Skip to main content

How We Shipped Live WebMCP Tools on a Production Next.js Site

How Ryzolv added three read-only WebMCP tools to Next.js, preserved unsupported browsers and kept enquiry submission with the person reviewing the form.

By Anshul KaushikManaging Director7 min read

Published Updated

We shipped three WebMCP tools on ryzolv.com: find a service, find a public example, and get an enquiry link. Each helps a browser agent answer a visitor's question using information we already publish. The enquiry tool returns a form link with the relevant practice selected. The visitor still needs to review and submit the form.

That was the scope we wanted for a first release. A visitor researching RAG should reach our Knowledge Systems page. An agent looking at a demonstration should receive its correct label. Neither task requires permission to create a lead or change a business record.

Our site uses Next.js App Router, React and TypeScript. Adding an experimental browser interface meant handling support detection and navigation lifecycles while preserving the ordinary website.

Keep the website independent of WebMCP

WebMCP is a proposed browser standard for exposing structured tools to agents. A page registers a named operation with a description, input schema and execution function. A supporting agent can discover that contract and call the operation.

It gave us a way to describe useful tasks explicitly. The existing pages, headings and links still do their normal work. A browser without WebMCP can browse the services, read examples and follow the contact path.

We placed registration in a small client component. Browser globals are accessed inside a React effect, after client mounting, so server rendering never needs a document or navigator object. The component checks for the browser feature before dynamically importing the registration code.

The adapter prefers document.modelContext when it provides registerTool. It also recognises the earlier experimental navigator.modelContext interface when that interface provides both registration and explicit unregistration. Keeping those differences in one adapter limits the code affected by an API revision. Chrome's imperative API documentation describes the current document interface and registration lifecycle.

Unsupported browsers skip registration. An import or registration failure leaves ordinary navigation intact. We kept the feature optional all the way through execution; a visitor never has to enable an experimental browser feature to contact us.

Treat registration as a React lifecycle

Client navigation can change the page without creating a new document. A registration that survives indiscriminately can leave tools available in the wrong context or cause duplicate registrations on remount.

Our registration effect follows the current pathname. Its cleanup disposes of the registrations it owns before the next registration set takes over. The current adapter uses an AbortController for registration lifetime; the legacy path also supports explicit unregistration. We avoid replacing the document's entire tool registry because another component could own entries there.

We also handle registration completing after cleanup has started. A late completion must not revive a disposed registration. If registration fails partway through, we clean up the tools already registered by that attempt.

These details are easy to miss when the first test consists of loading one page. We tested remounts and asynchronous cancellation separately, then checked actual public navigation through the browser. Private editorial pages are outside the public tool surface.

Give the agent a small public catalogue

The tools operate on a deliberate projection of existing public content. Each result includes a title, description, practice, content type and canonical URL. We select those fields explicitly rather than passing a broader internal content object into the browser interface.

The tool set is short:

Swipe or scroll horizontally to compare every column.

ToolWhat it returns
ryzolv_find_servicesPublic services, capabilities and industry applications matching the query.
ryzolv_find_public_examplesPublic examples with their case-study, demonstration or security-service labels.
ryzolv_get_enquiry_linkThe existing enquiry form URL with a selected practice.

Search uses deterministic keyword matching over that catalogue. It does not call a model or decide whether we are suitable for an engagement. Results point back to the public page where the visitor can inspect the description.

The example label matters. A demonstration illustrates a workflow; it does not establish that we delivered it for a client. Returning the type alongside the description gives the agent evidence for that distinction. We retain the same distinction in the visible source content.

We derive JSON Schemas from the validation definitions and validate inputs again inside execution. Outputs are validated before being returned as structured values. That keeps the advertised contract and handler behaviour connected, even when the caller supplies something unexpected.

Stop the enquiry tool before submission

An agent can help someone find the relevant enquiry form without receiving authority to send an enquiry on their behalf.

Our tool accepts a practice selection. It uses the same link-building helper as the website's existing enquiry path and returns that URL. It does not open the destination, fill personal details or submit the form. Its response explicitly says that no enquiry has been submitted and no appointment has been booked.

Calling this a pre-filled form needs precision: the practice is preselected. The tool is not preparing a complete lead from details inferred during a conversation. The person opening the link can check the choice and decide what to share.

All three tools carry readOnlyHint: true and consequentialHint: false. Those annotations describe the handlers. They do not enforce permissions or substitute for validation. The read-only property comes from what the code can do: search public entries or construct an existing link. It performs no submission or persistence operation.

Read-only access also needs a data boundary. A tool that returns private customer information could still cause harm without changing a record. We limited this release to selected public information. Chrome's tool-security guidance discusses both annotation hints and the risks of exposing read-only data.

Returning a review link preserves the human handoff within this tool's workflow. It does not claim to prevent an independently controlled browser agent from interacting with a public form by other means.

What the browser evidence shows

Our recorded native browser test discovered all three tools with their real input schemas. Searching for RAG returned Knowledge Systems. Searching public examples for insurance returned Carrier Appetite Search with kind: "demonstration". Those calls exercised registered handlers through the browser, separately from the simulated API tests used for lifecycle failures.

The enquiry schema exposed a required practice choice rather than contact fields. A fresh Lighthouse capture of the production page confirmed that contract. This excerpt retains its validation rules and omits the schema dialect and descriptive text:

{
  "type": "object",
  "properties": {
    "practice": {
      "type": "string",
      "enum": [
        "AI Engineering",
        "Ryzolv Automate",
        "Not sure"
      ]
    }
  },
  "required": ["practice"],
  "additionalProperties": false
}

HTML scraping gives an agent page text and interface elements from which to infer the next step. Here, the registered schema states the accepted choice, and the handler returns a structured result. The public source page remains available for checking what the result means. This establishes the interface we exposed; it does not measure a speed or accuracy improvement over scraping.

We ran Lighthouse 13.6.0 against the live AI Engineering page on 16 November 2026, using Chrome 157 with WebMCP enabled in an isolated test browser. Its Agentic Browsing report listed these imperative tools:

ryzolv_find_services
ryzolv_find_public_examples
ryzolv_get_enquiry_link

Registered tools (informative): 3
Schema validity: pass
Form coverage: not applicable

This is a condensed transcription of the report, not a separate test result. The report includes each tool's description and input schema. Form coverage was not applicable on that page; this run does not audit the hosted enquiry form.

In a control run where WebMCP was unavailable, those checks were not applicable. An unavailable check supplies no evidence that registration or schemas passed. Chrome documents these support requirements in its Agentic Browsing audit guidance.

The successful schema check confirms what the browser observed in that configuration. It does not certify the application as secure or establish compatibility with every agent.

Verify behaviour beyond discovery

A tool appearing in a registry proves only that it was registered. We also checked returned results and the enquiry handoff. The recorded form checks confirmed each supported practice selection while stopping before personal data entry or submission. A separate production browser call confirmed the live enquiry-link response after deployment.

Our automated tests cover validation and registration cleanup. Native browser calls provide different evidence: the tool contract was discoverable and executable in the tested environment. Neither establishes support across every browser or agent product.

In this release, an agent can find public information, preserve the distinction between a demonstration and client work, and return the relevant enquiry link. The person reviewing the form decides whether to send it.

Written by Anshul Kaushik, Managing Director at Ryzolv.

Back to Insights

Have a project in mind?

Discuss Your Project

    Google Analytics helps us measure page visits and site interactions. It is optional. You can use the site and contact us with analytics switched off.

    Analytics is off.

    We remember your choice for six months in this browser. Essential storage keeps that preference.

    Read our privacy policy