WebMCP TN 4: is the spec ready? https://web-mcp.net/

Anssi and everyone

thanks in advance for handling concerns with diligence and patience

Just venting some concerns for discussion, before catastrophic failures may
accidentally occur

I may or may not be able to flag these in Github (.....)

PDM


*Technical Note* *28 February 2026      *

*FROM PDM W3C AI KR CG *

*TO ALL*
------------------------------

WebMCP is a draft specification with empty security and accessibility
sections. It is not a W3C Standard. Commercial products and services are
already available

WebMCP is a proposed browser API being incubated by the W3C Web Machine
Learning Community Group. It allows websites to expose JavaScript functions
as structured tools that AI agents can discover and invoke. Chrome 146
ships an early preview behind a feature flag. The draft is dated 27
February 2026.

The spec itself states: "It is not a W3C Standard nor is it on the W3C
Standards Track." (Source: https://webmachinelearning.github.io/webmcp/)
Security and Privacy?

The specification's Security and Privacy section is empty. It contains only
a TODO comment linking to a separate document. The Accessibility section is
completely empty -- no text at all. All four core API method definitions
say "TODO: fill this out."

This means there is currently no normative guidance on how browsers should
handle prompt injection through tool descriptions, how users should be
informed when tools are registered on a page, how cross-origin tool data
should be isolated, what consent model should govern agent-to-tool
interaction, or how WebMCP tools relate to the existing accessibility tree.

(Source: https://github.com/webmachinelearning/webmcp/blob/main/index.bs)
Already on the Market?

Multiple commercial ventures have launched products and services around
WebMCP within days of the Chrome preview. These include paid "Agent
Readiness" assessments, enterprise security scanners, CLI audit tools, CMS
plugins, and partner programs -- all built on a specification that has not
defined its own security model.

Businesses are being told to annotate their forms with WebMCP attributes so
AI agents can submit them programmatically. Fear-of-missing-out marketing
frames this as "SEO for AI" and warns that companies who do not implement
WebMCP will be "skipped by agents."

This is premature. The security implications of exposing website
functionality to autonomous agents through a browser API without a defined
consent model, permission framework, or threat analysis have not been
resolved by the standards body. Selling security tooling for a threat model
that does not yet exist is not responsible engineering.
What WebMCP Is Not

WebMCP is not the Model Context Protocol (MCP).

 It does not implement the MCP wire protocol (JSON-RPC 2.0). It is not
interoperable with MCP client libraries. It borrows the tool abstraction --
functions with schemas and descriptions -- but implements everything
through browser-native mechanisms.

The name creates confusion that is being commercially exploited.

(See Technical Note 3:
https://github.com/Starborn/webmcp/blob/main/WebMCnotMCP.md)
What You Should Do

If you are a developer: contribute and experiment with the Chrome preview,
but do not deploy WebMCP tools on production sites until the security model
is defined.

If you are buying services: no commercial product can deliver WebMCP
security compliance because the spec has not defined what compliance means.

If you are a standards participant: the W3C Web Machine Learning Community
Group meets next on 5 March 2026. Comments can also be submitted via the
public mailing list (public-webmachinelearning@w3.org) or as GitHub issues (
https://github.com/webmachinelearning/webmcp/issues). The window for
meaningful input is now.
References

   - W3C Draft Spec: https://webmachinelearning.github.io/webmcp/
   - Spec Source (index.bs):
   https://github.com/webmachinelearning/webmcp/blob/main/index.bs
   - Security/Privacy Doc (separate, not in spec):
   https://github.com/webmachinelearning/webmcp/blob/main/docs/security-privacy-considerations.md
   - Technical Notes 1-3: https://github.com/Starborn/webmcp/
   - W3C CG Mailing List: public-webmachinelearning@w3.org
   - Issue Tracker: https://github.com/webmachinelearning/webmcp/issues

------------------------------

*. Corrections and discussion welcome.*