Your agent's been guessing what that button does. It reads the HTML, infers the intent, simulates a click, and hopes the DOM didn't change since last Tuesday. On August 25, 2026, OpenAI shipped Site tools - its implementation of the WebMCP standard - inside the ChatGPT desktop app. WebMCP is an open standard that lets websites declare named, callable functions that AI agents can use directly. Instead of an agent reading a page's HTML, guessing which button does what and simulating a mouse click, the site explicitly tells the agent which actions are available and how to call them.
That is a structural change in how agents interact with the web, and most teams building agent workflows haven't noticed it yet.
What WebMCP actually does, in one paragraph
WebMCP gives a website a way to publish a manifest of callable actions alongside its normal HTML. An agent visiting that page reads the manifest, sees something like add_to_cart(product_id, quantity) or find_section(heading_text), and calls it directly - no visual layout parsing required.
With Site tools, ChatGPT Work and Codex can use actions offered by a website in the desktop app's built-in browser - for example, a document editor can provide tools to find a section or add a comment.
The agent and the user share the same live session, so actions happen in context, not in a sandboxed headless copy of the page.
The analogy that holds: REST APIs did this for machine-to-machine integrations in the 2010s. They replaced screen-scraping with declared, stable contracts. WebMCP is doing the same thing one layer up - for agents interacting with web UIs rather than raw data endpoints.
Who shipped support on day zero
Millions of Shopify storefronts are already WebMCP-enabled, with Expedia, Instacart, and Target among the companies experimenting with the standard. That is not a pilot - Shopify's default Liquid theme generates the manifest automatically, so any merchant running a standard storefront got it without touching their code. A Distinguished Engineer at Shopify explained that ChatGPT's browser can use site-provided tools on Shopify storefronts to browse merchant catalogs and assemble shopping carts.
The day after OpenAI's announcement, Cloudflare introduced a developer preview that adds a WebMCP bridge right at the network edge. That matters: you don't need to modify your origin server. You can add a WebMCP manifest at the CDN layer, which means retrofitting existing apps is cheaper than most teams assume.
What changes for teams building agent workflows
Most internal agent workflows right now fall into one of two shapes: they use an API (fast, typed, version-controlled) or they use a browser agent that reads HTML and clicks things (slow, fragile, breaks on redesigns). WebMCP creates a third shape - structured actions attached to the live, authenticated, user-facing page - and that third shape dissolves the trade-off.
The concrete shift:
- Vendor apps you use today - if your team's procurement tool, analytics dashboard, or project tracker ships a WebMCP manifest, your agents stop needing a custom MCP server or a brittle scraper for that surface. The action contract comes from the vendor.
- Internal tools you build - publishing a
/webmcp.jsonmanifest alongside your internal app is now a reasonable engineering task, not a platform decision. Cloudflare's edge bridge means the origin doesn't need to change. - Authentication is inherited - you and the agent can work with the same live page and signed-in session. The agent gets the same permissions the logged-in user has, which removes a whole class of token-management problems and also means the agent cannot exceed the user's access level.
- Observability improves - because the agent is calling named functions instead of manipulating a DOM, logs show
add_comment("LGTM")instead ofclick(x=847, y=302).
One caveat worth naming: Site tools are not available with GPT-5.6 Luna or in Enterprise or Edu workspaces as of launch. Teams on enterprise plans are currently excluded, which is the usual early-adopter pattern but worth checking before you build a workflow dependency on it.
Beagle in action#ops-tools, 10:41amget_booking_details(booking_id) function, drafts a reply with the structured breakdown and a direct link to the bookingThe non-obvious consequence: agent-proofing is now a product decision
Here is the thing most coverage of WebMCP misses. Publishing a WebMCP manifest is not just a developer convenience - it is a product decision about what agents are allowed to do on your surface, and in what order. The manifest is a contract. It declares which actions are callable, what parameters they accept, and implicitly which ones aren't exposed.
That matters for two reasons. First, it means a vendor who ships a WebMCP manifest is actively choosing which agent workflows to support and which to block at the protocol layer. Second, it means teams evaluating SaaS tools will increasingly ask: does this vendor have a WebMCP manifest? Because "no" means their agents either get a brittle scraper or get nothing.
The MCP Tasks extension allows certain requests to be augmented with tasks - durable state machines that carry information about the underlying execution state of the request they augment, intended for client polling and deferred result retrieval. WebMCP and the MCP Tasks extension are complementary: Tasks handles long-running work that a WebMCP call might kick off (a document generation, a batch export), while WebMCP handles the structured surface the agent uses to trigger it. Expect them to show up together in the same workflows.
get_metric(name, period) function, gets a typed JSON response with a source referenceWebMCP site tools: common questions
What is WebMCP and how is it different from a regular MCP server?
WebMCP is an open standard for websites to declare named, callable functions that AI agents can invoke directly through a browser session. A regular MCP server is a separate backend a developer builds and hosts. WebMCP puts the tool manifest on the page itself, so agents pick it up automatically when visiting a compatible site, with no separate server required.
Which agents support WebMCP Site tools right now?
On August 25, OpenAI added Site tools to ChatGPT's built-in desktop browser. ChatGPT Work and Codex can discover and use tools provided by the current page. This requires the ChatGPT desktop app and GPT-5.6 Sol or Terra. Enterprise and Edu plans are currently excluded.
Do I need to rebuild my internal tools to support WebMCP?
Not necessarily from scratch.
Cloudflare introduced a developer preview that adds a WebMCP bridge right at the network edge
, which means you can add a manifest without touching your origin server. For green-field internal tools, publishing a /.well-known/webmcp.json manifest alongside the app is the cleaner path.
Is WebMCP the same as MCP?
They share a design philosophy but solve different problems. MCP (Model Context Protocol) standardizes how an agent connects to external tools, databases, and services via a dedicated server. WebMCP standardizes how an agent discovers and calls actions on a live web page the user is already visiting. Think of MCP as the backend integration layer and WebMCP as the front-end action layer.
What is the security model - can an agent take actions the user can't?
No. Each tool invocation goes through a safety review before it runs. And because the agent operates inside the user's existing authenticated session, it inherits exactly the user's permissions - it cannot escalate access or act on pages the user cannot reach themselves.