Re: [w3ctag/design-reviews] Incubation: WebMCP (Issue #1238)

christianliebel left a comment (w3ctag/design-reviews#1238)

@domfarolino Thanks for your response and the update. Our review won’t be finished this week, so we are able to take Mozilla’s updated position and the TPAC discussions into account. In the meantime, we want to share our current thinking with you. **This is not a resolution comment, and it does not replace the review that we will continue working on.**

Part of the reason why our review is taking longer than usual is that WebMCP is linked to Agentic Browsing, which has web architecture implications of its own (security, privacy, identity delegation). Therefore, we plan to publish a Finding with the concerns specifically related to Agentic Browsing, which are relevant for WebMCP as well.

Agentic Browsing, at least with the current generation of models, can be slow, brittle, and consume excessive tokens. WebMCP addresses these problems by providing a dynamic list of tools, bridging the gap between the unstructured, natural-language, non-deterministic world of LLMs and the structured, programmatic, deterministic world of web applications.

However, TAG members have expressed a collection of concerns, which ought to be addressed before WebMCP ships. The TAG hasn’t reached consensus on how each of these points affects WebMCP’s impact on the web’s architecture.

The TAG is concerned that WebMCP creates an agent-only description of a page's functionality that might be unreachable by human users or screen readers, and might duplicate content. One unintended consequence could be a "web for agents", where websites can no longer be used by humans through the usual input methods. Users might not know whether, and which, tools are available without asking the agent.

Some TAG members are concerned that WebMCP removes the incentive for authors to provide a semantically rich, accessible website, which would also aid agents. Specifically, they ask to fully work out the declarative version before shipping the imperative one.

WebMCP doesn’t completely remove the brittleness associated with LLMs: success depends on the model driving the agent, the schema, and the natural-language descriptions. Especially when dealing with a smaller/local LLM, the model needs to support the language the tool descriptions are written in. The imperative tools only carry the page’s overall language and direction information with no current ability to change this, while the declarative ones carry their individual node’s language and direction.

Some TAG members are concerned that this (like the Prompt API) would be the first time that JSON Schema is used by a web platform API. They’re also concerned that the W3C doesn’t have an existing relationship with its SDO.

Finally, we’re concerned that the name WebMCP implies a relationship to MCP, but the proposal only covers tool registration and is otherwise unrelated to this particular protocol. Is there time to rename the proposal to something like “Tools”?

All in all, while we see the value of WebMCP, we believe the points above will need more consideration as this proposal evolves.

-- 
Reply to this email directly or view it on GitHub:
https://github.com/w3ctag/design-reviews/issues/1238#issuecomment-6410471845
You are receiving this because you are subscribed to this thread.

Message ID: <w3ctag/design-reviews/issues/1238/6410471845@github.com>