DataStrict joins the first wave of Google WebMCP

A new kind of visitor is arriving on the web: the AI agent. Built into browsers like Chrome, it can read a page and act on it - run a search, fill in a form, complete a task - on a person's behalf. Until now, that agent had to guess its way through interfaces designed for human eyes, simulating clicks and reading the screen. WebMCP, a new web standard from Google, changes the contract: a site can declare what it can do, so an agent uses it directly and reliably. DataStrict is among the first wave of sites to adopt it, through Google's origin trial in Chrome 149 to 156.

Here is what that means if you never think about browser internals. Soon you will be able to tell your browser's assistant 'book a demo with DataStrict' or 'find their page on MCP gateways,' and it will simply work - because our site now speaks the agent's language instead of leaving the agent to fumble through the layout. The page does the same thing it always did for a person; it just also offers a clean path for the machine.

There is one line we deliberately did not cross, and it is the most DataStrict part of this. Our agent tools pre-fill the demo and contact forms and bring them to you - but a human always reviews and clicks submit. No agent can send a request, generate a lead, or take any action on its own. That restraint is not a limitation we forgot to remove; it is the whole company in miniature. We build the cutting-edge control layer that sits between AI and the actions it can take, and we hold this site to the same bar: agentic, but governed. An agent that can do anything, unsupervised, is exactly the problem we exist to solve.

For the curious, here is how it works. WebMCP is a proposed web standard, currently incubated in a W3C community group and shipping in Chrome behind an origin trial. A page registers 'tools' - named capabilities, each with a structured input schema and a function the browser's agent can call. Instead of scraping pixels and guessing which button does what, the agent reads the declared tools and invokes them against the site's own logic. It applies the same tool-calling model as the Model Context Protocol - named capabilities with typed inputs and explicit results - but binds those tools to the page in the browser, rather than connecting a model to an external server.

Our reason for moving early is narrow and on-brand: agents are beginning to take real actions on a person's behalf, and governing those actions is exactly what DataStrict does. The sooner a site can declare what an agent may do, the sooner that interaction can be governed instead of guessed at. WebMCP is still experimental and will keep changing, so we treat it as cutting-edge work in progress rather than a settled feature - and the problem it points at, agents acting in the world, is the problem we already build for. We would rather help shape how that interaction handles trust and control than adapt to it after the fact.

For the engineers reading this, the practical version is short. WebMCP is a browser API gated behind a Chrome origin trial: you register a tool with a name, a JSON input schema, and an execute function, and a participating agent can discover and call it. Two things we learned shipping it that the docs do not stress enough. First, register your tools outside your framework's render cycle and behind a try/catch - an experimental API that throws synchronously can otherwise take down your page. Second, make any tool that changes state fill-only: pre-fill, then hand control back to the user, and never auto-submit. The convenience of an agent does not justify letting it take irreversible actions unsupervised - and for anything sensitive, that is not a nicety, it is a requirement.

Governance is becoming infrastructure, and that now includes agents. When your AI can take actions, the real question is not whether it is capable - it is whether it is allowed, and whether you can prove what it did. That is the layer we build. WebMCP is a small milestone on the way there, and an honest preview of how we think the agentic web should work: genuinely useful for the user, and always under their control.