Make Websites Agent-Ready with WebMCP: Front-End Tools for Browser Agents, the Tool Menu, MCP vs WebMCP, and a New Service Offer

Every website has forms - search bars, date ranges, filters, sort - and browser agents handle them slowly: they snapshot the page, guess where to click, fumble drop-downs and date-pickers, and sometimes get stuck. As more people delegate booking and buying to agents, the site that lets the agent act reliably gets the traffic. Harshit frames this as a new service like SEO: an owner of a small hotel chain or shop does not know or care what an MCP is, but wants agent-driven customers. The pitch is not 'I will enable WebMCP' - it is 'your site is not ready for AI agents, and I can fix it or build a new one with this built in'.

WebMCP is a standard for exposing structured tools from a web page. Each operation on the page - the search field, the date picker, the guest counter, a filter - becomes a tool with a name, a description and a JSON schema of its inputs. Harshit likens it to accessibility standards: ARIA labels and image alt text exist so a screen reader can describe a page; WebMCP exists so an agent can operate one. Everything an agent needs is context, so clearly described fields are the whole game. A typical form-heavy site has 10-20 such tools; a marketplace like a directory or classifieds site could have around 100. The old term for what browser agents do today is 'actuation' - simulating mouse clicks and typing as though they were the human user; WebMCP turns that into a single structured call.

Ordinary MCP is provided from the vendor's backend by its engineers: you connect Gmail or a project-management tool to your coding agent and the data flows from the servers, no tab opened, available any time, running in the cloud if you like. WebMCP is client-side and UI-oriented: it exists only on the website in an open browser tab, built for browser agents, with a human in the loop (you still look at the product before buying). It reduces token use because the agent does not spend turns discovering where to click. Which to choose depends on ownership: Harshit's own marketing-metrics SaaS already has an MCP, so it is a poor WebMCP example; Amazon has no MCP for shoppers and huge filters, so it is a strong one. He also stresses you still need to be the site's front-end engineer - WebMCP does not let you add tools to someone else's site.

The demo isolates one variable. On branch 1 the site is plain HTML and the agent operates it like a person. On branch 2 the site registers a tool, and the identical request resolves in one call. On branch 3 a small set of tools supports a whole journey - search, compare, then fetch the booking code - with an on-page console proving which calls came from WebMCP rather than clicking. Harshit says the tools were written by asking a coding agent, not by hand.

Each field in the search tool gets a schema entry: destination is a string limited to three allowed values, described as the city to search; check-in and check-out are dates; guests is an integer with minimum 1 and maximum 12, described as total guests; further fields cover minimum bedrooms, maximum total price, amenities, free cancellation and sort order. The page loads this script with everything else; agents receive it behind the scenes as soon as the page loads, from the front end, and never see a visible change. The site owner does not hand-write it: the practical route is to tell a coding agent to add a JSON schema for the form fields and give it the WebMCP documentation. 'If it is a drop-down, it will be an enum with options'.

The standard supports discovery (a standard way for a page to register tools), JSON schemas for inputs, and state (a shared understanding of the current page context, so the agent knows what it can act on in real time). Two APIs: imperative - define tools with standard JavaScript for form inputs, navigation, state management or other functions; declarative - add annotations directly to standard HTML forms with no extra JavaScript. Harshit recommends the imperative API in most cases, calls the technical detail skippable for non-developers, and points people to the official WebMCP versus MCP and 'when to use WebMCP' guides. Both ChatGPT's browser and Chrome are implementing it.

Harshit's candidate list: marketplaces and classifieds (Craigslist, Zillow, Just Dial, eBay), job portals, Amazon-style catalogues, Booking.com-style travel, and support-heavy sites where finding the right form is the problem (he describes struggling to find the right GoDaddy support route for an expired domain). Official use cases: customer support (route the agent to the right form and fill the fields), travel booking, structured forms, support-agent interactions, and application debugging. The official 'shopping journey' example: a parent asks an agent to price a birthday shopping list across two or three stores and build a wishlist; the store's tools - product search, add to wishlist, refine search by price, repeat purchases - are what make that possible. Commercial ideas: a Shopify app that brand owners install to become WebMCP-enabled (Shopify may also do it natively); Instamart-style stores; a Chrome extension layer; audit reports as a sales lead-in. Native mobile apps are out of scope - WebMCP is browser-only; a follow-up on iOS/Android app building was announced.

Benefits: speed (one call versus a snapshot loop), reliability (tools use the browser's own internal systems), fewer tokens, and control over how agents interact with your site. Because tools attach to application logic, the agent still finds the feature after a redesign; only renaming or removing parameters breaks it. Support: in recent Chrome you enable WebMCP support and the developer-tools flag under Chrome flags and relaunch; the feature is in beta; ChatGPT's browser has enabled it; Chrome, ChatGPT and others ran a challenge exploring what people build. Limits stated: no persistent login or session (tab must remain open), browser only (apps will not work), only useful if the site owner wants agents to visit, and the human is still in the loop. To check whether a site has it, ask your coding agent to inspect the site for WebMCP tools.

Harshit polls interest in a beginner iOS/Android app session (Saturday) and Duolingo-style character animations (Sunday), with a guest expert he has hired before; possible add-on on analytics and user behaviour. A cohort member asks about AI filmmaking: the described method - begin and end frames per scene, roughly 8-second generations, then joining and editing - is possible with video models and an editor, but is a field in itself, so he defers to a former student. On whether agentic AI will slow down after lab calls for restraint, he thinks not - the labs shipped new models straight after and cannot slow given the money invested; he suspects some doom messaging is attention marketing. He asks the cohort to bring monetization and distribution actions - first sign-ups, outreach counts, tactics - to Thursday's office hours. The CSAT poll is launched; resources will be on the LMS.