Make Your Web App Agent-Ready with WebMCP
Your customers are arriving with an agent. Since 26 August 2026, ChatGPT can call the WebMCP tools your web app declares, instead of screenshotting it and guessing where to click, and Chrome and Edge are running the same standard in origin trials. Two weeks and a fixed price gets your product from something an agent fumbles through to something it operates properly, with you deciding what it is allowed to touch.
Technology Partners
Recognized by Clutch
Agent Readiness Sprint
Two weeks. Your product's actions shipped as agent-callable tools, tested against a real agent. Credited in full against a larger build if you engage us within 90 days.
What WebMCP Changes
Until recently an AI agent used your product the way a very determined stranger would. It took a screenshot, read the DOM, guessed which element was the submit button, and clicked. It worked often enough to be interesting and failed often enough to be useless for anything that mattered. You had no say in any of it.
WebMCP replaces the guessing with a declaration. Your page registers a list of tools, each with a name, a description written for a model, and a schema for its arguments, and the agent calls them. The tools run in the browser, in your visitor's own session, calling the same client-side code your buttons already call. Your page can be thought of as an in-page MCP server, except that the user is watching it happen in your interface rather than reading a summary of it somewhere else.
That last part is the reason this proposal exists, and it is the part product leaders should care about. The alternative route, where an agent platform integrates with your backend API, means the agent talks to your servers and your customer never sees your product. The specification names that outcome directly and says its goal is to prevent it, by adapting front ends for agents rather than replacing them. If you think your interface is worth something, this is the standard that keeps it in the transaction.
What You Get in Two Weeks
Six things, the same six every time. The price is fixed, so the shape is fixed.
A tool map of your product
Every action a user can take today, sorted into three groups: what becomes an agent-callable tool, what stays behind a human click, and what should not be reachable by an agent at all. The third group is the one that matters, and it is a product decision rather than an engineering one, so you make it with us in the room.
The tools, registered and working
Typically ten to twenty, each with a name, a description written for a model to choose from, and a JSON Schema for its arguments. They call the client-side code your UI already calls, so there is no second implementation of your product to keep in sync and no backend to stand up.
The exposure boundary, set deliberately
Which origins can discover which tools, what the Permissions Policy allows, and what an embedded third-party widget in your page can and cannot see. Tools run in your visitor's authenticated session, so this is the part that decides whether a prompt-injected agent can do damage in their name.
A confirmation model for anything that costs
Anything that spends money, sends a message, or cannot be undone gets a human in front of it. The pattern we use is that the agent prepares the action and fills in the interface, and your user is the one who confirms. That is also what the specification is designed for.
Proof that an agent finishes the job
A set of real tasks, run end to end against ChatGPT Desktop and against Chrome with the origin trial enabled, with pass and fail recorded per task. A tool an agent picks correctly nineteen times in twenty is a different product from one it picks correctly half the time, and the only way to know which you have is to run it.
A written handover
What we exposed, what we deliberately did not, and where the specification is still moving so your team knows what will need revisiting. Written to be read by your engineers after we have gone, not as a sales artifact.
This Site Runs WebMCP
The page you are reading registers WebMCP tools. Open vindler.solutions in ChatGPT Desktop's built-in browser, or in Chrome with the WebMCP flag enabled, and ask it to find our work in your industry or to write up your project for us. It will call these:
list-servicesandsearch-articlesread what we do and what we have written about it.list-case-studiesfilters our client work by industry or by technology.get-agent-readiness-offerreturns this engagement as structured data, and exists only while you are on this page.draft-project-enquirywrites your enquiry into our contact form and puts it in front of you. It does not send it. You press Send.
That last one is the whole design argument in one tool. An agent that could post directly to our contact endpoint would be an agent that emails us in a stranger's name the first time it reads an instruction planted on some other page. So it fills in the form and stops. Every engagement we run applies the same rule to anything that spends money, sends a message, or cannot be taken back.
What Is Actually True Today
One mainstream agent client calls these tools: ChatGPT Desktop, since 26 August 2026. Chrome and Edge are running origin trials, which are time limited by design. Firefox and Safari have open standards positions and no commitment. The specification has real open questions, including how tools handle binary media, how a tool responds when it navigates the page, and whether the declarative form syntax lands in the shape it is currently drafted in.
So this is an early bet, and the case for taking it now is not that WebMCP has won. It is that the work is small, the same tool map is what you would write for any agent interface you build later, and the alternative to declaring your tools is not privacy. It is agents clicking around your product without supervision. If that argument does not hold for your product, we will say so on the call rather than sell you two weeks.
What We Need From You
Four things, all on day one. The two weeks start when the last of them arrives, not when you sign, because chasing repository access midway through is what makes a fixed timeline slip.
- Read access to the front-end repository
- A 45 minute walkthrough where you drive the product and we watch
- A working environment we can point an agent at, staging is fine
- Your answer to one question: which ten things do users actually come here to do
Questions
Start With a Call
Twenty minutes, no charge. Show us the product, and we will tell you which of its actions are worth exposing, which are not, and whether two weeks is the right size for it.
Or Send Us the Details
Tell us what your product does and who uses it. We reply within 24 hours.