ChatGPT’s built-in desktop browser can now call structured “site tools” that a webpage exposes on its own, letting the assistant search, edit, or complete a task directly on the page instead of clicking and typing its way through the interface. OpenAI rolled the feature out in the ChatGPT desktop app on August 25, 2026, alongside a 10-day WebMCP Challenge inviting developers to build sites that use it.
This is separate from the server-based Model Context Protocol (MCP) that ChatGPT has supported since 2025. Server MCP connects ChatGPT to a remote service and works without a browser tab open at all. WebMCP is narrower and more local: the tools live in the JavaScript of the page a person is actually visiting, and they only exist while that tab is open.
How Site Tools Actually Work
A website registers named JavaScript functions as tools, each with a description and a defined input schema, using the WebMCP standard. When ChatGPT’s desktop browser lands on a page that exposes tools this way, an arrow icon appears in the address bar showing whether the available tools can read data or make changes. The assistant can then call those functions directly, inside the same signed-in session the person is already using, rather than guessing at buttons and form fields.
Because the tools are scoped to the page, they disappear the moment that tab closes and do not carry over to a different page on the same site. A document editor might expose a tool for leaving a comment; a dashboard could let the assistant set a date range and pull the underlying numbers; a travel or booking site could let it compare options and update a cart, all without OpenAI having to build a custom integration for each one.
Which ChatGPT Setups Actually Support It
Support is narrower than a blanket “ChatGPT now has this.” It depends on the model, the account type, and even the browser:
| Requirement | Status |
|---|---|
| GPT-5.6 Sol / GPT-5.6 Terra | Site tools supported |
| GPT-5.6 Luna | WebMCP currently disabled |
| Enterprise / Edu workspaces | Not available |
| ChatGPT’s built-in desktop browser | Supported (update to latest app version) |
| Chrome (regular) | Not supported; only behind an experimental flag in Chrome 146 or via Chrome’s origin trial |
| Firefox, Safari, Edge | No public implementation yet |
Even where the model and app version line up, a given page only offers tools if its developer has actually built them. Plenty of sites will show no arrow in the address bar at all for the foreseeable future, and tools discovered on one page won’t automatically appear on another.
Early Adopters and the WebMCP Challenge
Adoption isn’t purely theoretical. Millions of Shopify storefronts are already WebMCP-enabled, and Expedia, Instacart, and Target are among the companies reported to be experimenting with the standard. OpenAI is also using ChatGPT Sites, its tool for building websites through ChatGPT itself, to let anyone spin up a WebMCP-ready site just by describing what they want it to do.
To push adoption further, OpenAI launched a 10-day WebMCP Challenge with entries closing around September 3-4, 2026 (the exact cutoff depends on time zone) and winners expected roughly three weeks later, though that date can shift. The top entries split prizes that include cash, a year of ChatGPT Pro, and contributions from Shopify, Google Chrome, Netlify, Cloudflare, Vercel, and Render — a lineup that signals real interest from the hosting and commerce side of the web, not just from OpenAI.
What’s Missing Right Now
The current implementation is a subset of the full WebMCP proposal, and OpenAI has been explicit about what it leaves out:
- No declarative tools. Tools defined purely through HTML form attributes, rather than registered in JavaScript, aren’t picked up as site tools yet.
- No iframe discovery. Tools registered inside an iframe, same-origin or cross-origin, are invisible to the browser; they have to be registered on the top-level page.
- No resources, prompts, or sampling. Unlike the broader Model Context Protocol, the current WebMCP spec is limited to callable tools — named functions with a description and an input schema — and doesn’t yet cover the other primitives server-based MCP supports.
None of this stops ChatGPT from working on a page that has no site tools at all; it simply falls back to its ordinary browsing behavior of clicking and typing like a person would.
Permissions and the Trust Problem
Because site tools run inside a signed-in session, they can change real data. OpenAI asks for explicit permission before ChatGPT uses a site this way, and requires separate confirmation for anything sensitive: purchases, deletions, account or settings changes, sending messages, or sharing personal information. OpenAI has also flagged the two risks that come with any agent acting inside a live webpage — a malicious or misleading tool description tricking the assistant into doing the wrong thing, and content on the page itself attempting a prompt injection. A safety review runs on each tool call, but that review can’t vouch for the trustworthiness of the website or the data it returns, which keeps the underlying question of which sites should be trusted with this kind of access firmly unresolved.
Why This Matters Beyond ChatGPT Itself
Google’s own developer guidance has already been nudging site owners toward building for AI agents rather than only for human visitors, and WebMCP is the first concrete, shippable version of that idea from a major AI provider. For businesses in Nigeria and the US alike, this sits next to a broader shift already underway: assistants increasingly complete tasks and answer questions on a site’s behalf rather than simply linking out to it, which is the same dynamic behind rising interest in answer engine optimization and in structuring content so AI systems can use it directly.
WebMCP doesn’t change how a page gets discovered or cited — that’s still governed by the same crawler access and content-quality factors covered in our guides to getting cited by ChatGPT and Perplexity and preparing a website for AI platforms. What it changes is what happens after an assistant is already on a page: whether it can act there directly, or has to fall back on clicking around like a confused first-time visitor. OpenAI’s own documentation is upfront that it hasn’t connected WebMCP to search rankings, citations, or recommendations in any way, so exposing site tools is not currently an SEO or AEO tactic on its own — it’s a task-completion feature that happens to matter to the same site owners tracking whether AI search engines are mentioning their brand and reading up on how ChatGPT surfaces local businesses.
For most small and mid-sized sites, including the majority of Nigerian and US small-business sites, there’s no urgent action here: WebMCP is experimental, browser support outside ChatGPT’s own app is still limited to a Chrome flag, and the specification itself isn’t on a standards track yet. The sites worth watching this closely right now are ones with real on-page workflows worth automating — booking, ordering, account management, search — where a well-scoped tool could genuinely save a visitor steps. For a brochure-style service site, the more immediate lever is still making sure AI systems can find, understand, and cite what’s already there.
Frequently Asked Questions
Keeping a site ready for how AI systems actually find and use it — crawler access, structured content, and now an eye on emerging standards like WebMCP — is exactly the kind of ongoing work covered in our content and AEO strategy services.