AI agents are starting to use websites on people's behalf: comparing products, booking appointments, filling in forms. Today, most of them do it the hard way, by reading the page and clicking buttons like a very fast, slightly clumsy person.
WebMCP is a proposed web standard designed to make that easier and more reliable. Instead of guessing which button to click, an agent can ask the page what actions it offers, such as "search available appointments" or "add to basket", and call them directly.
Where WebMCP came from
It builds on the Model Context Protocol (MCP), the open standard that lets AI assistants connect to tools. MCP usually needs a separate server. WebMCP brings the same idea into the browser, so a website can describe its own actions to an AI agent running there, using the visitor's existing login and session.
Status in late 2026
WebMCP is still a proposal, not a finished standard. Chrome has offered it for local testing behind a flag and moved it into an origin trial for registered sites during 2026. Some platforms and vendors are experimenting with it, and claims about wide default adoption vary, so check the current state in Chrome's developer documentation before planning around it.
In short: worth understanding and prototyping, not worth rebuilding your site for yet.
How it works
A page can expose tools in two main ways:
| Approach | How | Best for |
|---|---|---|
| Declarative | Add attributes to existing HTML forms (a tool name and description) | Simple forms: search, contact, booking |
| Imperative | Register tools in JavaScript with a name, description and input schema | More complex actions: filtering, multi-step flows |
Each tool has:
- A name ("check_availability").
- A description in plain language, so the agent knows when to use it.
- An input schema defining what information it needs (date, service, number of people).
The agent calls the tool, the page runs its normal code and returns a result. Sensitive actions such as purchases can require the person to confirm.
What problems it solves
| Without WebMCP | With WebMCP |
|---|---|
| Agent reads the whole page and guesses | Agent sees a clear list of actions |
| Breaks when the design changes | Tools stay stable even if layout changes |
| Slow, multi-step clicking | One call per action |
| Errors on complex forms | Inputs defined and validated |
What to do now
You don't need WebMCP to be usable by AI agents. These steps help agents today and make WebMCP easy to add later:
- Make your site accessible. Agents rely heavily on the accessibility tree: proper labels, buttons that are real buttons, headings in order. Our WCAG 2.2 checklist covers it.
- Use clear, standard forms. Labelled fields, sensible validation and helpful error messages. Read form design best practices.
- Keep key information in the HTML: prices, availability, opening hours, policies.
- Add structured data for products, services, events and your organisation.
- Make important actions work without heavy JavaScript tricks.
- Publish an llms.txt file to point AI tools at your key pages. See llms.txt explained and our free llms.txt generator.
- Experiment with WebMCP on one simple form, such as search or a booking enquiry, using Chrome's testing flag.
For a broader list, read our AI agent-ready website checklist.
Example: describing a booking enquiry as a tool
Here's how a simple restaurant booking form might be described as a WebMCP tool, in plain terms (the exact syntax follows the specification and Chrome's documentation at the time you implement it):
| Part | Example |
|---|---|
| Tool name | request_table_booking |
| Description | "Request a table at Harbour Kitchen, London. Returns available times or confirms a request. Bookings for more than 8 people go to the events team." |
| Inputs | date (YYYY-MM-DD), time (HH:MM), party size (1–8), name, email, phone, notes (optional) |
| Validation | Date within the next 60 days; times within opening hours |
| Result | Confirmation reference, or a list of alternative times |
| Confirmation | The person confirms before the request is submitted |
Notice what makes this work: a clear name, a description that says exactly what the tool does and its limits, tightly defined inputs and a result the agent can report back. Those same qualities make a form better for human visitors too.
Write good tool descriptions
- Say what the tool does, for whom and any limits.
- Use the words a customer would use.
- Mention prerequisites ("customer must be signed in").
- Describe what's returned.
- Keep it short and specific.
Which businesses benefit most
- Booking businesses: restaurants, salons, clinics, tradespeople with online booking.
- Online shops: product search, filtering, stock checks, basket actions. See agentic commerce explained.
- Service businesses with quote forms.
- Software companies whose web apps users want to automate.
Risks to think about
- Confirm money and personal data actions. Never let an agent complete a payment without a clear confirmation from the person.
- Rate limits and abuse. Treat agent calls like any other traffic: validate inputs and protect against spam.
- Keep the human path working. WebMCP adds a faster route for agents; it must never replace the normal experience.
- Don't over-promise. A smooth agent action on a page with stale prices or unclear policies still produces a bad outcome.
The short version
WebMCP is a proposed standard that lets web pages declare actions, like searching, booking or adding to basket, so AI agents can use them directly and reliably. It's in testing and origin trials, not finished. Prepare now by making your site accessible, using clean forms, keeping key facts in HTML and adding structured data, then experiment with one simple tool.
Want your website ready for AI agents? Our web development team builds accessible, agent-friendly sites.




