, read the resulting text) or connect to a separately configured backend server via the Model Context Protocol. Neither is ideal. Simulating clicks is brittle and slow. Setting up a server-side MCP connector requires technical effort that most website owners won’t bother with.
WebMCP is a proposed web standard, currently a draft under the W3C’s Web Machine Learning Community Group, that offers a third path. A website registers JavaScript functions as structured tools directly on the page, with names, descriptions, and typed input schemas. The agent reads those definitions and can call them directly. No scraping. No click simulation. No separate server.
The critical detail is that it all runs inside the user’s existing browser session. ChatGPT isn’t logging into anything separately or impersonating you in the background. It’s working with the same page you’re already signed in to, which means it has access to your account’s data and permissions, and nothing more.
What you actually see as a user
When you open a website in the ChatGPT desktop app’s built-in browser and that site has implemented WebMCP, a small arrow appears in the address bar. It’s grey when tools are available and turns blue when ChatGPT is actively using them. You can select the arrow at any time to see exactly which tools the site has offered up.
From there, you just tell ChatGPT what you want. Search the documentation. Edit the shared document. Compare the available flights. Update the shopping cart. If a relevant tool exists on that site, ChatGPT can discover and use it automatically.
Before ChatGPT acts on anything consequential, you’ll see a website-access prompt to review. Sensitive actions still require your confirmation, so you’re not handing over unattended control.
A few constraints worth knowing:
- Site Tools require GPT-5.6 Sol or GPT-5.6 Terra. GPT-5.6 Luna has WebMCP disabled.
- The feature is not available in Enterprise or Edu workspaces.
- You can turn it off entirely in the desktop app under Browser settings, then Permissions, then “Enable site tools”.
What this means for developers
WebMCP is deliberately narrower than the full Model Context Protocol. The current specification covers callable tools only: named functions with descriptions and typed input schemas. The resources, prompts, and sampling primitives found in MCP are not part of the WebMCP spec yet. That narrowness is intentional; the goal is a minimal, safe browser surface that’s easy to implement.
Two current limitations in ChatGPT’s implementation are worth noting if you’re building now. The declarative API (defining tools through HTML form attributes) isn’t supported, so you’ll need to register tools via JavaScript. Tools registered inside iframes also aren’t discovered; they need to be on the top-level page.
Chrome 146 supports WebMCP behind a developer flag. Firefox, Safari, and Edge had not shipped implementations as of 25 August 2026. For testing, ChatGPT’s in-app browser supports WebMCP out of the box, which is the most straightforward path.
Shopify confirmed that WebMCP tools were live on every Liquid storefront. Cloudflare has introduced a developer preview that adds a WebMCP bridge at the network edge, meaning you can add WebMCP support to an existing app without touching your application code directly. Expedia, Instacart, and Target are among the companies experimenting with the standard.
The WebMCP Challenge
OpenAI launched a WebMCP Challenge alongside this feature, in partnership with Google Chrome, Cloudflare, Shopify, Vercel, Render, and Netlify. There’s $35,000 in cash prizes on offer, alongside Codex Micros, ChatGPT Pro subscriptions, and other rewards. Submissions close on 3 September 2026 at 5pm PT, with ten winning projects announced on 23 September 2026.
The bigger picture
It’s worth separating WebMCP from MCP clearly, because the names invite confusion. MCP (Model Context Protocol), originated by Anthropic and now governed by the Linux Foundation, connects an AI application to a local or remote server. It works independently of any open webpage. WebMCP works through the browser, on the page you’re visiting, using your session. They’re complementary, not competing.
The broader question is whether WebMCP gets traction beyond OpenAI’s own browser. A proposed standard only becomes meaningful when multiple browsers implement it. Chrome’s developer flag is a start. The W3C community group process means other vendors can participate in shaping the specification before it stabilises. OpenAI’s choice to build on a proposed open standard rather than a proprietary protocol is at least structurally the right move for long-term adoption.
For now, if you use the ChatGPT desktop app and visit a site that has implemented WebMCP, you’ll get a noticeably more capable agent experience without doing anything different. And if you’re building a web application, WebMCP is worth understanding sooner rather than later. The plumbing for agents to interact with your site in a structured, permissioned way is being laid now.