AI agents can already open websites, read pages and try to complete tasks. The problem is that most agents interact with a site much like a hurried visitor: they inspect the page, guess what each control does, click buttons and hope the layout has not changed.
WebMCP proposes a cleaner arrangement. A website could tell an AI agent which actions it supports, what information each action needs and what result it returns. Instead of guessing how to navigate a booking form, an agent could discover a clearly defined “prepare appointment request” tool.
Microsoft is one of the contributors to this work, alongside Google. For a small business, the interesting question is not whether to rebuild the website immediately. It is whether a handful of high-friction customer tasks could become easier when browser agents can understand them properly.
The short version: WebMCP could make selected website tasks easier for AI agents to complete, but it remains experimental. Start by mapping useful, low-risk actions. Keep the ordinary website working. Do not hand an agent permission to complete purchases, send messages or change accounts without confirmation.
What WebMCP is
WebMCP is an experimental web-platform proposal from the W3C Web Machine Learning Community Group. The current specification lists editors from Microsoft and Google. It enables a web application to expose JavaScript functions or HTML forms as structured tools that an AI agent can discover and invoke through the browser.
A tool can describe:
- its name and purpose;
- the information it needs;
- the action the site will perform; and
- the structured result returned to the agent.
The draft explores two approaches. A declarative API can describe standard actions built around HTML forms. An imperative JavaScript API can represent more dynamic workflows.
This matters because an agent no longer has to infer every action from buttons, labels and page layout. The website provides a task-level contract.
WebMCP is not a finished Microsoft product, and it is not a general switch inside Microsoft 365. It is also not a finalized W3C standard. As of 26 August 2026, the project’s implementation tracker lists origin trials in Edge 150 and Chrome 149. Small businesses should treat it as an emerging capability and use progressive enhancement rather than assuming every customer’s browser supports it.
A simple example: booking a service
Imagine a local heating company with an online booking form.
Today, an AI assistant may need to find the booking page, identify the correct fields, interpret the date picker, understand which services are available and work out whether clicking “Book” creates a confirmed appointment.
With WebMCP, the site could expose a safer sequence:
search-appointment-slotsprepare-appointment-requestconfirm-appointment-request
The first action is read-only. The second prepares a request and returns a summary. The third creates the booking, but only after the customer has seen and approved the details.
The business still controls availability, pricing, validation and account permissions on its server. WebMCP makes the supported actions easier for the agent to understand; it does not replace the booking system.
Five useful small-business starting points
1. Help customers find the right product or service
A shop, wholesaler or service company could expose a structured search action using criteria such as budget, location, availability, size or service type.
The agent could return a short list without trying to scrape every card on a catalogue page. The normal site remains available for browsing and comparison.
Good first test: read-only catalogue or service search.
2. Prepare a quote request
Many small businesses lose time because enquiry forms arrive without the details needed to price the job.
A WebMCP tool could guide an agent through the required inputs, validate obvious omissions and prepare a complete quote request. The customer reviews it before submission.
Good first test: prepare the request, but keep final submission behind confirmation.
3. Find and prepare an appointment
Salons, clinics, tradespeople, tutors and consultants could make availability easier for agents to query. The agent can identify suitable times and prepare the appointment details.
Good first test: slot search and appointment preparation. Add confirmed booking only after permissions, duplicate prevention and cancellation rules are tested.
4. Create a better support request
A support tool could collect the product, order number, symptoms, device details and steps already tried. It could then prepare a structured ticket for the customer to approve.
That gives a small team a cleaner starting point without promising that AI will solve every issue automatically.
Good first test: ticket preparation with sensitive fields minimized.
5. Check order or job status
A signed-in customer could ask an agent to retrieve the current state of an order, repair, delivery or application. The tool should return only data that the authenticated account is allowed to see.
Good first test: a read-only status lookup with clear tenant and account checks.
What WebMCP does not solve
Structured tools reduce ambiguity, but they do not remove the hard parts of running a reliable service.
Your server must still enforce:
- authentication and account permissions;
- input validation and business rules;
- rate limits and abuse controls;
- privacy and data minimization;
- duplicate-request protection;
- audit logs; and
- confirmation for consequential actions.
A tool description is not a security boundary. An AI agent can misunderstand a request, provide incorrect values or act on misleading page content. The WebMCP specification itself discusses risks including prompt injection, tool poisoning, misrepresentation of intent and privacy leakage.
Purchases, bookings, messages, account changes and deletions need a meaningful approval step. A broad “allow my assistant” setting is not informed consent for every future action.
WebMCP compared with MCP, APIs and browser automation
The similar names can cause confusion.
Regular MCP connects an AI host to separately deployed tools, services and data sources. It works well when an agent needs access outside the current browser page or across several systems.
WebMCP lets the website currently open in the browser expose selected actions in the context of that page and session.
A conventional API supports system-to-system access and usually has its own credentials, quotas, versioning and documentation. WebMCP can sit over workflows backed by an API, but it does not replace one when independent or high-volume access is required.
Browser automation clicks controls or manipulates the page. It remains useful for sites without WebMCP, but it can be fragile when layouts and selectors change. WebMCP gives supported tasks a clearer semantic interface.
What a small business should do now
You probably do not need a WebMCP development project this quarter. You can still prepare for an agent-facing web without committing to an unstable API.
Map the customer jobs
List the five most common things customers try to complete on your website. Use actions, not pages:
- find a suitable service;
- check availability;
- prepare a quote request;
- find an order; or
- create a support ticket.
Separate preparation from commitment
Design read-only and preview actions first. A tool that searches, checks or prepares is easier to test than one that buys, books, sends or deletes.
Fix the underlying workflow
If a form is confusing for people, exposing it to an agent will not repair the business process. Clean up required fields, validation, confirmation screens, error messages and backend permissions first.
Run one controlled experiment
Choose one repetitive, low-risk task. Test whether structured inputs improve completion quality compared with ordinary browser automation. Measure:
- successful completion rate;
- missing or incorrect information;
- time spent by staff correcting requests;
- duplicate or unintended actions; and
- how often a customer needs to take over.
Keep a fallback
WebMCP support is not universal. The human-facing website, accessible forms and existing integration paths must continue to work.
The practical opportunity
Small businesses do not need more AI theatre. They need fewer incomplete enquiries, less rekeying, cleaner bookings and quicker answers to routine questions.
WebMCP is interesting because it could give agents a clearer way to perform those jobs through a website. The near-term move is modest: identify one useful action, make the underlying workflow reliable, expose the minimum data and keep the final decision with the customer.
Sources and standards note
WebMCP remains under active incubation. API syntax, security requirements and browser availability may change.