updates on WebMCP specs. PRs for discussion

AI KR CG and adjacent groups

I note the difficulty *for everyone in keeping up with development in
agentic AI

The main mission for  W3C AI KR CG so far has been to be able to support
explainability, cognition and reasoning of whatever is taking place that
drives

advancements in the space,


We are now trying to figure out* MCP and WebMCP*  *I personally do not like
github at all, just another layer of headache to the stack

but here some readers, for those who like to keep up and chip in their
respective CG/WG/IG

- https://starborn.github.io/webmcp/ (Model Card Generator)
-
- https://starborn.github.io/webmcp/webmcp-complete-guide.html (Guide)
- https://starborn.github.io/webmcp/webmcp-quiz.html (Quiz)
-
- https://starborn.github.io/webmcp/whatotest.md
- https://starborn.github.io/webmcp/browserhack.md
- https://starborn.github.io/webmcp/WebMCnotMCP.md

and PRs to the main webmcp spec repo for discussion

https://github.com/webmachinelearning/webmcp/issues/9




https://github.com/webmachinelearning/webmcp/issues/100



*Subject: WebMCP spec development at WebML CG -- relevance for AI KR CG and
related groups*

The W3C Web Machine Learning Community Group has been developing WebMCP, a
browser-native API for exposing web page tools to AI agents via the Model
Context Protocol. Yesterday's meeting (19 Feb 2026) marked a significant
milestone: an early implementation is now available behind a flag in Chrome
Canary 146, and the spec is actively being shaped by browser vendors
including Google and Mozilla.

For those not yet tracking this work, WebMCP is essentially standardizing
how AI agents discover and invoke capabilities exposed by web pages --
making MCP a first-class browser primitive. The implications for agent
interoperability, trust architecture, and knowledge representation are
substantial, and the window for influencing the spec's foundational
decisions is open right now.

Several issues on the table are directly relevant to our communities:

The group is actively debating agent identity -- how tools know which agent
is calling and with what authority. This is currently framed as an
authentication problem, but it has an equally important knowledge
representation dimension: what vocabulary do tools and agents share to
declare and interpret capability scopes? Without coordination at the
semantic layer, permission systems will fragment by publisher.

A related gap concerns tool capability representation. The spec needs a way
for tools to declare not just what they are called but what they do, what
they assume, and what their side effects are -- the kind of invocable
contract that makes user consent meaningful and audit trails actionable.
This is precisely the territory where AI KR, WebAgents, and related CGs
have relevant expertise.

There is also an open resolution from yesterday's meeting to document use
cases and requirements for an "agent allowlist" -- enterprise-controlled
access policies for which agents can invoke which tools. This is a trust
and policy problem with significant KR dimensions.

I am engaging directly in the WebMCP GitHub repository and would welcome
interest from CG members in doing the same. The repository is at
https://github.com/webmachinelearning/webmcp and the meeting minutes are at
https://www.w3.org/2026/02/19-webmachinelearning-minutes.html.

 I can propose a liaison item for our next CG call.


Paola Di Maio Chair, W3C AI Knowledge Representation Community Group