First a brief explanation – MCP stands for model context protocol, and is an open standard to let AI connect to external systems, simplified. WebMCP is similar, but it’s promise (it’s not a standard yet) is to work directly in the web browser context – allowing AI agents to “talk” to the webpage or web application.
What’s the point? Well, if it will be accepted, we will be able to use AI agents / AI chat-bots and similar in-browser AI powered technology to communicate with a website that will use WebMCP tools. Website owners can provide WebMCP tool for all functionality that is already on the page and extend it even more. So – we have a choice – use the website as we have been using for decades – or use the exposed tools directly, via our AI model (if we want to).
Sure, we already have the technology for AI to “talk to” and control websites, but it involves a lot of guesswork – and when accessibility is poor it fails more often – as it relies on the semantics, just like assistive technologies – and computer vision working out of website screenshots is also a guessing game more often than not.
So – WebMCP helps the AI agent with predefined tool possibilities that webpage owners implemented to make people more efficient, but without AI needing to guess.
To make this more concrete – imagine an e-commerce page for clothing and shoes with a filter and search pages offering abundance of possibilities – we can go through all of the possibilities manually and it often takes some time. If we are not familiar with the website (or they changed it), it takes some time to find all the possibilities and then also use them. With WebMCP, when available, our AI agent would get the possibilities and let us know what we can filter by, then we can just provide our criteria and AI agent will then send them to the webpage and webpage will trigger it’s search/filtering based on our criteria. Making the path potentially way faster.
Future beyond auto-fill
Now, I will look into (near?) future for a moment and I think it is actually very possible (some are already testing this), but imagine that the AI agent would know you better. A secure, private, local and affordable (open source) AI agent that would know your shoe and clothing sizes, color preferences, existing wardrobe and color combinations in it, age of existing wardrobe articles, just to limit myself to an online clothes shop context. That could mean that your AI agent could be very helpful with the filtering, as it would know you better and had the possibilities to suggest personalized filter/search criteria based on your wishes and your situation. Please let your imagination fill in the gaps where such context would be safe to use and making you way more efficient and minimizing manual data inputs.
We all know HTML autocomplete attributes, allowing us to auto-fill known form fields (when we allow browser to), like name, family name, address, email, and even credit card and so on. With this WebMCP, in combination with AI agent that you can trust – we can see how this would make everybody using AI more efficient
Well – if webpage, AI agent interface and the WebMCP are accessible, that is. Here I reach the wishful thinking made in the title of this post.
Human in the loop needs accessibility of the whole loop
I can not see a scenario where AI will be fully automated at all times. Perhaps I am wrong, time will show, but personally I want control. Some claim that AI agents don’t need user interfaces. But I am not very comfortable with this idea for multiple reasons. So I believe that “human in the loop” – human that can take control at any time, that can check everything is and will be essential. Which means that human needs to be able to use both the website itself, the AI agent interface and that developers need to consider the WebMCP interface facing the user to be accessible as well.
Only with all parts accessible humans can really be in the loop.
I do not see WebMCP as a magical silver bullet where we don’t need the semantics and accessibility of the underlying webpage. Not all people want or can use AI agents. And people that want and use AI agents still need to be able to use the webpage if they wish so.
WebMCP is not a replacement for accessibility. I treat it as an additional tool that can make people that want to use AI more efficient – but once again I need to say that all moving parts need to be accessible.
If, for example we use WebMCP on an inaccessible webpage – how can then user get the needed feedback from the webpage? How can a user check the rest of the webpage manually? Some may say that WebMCP can cover all functionality of a webpage, which is technically possible, yes. But as a webpage owner, I do not want my WebMCP to make my webpage obsolete. I want to offer additional information to the user. I want the user to feel my brand. I want the search engines to find my page, I do not want to replace my brand with a tool for AI.
Standards in making need co-design or they may exclude
I see a giant potential in WebMCP, but standardization needs to be discussed not only by developers and marketers, but also by people with disabilities and accessibility specialists. And there are a couple of things I can suggest…
Communication channels
GitHub is the default and offers an open forum indeed, but it can be a barrier for some people and is mostly known to developers. Initiators could do better here, with more information on how to contribute and how to test.
Perhaps that is just behind the corner or it is just my searching that is not optimal.
How to actually test can introduce many barriers
Testing of the standards proposal is essential if we want proper feedback.
I don’t think it is an issue for developers, but if we want diverse user base it can be an extremely large barrier.
Users need to be quite technical to use the browser innovative feature flags, then find the example apps, get the AI agent security key and free tier is quite limited (so they need to pay for it if they do not have it from before), use it in the testing browser extension and then operate the whole thing.
A lot of complexity for a non-technical user.
And some examples are not very accessible, therefore potentially difficult to use.
There are descriptions on how to make this work, but two major barriers still exist:
- Some demo examples are not very accessible. How can we then have people with disabilities to really test them. I have not done full audits, but there are issues in them that we do not need a lot of time to notice. I would like the initiators to have more focus on providing accessible examples. Inaccessible example can exclude vital feedback!
- Extension requires online AI model that requires registration and has a free tier that is practically unusable for any real testing. This means we need to pay to test. I am not aware of any alternatives, but I suggest initiators to provide some alternatives, so that testing would not cost money (perhaps free tier could have an allow list and more generous free tokens when requests come from well known examples at least).
Welcoming feedback from everyone is super, but enabling more people to have a chance to provide real feedback require more from initiators. Once again, I write about this from my standpoint and perhaps (I hope) these problems are already worked on.
(Testing) resources
- GitHub homepage webMachineLearning / WebMCP (opens in new window), allows for discussion (please check issues pane) and more context.
- Google Doc WebMCP Early Preview (opens in new window), has a lot of context, and under Setup heading you can find how to actually test it.
- Code for some demos on GitHub (opens in new window), here we can see the source code and also open the demos in the browser.
To conclude – WebMCP is not a replacement for accessible webpages
One of the goals of WebMCP mentions accessibility:
Improve accessibility: Provide a standardized way for assistive technologies to access web application functionality beyond what’s available through traditional accessibility trees which are not widely implemented.
WebMCP goal that mentions accessibility (opens in new window)
This goal can be interpreted in many ways and some of the interpretations can be understood that WebMCP can be an alternative to the webpage itself. This is extremely dangerous if it allows for underlying webpages to be inaccessible and only WebMCP interaction that would be accessible.
If we have an inaccessible webpage that has a WebMCP – chances are enormous that the WebMCP experience will also be inaccessible, even if the website owner believes that they are accessible. And copying all the webpage functionality inside WebMCP is just introducing a lot of unneeded work that could be used to make the underlying webpage more accessible instead.
Another giant issue is – it can mean that only users of AI agents can get accessibility, while others don’t.
Sure, some may argue, that website owner will also implement the AI agent, so that all users can use it. Well, in that case – we can’t really benefit from the full potential of the WebMCP – as then we don’t have the private context that makes us most efficient (auto-filling and context-aware suggestions).
And I also strongly recommend against “conforming alternate version” (opens in new window), which WebMCP could bring – as it is against the universal design approach we really should offer to our users.
Humans needs the website to be accessible, and if they choose to use AI agent and WebMCP tools, so must these be accessible, or they are robbed of full control and transparency (and website owners of all their tangential and branding context).
