The website exposes a tool
For example search_products, check_stock or book_appointment, with a structured input schema.
Practical AI / Agent-ready websites
WebMCP & Site Tools
Give compatible AI agents a structured way to use useful parts of your website or web application instead of making them guess their way through buttons, forms and page layouts.
WebMCP is an emerging proposed web standard. We build it as progressive enhancement so the normal website continues to work for people and non-supporting browsers.
Live on Atinamos:
our AI Website Check exposes a WebMCP
check_website site tool
·
human interface and agent interface use the same backend
What is WebMCP?
Most websites are designed for a person to read the page, understand the controls and click through a sequence of steps. AI agents can try to do the same, but every button, label and layout change creates another opportunity for the agent to misunderstand what the website means.
WebMCP adds a structured layer. A page can describe useful actions as tools with clear names, descriptions and inputs. A compatible browser or AI agent can discover those tools and call them directly while the website remains the place where the action happens.
It is not another chatbot and it does not replace the normal website. It is a way of making selected website capabilities easier for agents to understand and use.
The agent discovers it
A compatible browser or agent can see what the page allows instead of inferring every action from the visual interface.
The existing website does the work
The tool can reuse the site's current APIs, forms, session state and business logic, subject to the same controls you choose to enforce.
A live example on this site
The Atinamos AI Website Check is a normal browser tool for people. The same page also registers a WebMCP site tool called check_website.
A compatible agent can supply a public website URL and receive the same six-page evidence-based inspection used by the human interface. We do not maintain a second "AI version" of the product; both routes use the same hardened backend.
> tool: check_website
> input: {"url":"https://example.com"}
> scope: up to 6 representative pages
> checking crawlability, structure, machine signals...
> structured findings returned
Possibilities
What could WebMCP do for a business?
The useful question is not “how many tools can we expose?” It is “which tasks would become clearer and more reliable if an agent could call a well-defined action?”
Online shops
Help an agent work with the catalogue and basket using actions the shop actually understands.
search_productscheck_stockcompare_productsconfigure_bundleadd_to_cartestimate_delivery
Bookings & services
Make common service journeys easier for an agent to complete without guessing at form fields or date controls.
find_servicecheck_availabilityrequest_quotebook_appointmentsubmit_enquiry
Portals & internal systems
Expose tightly controlled actions inside authenticated applications where an agent can help the signed-in user.
- find a record
- create a task
- update a workflow stage
- run a diagnostic
- prepare a report
Example: a WooCommerce shop
A customer could ask an AI agent to find a product in stock, compare two options and add the chosen variant to the basket. WebMCP can give the agent explicit shop actions while the existing WooCommerce catalogue, pricing, stock rules and checkout remain authoritative.
Try it
How can WebMCP be accessed today?
Support is still evolving. The routes below reflect the state of the ecosystem in September 2026 and may change as the proposed standard develops.
ChatGPT desktop Site Tools
OpenAI's built-in browser can discover WebMCP Site Tools on supported pages when the account and selected model have access. The address bar shows when site tools are available and ChatGPT can use matching tools with the normal website-access and safety checks.
Chrome developer testing
Chrome documents WebMCP as an experimental web platform feature. Developers can use the origin trial, or enable the local testing flag at chrome://flags/#enable-webmcp-testing. Availability can vary across Chrome release channels while the feature is being developed.
Chrome inspection & manual calls
Chrome's WebMCP tooling can inspect registered tools, schemas and outputs. Developers can also retrieve tools with document.modelContext.getTools() and execute a discovered tool with document.modelContext.executeTool().
Human website remains first-class
WebMCP is designed as progressive enhancement. A visitor without WebMCP support should still get the normal website, forms, shop and navigation. Agent access is an additional interface, not a replacement for people.
Experimental technology: WebMCP is a proposed standard under active development. Client support, API details and safety behaviour can change. We therefore avoid making a business workflow depend exclusively on WebMCP at this stage.
Atinamos service
We can help make an existing website agent-ready.
This is not about adding AI for the sake of it. We identify useful actions, reuse the systems you already have and expose only what has a clear purpose.
Opportunity review
Identify website tasks that are genuinely worth exposing to an agent and separate useful actions from novelty.
Tool design
Define clear tool names, inputs, outputs and read/write behaviour so agents understand the action correctly.
Implementation
Connect WebMCP tools to existing forms, APIs, WooCommerce, booking logic or bespoke application functions without replacing the human interface.
Safety & permissions
Keep read-only and write actions distinct, minimise data passed to agents, preserve authentication boundaries and add confirmation where an action is sensitive.
Testing & rollout
Test discovery and execution in available clients, retain graceful fallback for ordinary browsers and document what is experimental.
WebMCP is not the same thing as an MCP server.
WebMCP exposes tools from the live webpage and can use the page's current state and signed-in session. A conventional MCP server is a separate machine-facing service that can be available without a user having that webpage open.
They can complement each other. Which approach makes sense depends on whether the job belongs inside the user's browser session, in a backend integration, or both.
Make the useful action explicit. Keep the business logic authoritative.
We see WebMCP as another interface to a well-designed website or system, not a reason to duplicate the underlying workflow or hand control to an agent without boundaries.
Could your website expose useful Site Tools?
Tell us what visitors or staff currently need to do. We can look at whether WebMCP would make any of those journeys easier for compatible AI agents without making the normal website worse.