Article

We made umage.ai agent-ready

ChatGPT can now call functions on web pages through the WebMCP draft. We put two of them on umage.ai — one lists our open roles, the other sends us a message — and here is what it took, what it costs, and why we did it anyway.

The moment that sold me was unremarkable, which is the point. While testing the change this post describes, I had our contact page open in the browser the ChatGPT desktop app was driving, and I asked the app: What open roles does umage.ai have? The app called a function our page had handed it, got a JSON answer, and read me the list.

No clicking. No typing into a form. No screenshot of a page it was trying to interpret. The page had said here is what I can do, and here is how to ask, and the agent had used it.

That sentence — a page telling an agent what it can do — is the WebMCP draft. This post is about what it took us to adopt it, what it costs, and what it is actually for.

What WebMCP is, in two paragraphs

WebMCP (the Web Machine Learning Community Group, run by Google and Microsoft) is a draft that lets a web page register its own JavaScript functions as tools: a name, a natural-language description, and a structured input schema. Agent-capable browsers — today, Chrome’s origin trial, and ChatGPT Desktop, which has called them on by default since late August — discover those tools and call them the way a person would use a form, except without simulating a person.

The API itself is one line deep: document.modelContext.registerTool({ name, description, inputSchema, execute }). Everything else is ordinary JavaScript running on your own page, with your own data. Nothing is sent to a server you don’t control; the agent calls a function, the function does what it does, and the browser returns the result.

What we exposed: two tools, about 120 lines

We resisted making this a showcase. The right size is two tools:

list_open_roles — read-only. Returns the open roles: title, summary, workplace, employment type, location, and the URL of each posting. The data already lived in the site’s content; the tool just hands it over.

send_message — the interesting one. The same name / email / company / topic / message payload as our contact form, POSTed to the same trigger we already run. No new backend, no new endpoint, no new server code path. When an agent calls it, the message lands in the same queue as a typed one, and the one-business-day clock starts the same way.

Both live in a single small script that ships only on the contact and careers pages. In every browser without the API — which, outside the trial, is still all of them — the script checks one property and exits after three lines. Human visitors see nothing new.

Why we did it, honestly

It is not a traffic play. The trial is Chrome-only (currently running to 30 March 2027), Safari has publicly declined, and nobody — including us — can yet tell you how many agent-initiated contacts are worth in raw visits.

We did it for three reasons, all of which we give to clients:

  1. Proof over claims. We sell agentic development. Your site can be agent-ready is a claim until it can be checked. A page that hands ChatGPT our open roles and takes a message on our behalf is the same proof-not-claim move as our Studio work: the reader is invited to verify, not to believe.
  2. The window is the point. ChatGPT has called these tools on by default since 25 August. Gemini is expected to follow in 2027. The cost of doing this now is one header line on the server and one small script; the cost of waiting is arriving late to the pattern that defines the next version of the web. First movers get the case study; latecomers get the retrospective.
  3. It is reversible. If the draft settles differently, or the trial never graduates, we take the script out and remove the header. Nothing else about the site changes.

Try it yourself

  • Install the ChatGPT desktop app (macOS or Windows) and sign in.
  • Have https://umage.ai/contact/ open in the browser the app drives.
  • Ask: What open roles does umage.ai have? — it calls list_open_roles and reads back the list, with links to each posting.
  • If you actually want one of them: ask it to draft and send a message. It calls send_message, the reply lands with us, and the one-business-day clock starts exactly as it would for anyone else.

What this means for your site

The part worth copying is not the tools — it is the pattern. Find the one or two actions a visitor, or an agent acting for a visitor, most needs — contact, search, status, checkout — and expose them as tools with honest descriptions and structured inputs. On a marketing site like ours that was an afternoon of work, most of it spent reading the draft. On a product, it is the same shape at a different scale: a tool is a thin wrapper around code you already have.

If you want us to map which of your site’s actions are worth exposing, that is a conversation we are happy to start. On the same contact page, by the way — which now accepts both of you.