This seems backwards, somehow. Like you're asking for an nth view and an nth API, and services are being asked to provide accessibility bridges redundant with our extant offerings.
Sites are now expected duplicate effort by manually defining schemas for the same actions — like re-describing a button's purpose in JSON when it's already semantically marked up?
No, I don't think you're thinking about this right. It's more like hacker news would expose an MCP when you visit it that would present an alternative and parallel interface to the page, not "click button" tools.
You're both right. The page can expose MCP tools like via a form element which is as simple as adding an attribute to an existing form and completely aligns with existing semantic HTML - eg submitting an HN "comment". Additionally, the page can define additional tools in javascript that aren't in forms - eg YouTube could provide a transcript MCP defined in JS which fetches the video's transcript
Great to see people thinking about this. But it feels like a step on the road to something simpler.
For example, web accessibility has potential as a starting point for making actions automatable, with the advantage that the automatable things are visible to humans, so are less likely to drift / break over time.
In theory you could use a protocol like this, one where the tools are specified in the page, to build a human readable but structured dashboard of functionality.
I'm not sure if this is really all that much better than, say, a swagger API.
The js interface has the double edge of access to your cookies and such.
We're building an app that automatically generates machine/human readable JSON by parsing semantic HTML tags and then by using a reverse proxy we serve those instead of HTML to agents
As someone heavily involved in a11y testing and improvement, the status quo, for better or worse, is to do it the other way around. Most people use automated, LLM based tooling with Playwright to improve accessibility.
Hmmm... so are we imagining a future where every website has a vector to mainline prompt injection text directly from an otherwise benign looking web page?
This GitHub readme was helpful in understanding their motivation, cheers for sharing it.
> Integrating agents into it prevents fragmentation of their service and allows them to keep ownership of their interface, branding and connection with their users
Looking at the contrived examples given, I just don't see how they're achieving this. In fact it looks like creating MCP specific tools will achieve exactly the opposite. There will immediately be two ways to accomplish a thing and this will result in a drift over time as developers need to take into account two ways of interacting with a component on screen. There should be no difference, but there will be.
Having the LLM interpret and understand a page context would be much more in line with assistive technologies. It would require site owners to provide a more useful interface for people in need of assistance.
> Having the LLM interpret and understand a page context
The problem is fundamentally that it's difficult to create structured data that's easily presentable to both humans and machines. Consider: ARIA doesn't really help llms. What you're suggesting is much more in line with microformats and schema.org, both of which were essentially complete failures.
LLMs can already read web pages, just not efficiently. It's not an understanding problem, it's a usability problem. You can give a computer a schema and ask it to make valid API calls and it'll do a pretty decent job. You can't tell a blind person or their screen reader to do that. It's a different problem space entirely.
This is great. I'm all for agents calling structured tools on sites instead of poking at DOM/screenshots.
But no MCP server today has tools that appear on page load, change with every SPA route, and die when you close the tab. Client support for this would have to be tightly coupled to whatever is controlling the browser.
What they really built is a browser-native tool API borrowing MCP's shape. If calling it "MCP" is what gets web developers to start exposing structured tools for agents, I'll take it.
The web was initially meant to be browsed by desktop computers.
Then came mobile phones with their small screens and touch control which forced the web to adapt: responsive design.
Now it’s the turn of agents that need to see and interact with websites.
Sure you could keep on feeding them html/js and have them write logic to interact with the page, just like you can open a website in desktop mode and still navigate it: but it’s clunky.
Don’t stop at the name “MCP” that is debased: it’s much bigger than that
Skills are great for static stuff but they kinda fall apart when the agent needs to interact with live state. WebMCP actually fills a real gap there imo.
What prevents them with working with live state. Coding agents deal with the live state of source code evolving fine. So why can't they watch a web page or whatever update over time? This seems to be a micro optimization that requires explicit work from the site developer to make work. Long term I just don't see this taking off versus agents just using sites directly. A more long term viable feature would be a way to allow agents to scroll the page or hover over menus without the user's own view being affected.
I really like how the shell and regular API calls has basically wholesale replaced tools. Real life example of worse-is-better working in the real world.
Just give your AI agent a little linux VM to play around that it already knows how to use rather than some specialized protocol that has to predict everything an agent might want to do.
The next one would be to also decouple the visual part of a website from the data/interactions: Let the users tell their in-browser agent how to render - or even offer different views on the same data. (And possibly also WHAT to render: So your LLM could work as an in-website adblocker for example; Similar to browser extensions such as a LinkedIn/Facebook feed blocker)
I’m just personally really excited about building cli tools that are deployed with uvx. One line, instructions to add a skill, no faffing about with the mcp spec and server implementations. Feels like so much less dev friction.
I really like the way you can expose your schema through adding fields to a web form, that feels like a really nice extension and a great way to piggyback on your existing logic.
To me this seems much more promising than either needing an MCP server or the MCP Apps proposal.
Demo I built 5 months ago: https://www.youtube.com/watch?v=02O2OaNsLIk
This exposes ecommerce specific tool calls as regular javascript functions as it is more lightweight than going the MCP route.
It's great they are working on standardizing this so websites don't have to integrate with LLMs. The real opportunity seems to be able to automatically generate the tool calls / MCP schema by inspecting the website offline - I automated this using PLayright MCP.
You could get rid of the need for the browser completely just by publishing an OpenAPI spec for the API your frontend calls. Why introduce this and add a massive dependency on a browser with a JavaScript engine and all the security nightmares that comes with?
The problem with agents browsing the web, is that most interesting things on the web are either information or actions, and for mostly static information (resources that change on the scale of days) the format doesn't matter so MCP is pointless, and for actions, the owner of the system will likely want to run the MCP server as an external API... so this is cool but does not have room.
I disagree. I run a sudoku site. It’s completely static, and it gets a few tens of thousands of hits per day, as users only download the js bundle & a tiny html page. It costs me a rounding error on my monthly hosting to keep it running. To add an api or hosted mcp server to this app would massively overcomplicate it, double the hosting costs (at least), and create a needless attack surface.
But I’d happily add a little mcp server to it in js, if that means someone else can point their LLM at it and be taught how to play sudoku.
Very cool! I imagine it'll be possible to start a static webserver + WebMCP app then use browser as virtualization layer instead of npm/uvx.
The browser has tons of functionality baked in, everything from web workers to persistence.
This would also allow for interesting ways of authenticating/manipulating data from existing sites. Say I'm logged into image-website-x. I can then use the WebMCP to allow agents to interact with the images I've stored there. The WebMCP becomes a much more intuitive way than interpreting the DOM elements
I’m working on a DOM agent and I think MCP is overkill. You have a few “layers” you can imply by just executing some simple JS (eg: visible text, clickable surfaces, forms, etc). 90% of the time, the agent can imply the full functionality, except for the obvious edge cases (which trip up even humans): infinite scrolling, hijacking navigation, etc.
Question: Are you writing this under the assumption that the proposed WebMCP is for navigating websites? If so: It is not. From what I've gathered, this is an alternative to providing an MCP server.
Instead of letting the agent call a server (MCP), the agent downloads javascript and executes it itself (WebMCP).
So usually MCP tool calls a sequential and therefore waste a lot of tokens. There is some research from Antrophic (I think there was also some blog post from cloudflare) on how code sandboxes are actually a more efficient interface for llm agents because they are really good at writing code and combining multiple "calls" into one piece of code. Another data point is that code is more deterministic and reliable so you reduce the hallucination of llms.
What do the calls being sequential have to do with tokens? Do you just mean that the LLM has to think everytime they get a response (as opposed to being able to compose them)?
LLMs can use CLI interfaces to compose multiple tool calls, filter the outputs etc. instead of polluting their own context with a full response they know they won't care about. Command line access ends up being cleaner than the usual MCP-and-tool-calls workflow. It's not just Anthropic, the Moltbot folks found this to be the case too.
That makes sense! The only flaw here imo is that sometimes that thinking is useful. Sub-agents for tool calls imo make a nice sort of middle ground where they can both be flexible and save context. Maybe we need some tool call composing feature, a la io_uring :)
Do expose the accessibility tree of a website to llms? What do you do with websites that lack that? Some agents I saw use screenshots, but that seems also kind of wasteful. Something in-between would be interesting.
I actually do use cross-platform accessibility shenanigans, but for websites this is rarely as good as just doing like two passes on the DOM, it even figures out hard stuff like Google search (where ids/classes are mangled).
Cannot wait to be able to have a browser that show me the web as if it were a gopher website and i don't have to deal with ever changing to worse JavaScript heavy UX.
I've prepared a thoughtful reply saved to /Users/yoshikondo/HN_REPLY.md
HN Thread Link: https://news.ycombinator.com/item?id=47037501
Quick summary of my reply:
- Your 70+ MCP tools show exactly what WebMCP aims to solve
- Key insight: MCP for APIs vs MCP for consumer apps are different
- WebMCP makes sense for complex sites (Amazon, Booking.com)
- The "drift problem" is real - WebMCP should be source of truth
- Suggested embed pattern for in-page tools
This seems backwards, somehow. Like you're asking for an nth view and an nth API, and services are being asked to provide accessibility bridges redundant with our extant offerings.
Sites are now expected duplicate effort by manually defining schemas for the same actions — like re-describing a button's purpose in JSON when it's already semantically marked up?
No, I don't think you're thinking about this right. It's more like hacker news would expose an MCP when you visit it that would present an alternative and parallel interface to the page, not "click button" tools.
You're both right. The page can expose MCP tools like via a form element which is as simple as adding an attribute to an existing form and completely aligns with existing semantic HTML - eg submitting an HN "comment". Additionally, the page can define additional tools in javascript that aren't in forms - eg YouTube could provide a transcript MCP defined in JS which fetches the video's transcript
https://developer.chrome.com/blog/webmcp-epp
Great to see people thinking about this. But it feels like a step on the road to something simpler.
For example, web accessibility has potential as a starting point for making actions automatable, with the advantage that the automatable things are visible to humans, so are less likely to drift / break over time.
Any work happening in that space?
In theory you could use a protocol like this, one where the tools are specified in the page, to build a human readable but structured dashboard of functionality.
I'm not sure if this is really all that much better than, say, a swagger API. The js interface has the double edge of access to your cookies and such.
We're building an app that automatically generates machine/human readable JSON by parsing semantic HTML tags and then by using a reverse proxy we serve those instead of HTML to agents
As someone heavily involved in a11y testing and improvement, the status quo, for better or worse, is to do it the other way around. Most people use automated, LLM based tooling with Playwright to improve accessibility.
I certainly do - it’s wonderful that making your site accessible is a single prompt away!
Hmmm... so are we imagining a future where every website has a vector to mainline prompt injection text directly from an otherwise benign looking web page?
Most teams that want their data to be operated programmatically expose an API. For who does this solve a problem?
This was announced in early preview a few days ago by Chrome as well: https://developer.chrome.com/blog/webmcp-epp
I think that the github repo's README may be more useful: https://github.com/webmachinelearning/webmcp?tab=readme-ov-f...
Also, the prior implementations may be useful to look at: https://github.com/MiguelsPizza/WebMCP and https://github.com/jasonjmcghee/WebMCP
This GitHub readme was helpful in understanding their motivation, cheers for sharing it.
> Integrating agents into it prevents fragmentation of their service and allows them to keep ownership of their interface, branding and connection with their users
Looking at the contrived examples given, I just don't see how they're achieving this. In fact it looks like creating MCP specific tools will achieve exactly the opposite. There will immediately be two ways to accomplish a thing and this will result in a drift over time as developers need to take into account two ways of interacting with a component on screen. There should be no difference, but there will be.
Having the LLM interpret and understand a page context would be much more in line with assistive technologies. It would require site owners to provide a more useful interface for people in need of assistance.
> Having the LLM interpret and understand a page context
The problem is fundamentally that it's difficult to create structured data that's easily presentable to both humans and machines. Consider: ARIA doesn't really help llms. What you're suggesting is much more in line with microformats and schema.org, both of which were essentially complete failures.
LLMs can already read web pages, just not efficiently. It's not an understanding problem, it's a usability problem. You can give a computer a schema and ask it to make valid API calls and it'll do a pretty decent job. You can't tell a blind person or their screen reader to do that. It's a different problem space entirely.
This is great. I'm all for agents calling structured tools on sites instead of poking at DOM/screenshots.
But no MCP server today has tools that appear on page load, change with every SPA route, and die when you close the tab. Client support for this would have to be tightly coupled to whatever is controlling the browser.
What they really built is a browser-native tool API borrowing MCP's shape. If calling it "MCP" is what gets web developers to start exposing structured tools for agents, I'll take it.
Yeah, this seems like a weird niche where an agent has to interact with an existing browser session.
That, or they expect that MCP clients should also be running a headless Chrome to detect JS-only MCP endpoints.
The web was initially meant to be browsed by desktop computers.
Then came mobile phones with their small screens and touch control which forced the web to adapt: responsive design.
Now it’s the turn of agents that need to see and interact with websites.
Sure you could keep on feeding them html/js and have them write logic to interact with the page, just like you can open a website in desktop mode and still navigate it: but it’s clunky.
Don’t stop at the name “MCP” that is debased: it’s much bigger than that
This is coming late as skills have largely replaced MCP. Now your site can just host a SKILL.md to tell agents how to use the site.
The purpose of this appears to be for sites that cannot be controlled via prompt instructions alone.
I do like agent skills, but I’m really not convinced by the hype that they make MCP redundant.
seems like skill is a better interface, but state still needs to be externally managed, even if not using mcp as the protocol
Skills are great for static stuff but they kinda fall apart when the agent needs to interact with live state. WebMCP actually fills a real gap there imo.
What prevents them with working with live state. Coding agents deal with the live state of source code evolving fine. So why can't they watch a web page or whatever update over time? This seems to be a micro optimization that requires explicit work from the site developer to make work. Long term I just don't see this taking off versus agents just using sites directly. A more long term viable feature would be a way to allow agents to scroll the page or hover over menus without the user's own view being affected.
Interesting. I'd appreciate an example. Thanks!
check out https://moltbook.com/skill.md
I really like how the shell and regular API calls has basically wholesale replaced tools. Real life example of worse-is-better working in the real world.
Just give your AI agent a little linux VM to play around that it already knows how to use rather than some specialized protocol that has to predict everything an agent might want to do.
I think this is a good idea.
The next one would be to also decouple the visual part of a website from the data/interactions: Let the users tell their in-browser agent how to render - or even offer different views on the same data. (And possibly also WHAT to render: So your LLM could work as an in-website adblocker for example; Similar to browser extensions such as a LinkedIn/Facebook feed blocker)
Why would Facebook or LinkedIn ever give you this?
I’m just personally really excited about building cli tools that are deployed with uvx. One line, instructions to add a skill, no faffing about with the mcp spec and server implementations. Feels like so much less dev friction.
[dead]
Wes Bos has a pretty cool demo of this: https://www.youtube.com/watch?v=sOPhVSeimtI
I really like the way you can expose your schema through adding fields to a web form, that feels like a really nice extension and a great way to piggyback on your existing logic.
To me this seems much more promising than either needing an MCP server or the MCP Apps proposal.
Demo I built 5 months ago: https://www.youtube.com/watch?v=02O2OaNsLIk This exposes ecommerce specific tool calls as regular javascript functions as it is more lightweight than going the MCP route.
It's great they are working on standardizing this so websites don't have to integrate with LLMs. The real opportunity seems to be able to automatically generate the tool calls / MCP schema by inspecting the website offline - I automated this using PLayright MCP.
Have any sickos tried to point AI at SOAP APIs with WSDL definitions, yet?
Likely no.
Every generation needs its own acronyms and specifications. If a new one looks like an old one likely the old one was ahead of its time.
You could get rid of the need for the browser completely just by publishing an OpenAPI spec for the API your frontend calls. Why introduce this and add a massive dependency on a browser with a JavaScript engine and all the security nightmares that comes with?
The problem with agents browsing the web, is that most interesting things on the web are either information or actions, and for mostly static information (resources that change on the scale of days) the format doesn't matter so MCP is pointless, and for actions, the owner of the system will likely want to run the MCP server as an external API... so this is cool but does not have room.
I disagree. I run a sudoku site. It’s completely static, and it gets a few tens of thousands of hits per day, as users only download the js bundle & a tiny html page. It costs me a rounding error on my monthly hosting to keep it running. To add an api or hosted mcp server to this app would massively overcomplicate it, double the hosting costs (at least), and create a needless attack surface.
But I’d happily add a little mcp server to it in js, if that means someone else can point their LLM at it and be taught how to play sudoku.
Very cool! I imagine it'll be possible to start a static webserver + WebMCP app then use browser as virtualization layer instead of npm/uvx.
The browser has tons of functionality baked in, everything from web workers to persistence.
This would also allow for interesting ways of authenticating/manipulating data from existing sites. Say I'm logged into image-website-x. I can then use the WebMCP to allow agents to interact with the images I've stored there. The WebMCP becomes a much more intuitive way than interpreting the DOM elements
I’m working on a DOM agent and I think MCP is overkill. You have a few “layers” you can imply by just executing some simple JS (eg: visible text, clickable surfaces, forms, etc). 90% of the time, the agent can imply the full functionality, except for the obvious edge cases (which trip up even humans): infinite scrolling, hijacking navigation, etc.
Question: Are you writing this under the assumption that the proposed WebMCP is for navigating websites? If so: It is not. From what I've gathered, this is an alternative to providing an MCP server.
Instead of letting the agent call a server (MCP), the agent downloads javascript and executes it itself (WebMCP).
In what world is this simpler than just giving the agent a list of functions it can call?
So usually MCP tool calls a sequential and therefore waste a lot of tokens. There is some research from Antrophic (I think there was also some blog post from cloudflare) on how code sandboxes are actually a more efficient interface for llm agents because they are really good at writing code and combining multiple "calls" into one piece of code. Another data point is that code is more deterministic and reliable so you reduce the hallucination of llms.
What do the calls being sequential have to do with tokens? Do you just mean that the LLM has to think everytime they get a response (as opposed to being able to compose them)?
LLMs can use CLI interfaces to compose multiple tool calls, filter the outputs etc. instead of polluting their own context with a full response they know they won't care about. Command line access ends up being cleaner than the usual MCP-and-tool-calls workflow. It's not just Anthropic, the Moltbot folks found this to be the case too.
That makes sense! The only flaw here imo is that sometimes that thinking is useful. Sub-agents for tool calls imo make a nice sort of middle ground where they can both be flexible and save context. Maybe we need some tool call composing feature, a la io_uring :)
Who implements those functions? E.g., store.order has to have its logic somewhere.
Do expose the accessibility tree of a website to llms? What do you do with websites that lack that? Some agents I saw use screenshots, but that seems also kind of wasteful. Something in-between would be interesting.
I actually do use cross-platform accessibility shenanigans, but for websites this is rarely as good as just doing like two passes on the DOM, it even figures out hard stuff like Google search (where ids/classes are mangled).
Finally, I was hoping for this to be implemented in 2026. Rendered DOM is for humans, not for agents.
What problem does this solve?
MCP is cool, but it's too open ended security wise.
People should be mindful of using magic that has no protection of their data and then discover it's too late.
That's not a gap in the technology, it's just early.
Cannot wait to be able to have a browser that show me the web as if it were a gopher website and i don't have to deal with ever changing to worse JavaScript heavy UX.
This is true excitement. I am not being ironic.
Now we just need a proxy server that automatically turns any API with published openapi spec into a WebMCP server, and we've completed the loop
I'm building this. Initially it was to do codegen for tools/sdks/docs but will incorporate webmcp as part of it.
I wanted to make FOSS codegen that was not locked behind paywalls + had wasm plugins to extend it.
I've prepared a thoughtful reply saved to /Users/yoshikondo/HN_REPLY.md