For two years the question was whether AI could find your website. Since August 25 the question is whether an AI agent can use it. WebMCP explained for business owners: what it is, what a site can expose, who needs it now, and five questions for your developer.
For two years the question about AI and your website was whether AI could find it. On August 25, 2026, the question changed. OpenAI added support for a proposed web standard called WebMCP to ChatGPT, which means an AI agent can now use a website directly, if the website tells it how. Not read it, not summarize it. Book the table, build the cart, request the quote.
This post explains what WebMCP is in terms a business owner can repeat, what a site has to do to support it, who it matters for right now and who can wait, and the questions to ask whoever builds your website. It is written by people who build websites for a living, so it is biased toward the practical.
What happened, in order
- May 18 and 19, 2026. Chrome shipped a developer trial of WebMCP, and Google announced a public origin trial in Chrome 149 at its I/O conference. The specification is a draft from the W3C Web Machine Learning Community Group, co-edited by Google and Microsoft. It is not yet on the standards track.
- August 25, 2026. OpenAI introduced Site tools, its implementation of WebMCP, in the ChatGPT desktop app's built-in browser, with ChatGPT Work and Codex able to discover and use a site's tools. OpenAI's framing: “With WebMCP, a website can offer useful actions directly to an AI agent alongside the interface people already use.”
- The same week. OpenAI launched a ten-day WebMCP Challenge with a $35,000 prize pool and sponsors including Google Chrome, Shopify, Cloudflare, Vercel, Netlify, and Render. The sponsor list is the tell: a browser maker, the largest hosted storefront platform, and the hosts most modern websites run on.
What WebMCP is, in one paragraph
WebMCP is a way for a web page to publish a short menu of things an AI agent is allowed to do on it. Each item on the menu has a name, a plain-language description, a definition of what information it needs, and a piece of code that performs the action using the same logic the page already uses for humans. When an agent that understands WebMCP opens the page, it reads the menu instead of guessing from the layout. The person and the agent share the same page and the same logged-in session, so the agent is acting as the user, on the user's screen, with the user watching.
The alternative, which is how agents have worked until now, is to look at the page the way a person does and try to click the right things. That is slow, it breaks whenever a layout changes, and it fails on anything complicated. WebMCP replaces guessing with a contract.
How it works, without the code
Technically, a page registers each tool through a browser interface called document.modelContext.registerTool. That is a developer detail; what matters is the shape of it. A tool is four things:
- A name and a description. “check_availability: returns open tables for a party size on a given date.” The description is what the agent reads to decide whether this tool fits the user's request, so it is written for a reader, not a compiler.
- An input schema. The exact fields the tool needs and their types: date, party size, time window. The agent fills them from the conversation instead of hunting for form fields.
- A handler. The code that runs. On a well-built site this calls the same function the human-facing button calls, so there is one source of truth.
- Annotations. Flags such as read-only, so the agent and the browser know which tools only look and which ones change something.
In ChatGPT's browser, an icon appears in the address bar when a page has tools, and the tools disappear when the page closes. Two limits worth knowing from OpenAI's own documentation: tools defined through plain HTML form attributes are not picked up, and tools inside embedded iframes are not discovered. In practice that means the site itself, not a third-party widget dropped into it, has to expose the actions.
What a business website can expose
The right way to think about this is to list the things a customer already does on your site that involve more than reading. Each one is a candidate tool.
- Restaurants and venues: check availability, hold a table, book an event space. The pick-your-table booking system we built for a Glendale restaurant is exactly the kind of flow that becomes a tool: the logic already exists, it just needs a menu entry.
- Food and retail ordering: search the menu or catalog, build a cart, apply a pickup time. Our online ordering build for a bionicos shop took orders off a third-party app and onto the shop's own site; an agent tool is the next step on the same path.
- Contractors and home services: get a quote, request an estimate with photos, check service-area coverage, book an inspection window.
- Practices and clinics: check appointment availability, request a consultation, submit a non-clinical intake. Anything with health information stays behind your compliant intake vendor, not in a tool.
- B2B and services: look up an order status, download an invoice, open a support request, schedule a call.
Who this matters for now, and who can wait
Be honest about the state of it. WebMCP is a draft standard. It works today in one browser's origin trial and one AI product's desktop app, on specific models, and OpenAI's documentation excludes Enterprise and Edu workspaces for now. The number of customers who will book your restaurant through a ChatGPT agent this month is small.
The reason to pay attention anyway is who is pushing it. Google put it in Chrome, Microsoft co-edits the spec, OpenAI built it into ChatGPT, and Shopify is sponsoring developers to adopt it. When the browser, the search engine, the assistant, and the storefront platform agree on a plumbing standard, it tends to stick. The businesses that will feel it first are the ones whose website already does something transactional: booking, ordering, quoting, scheduling, account lookups. If your site is a brochure with a phone number, there is nothing for an agent to do yet, and you can wait.
What WebMCP is not
It is not SEO. Registering tools does not make your site rank higher or get cited more in AI answers. That job belongs to the work we cover in our piece on AI SEO services: structured data, clearly answerable pages, entity clarity, reviews. Think of it as two different questions an agent asks. “Which business should I recommend?” is answered by your content and your reputation. “Can I complete the task here?” is answered by WebMCP. Winning the first and failing the second means the agent recommends you and then books your competitor because their site let it.
It is also not an llms.txt file or a schema block. Those describe your business to a model. WebMCP hands the model a working button.
The security question, which is the real question
The agent acts as the signed-in user. That is the feature and the risk. A tool that reads availability is harmless. A tool that places an order, cancels an appointment, or changes an address needs the same care you would give a human-facing button, plus one more thing: an explicit confirmation step designed for the case where the user asked for one thing and the agent understood another.
- Expose read-only tools first, and mark them read-only so the agent and the browser treat them that way.
- Keep every tool inside the permissions of the current user. An agent must never be able to do something the logged-in person could not do by clicking.
- Require a visible confirmation for anything that moves money, books a slot, or changes account data, and log every tool call with what the agent sent.
- Do not put protected health information, payment card entry, or anything covered by your privacy policy into a tool unless the same protections that apply to the form apply to the tool.
Template builder or custom site: who can actually do this
WebMCP is JavaScript running on your own page, calling your own logic. On a custom-built site, adding a tool for an action that already exists is a bounded piece of work: define the schema, wrap the existing function, mark it read-only or not, test it in the browser. On a hosted template builder, you are waiting for the platform to add it, and you will get whatever tools the platform decides to expose for everyone. Shopify sponsoring the challenge suggests storefront platforms will move; drag-and-drop site builders will move later, if at all.
This is the same split we describe in our WordPress versus custom comparison: when the web changes, a site you own can change with it, and a site you rent changes when the landlord gets around to it. Every website we build is code we control, which is why we can add an agent tool to a booking or quote flow as a normal feature rather than a platform request.
Five questions to ask your developer
- “Which actions on our site could be exposed as tools today?” A good answer is a short list of things customers already do, with the read-only ones first.
- “Does our platform let us register tools on the page, or do we wait for a vendor?” This tells you whether you own the site or rent it.
- “Would the tool call the same code as the button, or a copy?” Copies drift. One source of truth is the only maintainable answer.
- “How do we confirm and log an agent action?” If the answer is a shrug, do not ship anything that changes data.
- “How will we know if anyone uses it?” Tool calls should show up in your analytics as their own source, the way we separate paid and organic form submissions today.
Where we land
If your website takes bookings, orders, or quote requests, put WebMCP on the roadmap for the next build or the next meaningful update, starting with a read-only tool and one transactional tool behind a confirmation. If your site is informational, spend the money on the content and structure that gets you recommended, and revisit this when your site does something an agent could do for a customer. Either way, ask the five questions above; the answers tell you a lot about the site you already have. Our website development service covers both the transactional builds and the rebuilds off platforms that cannot keep up.
Frequently asked questions
What is WebMCP?
WebMCP is a proposed web standard that lets a web page register a set of named, described, typed actions that an AI agent can call directly, instead of the agent guessing from the page layout. It is a W3C Web Machine Learning Community Group draft co-edited by Google and Microsoft. Chrome ships it in an origin trial, and ChatGPT's desktop browser supports it as Site tools.
Does WebMCP help SEO or AI search rankings?
No. It does not affect whether your site ranks or gets cited. It affects whether an agent can complete a task on your site once it is there. Ranking and citation are still driven by content, structure, structured data, and reputation.
Do I need WebMCP on my website right now?
If your site takes bookings, orders, quotes, or account actions, it belongs on the roadmap for your next build or update. If your site is informational, you can wait and spend on content and structure first. The standard is a draft and support is limited to specific browsers and apps today.
Which browsers and AI apps support WebMCP?
As of early September 2026: Chrome through an origin trial that began with Chrome 149, and ChatGPT's desktop app built-in browser, where ChatGPT Work and Codex can discover a page's tools on supported models. OpenAI's documentation excludes Enterprise and Edu workspaces for now. Expect this list to change quickly.
How much does it cost to add WebMCP to a website?
On a custom site where the action already exists as code, a tool is a bounded feature: schema, wrapper, annotation, test. It is quoted like any other feature on our fixed-price builds. On a hosted template builder the cost is zero and the timeline is whenever the platform adds it, with whatever tools the platform chooses.
Is it safe to let an AI agent take actions on my site?
It is as safe as the design. The agent acts as the signed-in user, so tools must stay inside that user's permissions, read-only tools should be marked read-only, anything that moves money or changes data needs a visible confirmation, and every call should be logged. Regulated data stays out of tools unless the same protections that cover your forms cover the tool.
Sources: OpenAI, Build agent-ready websites with ChatGPT (August 25, 2026) and the Site tools documentation; Search Engine Journal, ChatGPT adds WebMCP support for Site tools (August 27, 2026); Chrome developer trial and I/O origin trial announcements, May 2026.