Meeting minutes
Anssi: as a reminder, we'll use IRC-based queue management in this meeting:
https://
Anssi: to suggest agenda topics, use Agenda+ label, e.g.:
Anssi: please welcome our latest new participants
… Nathan Schloss from Meta
… Idan Levin from nekuda
… Sohail Hooda from LogMeIn Inc.
… Ian Holmes, Eyup Ucmaz, Melvin Carvalho as individual contributors
… welcome all!
Announcements
WebMCP polyfill
Anssi: WebMCP polyfill project has been launched
… congratulations to the team with special thanks to Alex Nahas and Francois Beaufort for the initial contributions to establishing the project
webmachinelearning/
Anssi: this WebMCP polyfill project directly supports W3C's spec work and is developed in an open and transparent manner
… and thus qualifies as an official W3C Open Source Program pilot project
Anssi: I have invited Francois Daoust who co-leads W3C's Open Source Program to tell more about the pilot and next steps
Francois: three high-level things, the scope of OSS Program is to recognize W3C is always grounded on open source, we like to embrace that and get more things out that work hand in hand with W3C standards
… WebMCP polyfill is 100% aligned with that
… I saw discussion around naming of an npm package, so @w3c/webmcp-polyfill would be available to the project in this pilot
<gb> @w3c
Francois: during this pilot, we are working closely with the projects that are onboarding
Anssi: thank you Francois for the guidance
Anssi: I will work with the polyfill team on practicalities to advance the WebMCP polyfill project into the W3C OSS Program pilot project
WebMCP momentum and next steps
Anssi: we're all heads down working so it is easy to overlook the remarkable growth and achievements of everyone working on WebMCP in this group
Anssi: first, growth in the WebMCP ecosystem is remarkable
… this Community Group that initiated the WebMCP effort a year ago has grown +55% YoY, we're currently 314 participants across 84 organizations
https://
Anssi: the WebMCP repo has grown from 0 to 4.4k GH stars YoY, remarkable for a web spec:
Anssi: the recent WebMCP Challenge attracted over 5k submissions from developers:
https://
Anssi: the broader web community continues to impress us with their engagement, contributions, and innovative use cases for WebMCP, some of which are captured in the Awesome WebMCP repo:
webmachinelearning/
Anssi: WebMCP implementation experience is growing steadily and most importantly diversifying to different platforms and environments:
https://
Anssi: during this rapid growth phase, this Community Group has proactively engaged with various stakeholders to ensure this growth of the WebMCP ecosystem happens in a responsible manner, focused on privacy, security, the needs of humans and agents
… the group is actively working with W3C's horizontal groups; the TAG, Privacy, Security, Accessibility, and Internationalization, helping these groups understand WebMCP and in turn improve WebMCP in these critical areas
… I want to thank all of you for your dedication and contributions to the WebMCP effort
Anssi: I've invited W3C's Philippe Le Hégaret who's responsible for W3C's tech strategy to outline the next steps on the agentic web journey
Philippe: as we all know, humans are relying on AI agents to do their work, on the web, and we hope it matches the user intent, we want the web to be accessible to agents, WebMCP allows actuation by AI agents
<plh> https://
Philippe: we announced yesterday we're working on Agentic Web WG to deal with that problem space
… intersection of agents and web interactions
… this is an early draft, I expect conversations at TPAC, maybe in this CG
… TPAC breakouts have many proposals that are in the space of agentic web
… the idea would be that the WebML WG would focus on ML in the browser, and this proposed WG would focus on agent interactions on the web
… potentially also best practices docs for developers to help them design their web sites for better consumption by agents
Dominic: thanks for getting this off the ground, we've been interested in the venue, to have security and harness issues focus, this sounds a nice way to also work on those
… how do I join, the charter draft join button is not working
Philippe: it is just a draft, it will take a few months to set this up
… some issues you have, need some incubation, so we expect incubution, using WebML CG for incubating those ideas prior
Dominic: one of the deliverables is WebMCP, I guess the plan is to adopt it
Philippe: that is the current plan
<plh> w3c/
Johann: can you paste the link how to follow along the process?
<gb> Issue 576 [wg/agenticweb] Agentic Web Group Charter (by plehegar) [AI] [In Charter Refinement] [charter]
<plh> w3c/
Philippe: that's the link to the strategy issue
Philippe: I can organize a breakout and discuss during the WebML CG TPAC F2F this proposal
Philippe: the plan is we'd be happy with the charter by TPAC end
<Zakim> tomayac, you wanted to briefly mention the WebMCP extension
<domfarolino> (sorry for jumping you tom)
<tomayac> googlechromelabs/
Tom: on WebMCP, brief announcement, our team, we introduced WebMCP extension
… working in the Prompt API and WebMCP intersection
<tomayac> In the extension, you can choose the Prompt API to back the WebMCP tool calling.
Philippe: keep up the great work on WebMCP!
Dominic: one thing that makes the WG successful is when we define normative requirements for browser use agent harness, have frontier labs in the room
Philippe: I have a call with OpenAI next week, they presented at W3C workshop last month
WebMCP
Repository: webmachinelearning/webmcp
Wide review
Anssi: per our resolution, we have initiated web accessibility (a11y) and internationalization (i18n) reviews:
<gb> Issue 189 WebMCP 2026-10-01 (by anssiko) [SR] [WD] [REVIEW REQUESTED] [pending] [s:webmcp]
<gb> Issue 338 WebMCP 2026-10-01 > 2026-10-27 (by anssiko) [REVIEW REQUESTED] [SR] [pending]
Anssi: as part of these review requests, we have provided as references a11y and i18n checklists the group prepared:
<gb> Issue 272 WebMCP accessibility review checklist (by anssiko) [Agenda+] [a11y-tracker]
<gb> Issue 273 WebMCP internationalization review checklist (by anssiko) [Agenda+] [i18n-tracker]
Anssi: as discussed, both of these reviews and related issues remain open for additional community feedback
… these wide reviews are an ongoing process, not one-offs
… any comments, questions about a11y and i18n reviews so far?
Anssi: in June, we reached out to the Privacy WG for privacy review and feedback on WebMCP's privacy implications:
<gb> CLOSED Issue 213 WebMCP 2026-06-25 > 2026-08-01 (by domfarolino) [WD] [REVIEW REQUESTED]
Anssi: the Privacy WG group first discussed WebMCP on its 16 July call:
https://
Anssi: last week, two privacy-related issues were opened for WebMCP for discussion:
<gb> Issue 313 Mitigations for profound privacy and security issues seem to be missing? (by npdoty) [Agenda+] [privacy-needs-resolution]
Anssi: today, I'd like to discuss these two issues to converge on consensus on the group's response
Anssi: first, in issue #313, two pieces of feedback:
<gb> Issue 313 Mitigations for profound privacy and security issues seem to be missing? (by npdoty) [Agenda+] [privacy-needs-resolution]
Anssi: i) an observation that an agent with broad authority can create risks
… ii) suggestion to consider threats where the site author is the attacker
… for the first, the initial response from Dominic suggests this concern is relevant regardless of WebMCP and would be best addressed on a different layer
… for the second, I wonder if some non-normative "Guidelines for Agent Harness Implementers", possibly a separate document, or deliverable, would be helpful
… Google folks have published materials that could feed into these guidelines to turn this into community effort:
<Julia> https://
Anssi: thoughts on the privacy feedback, specifically these two issues #313 and #314?
<gb> CLOSED Issue 314 visible UI and forms might have fewer open-ended prompt injection attacks (by npdoty) [Agenda+] [security-tracker] [privacy-needs-resolution]
Dominic: well covered, Anssi
… we've been using a new label, "beyond scope agent identity" as a collection of issues that should be solved in Agentic Web WG scope
… I think it'd be great to point to the resources Julia shared and work on them
Victor: it seems the concerns from Privacy WG are similar to discussion we had with Security WG
… perhaps we could collapse the both groups and work on both privacy and security considerations together
… prompt injections is something we've considered, also need to discuss this with both security and privacy groups present
<domfarolino> +1
Julia: I agree with Victor and Dom, agent security is a bit out of scope as of now, we say that is pending for new WG where to discuss those aspects
MarkF: one topic which is somewhat in scope of this group is issue #173
<gb> CLOSED Issue 173 Make tools Document-scoped instead of Window-scoped (by domfarolino)
MarkF: guardrails for harnesses, internal discussions around that
Anssi: I propose we will have a joint discussion with the Privacy WG at TPAC to chart the path forward for new deliverables
TPAC F2F
Anssi: I want to use this opportunity to gauge F2F issue priorities, cross-group coordination and breakout interest
… we have roughly 4 hours of F2F time for WebMCP specifically and 1 hour for built-in AI APIs
… first, about F2F issues priorities
… I'm suggesting the group to assign labels to issues they think are worth discussing during the F2F meeting, editors but anyone else as well:
https://
Anssi: and respectively for built-in AI APIs
Anssi: please go to issues of your favorite repositories, use the existing labels or mint new labels to help indicate F2F discussion priorities and interest, or add a comment if you can't add labels
… alternatively you can add comments to our meeting issue:
webmachinelearning/
<gb> Issue 39 WebML WG/CG F2F Agenda - TPAC 2026 (Dublin, Ireland) (by anssiko)
Anssi: the group is executing like a well-oiled machine, so I don't want to freeze our agenda too early but consider dynamic adjustments based on ongoing discussions and emerging priorities
Dominic: how many groups we are meeting at TPAC?
… we have joint meetings with horizontals, and breakout with Web Extensions WG
… I can propose WebMCP breakout session, if that is helpful
Dominic: perhaps focusing on technical discussion in this CG is adequate
Cross-document tool response mechanism
<gb> PR 327 Add explainer for WebMCP continuations. (by markafoltz)
<gb> Issue 135 Cross-document tool response mechanism (by domfarolino) [Agenda+] [declarative]
Anssi: prior, we resolved this issue to explore a solution for the declarative API
… MarkF has prepared an explainer for a solution for the imperative API
… building upon a proposed new higher-level abstraction based on token creation and token redemption, called Continuation Tokens
… the explainer PR is out for review:
https://
MarkF: problem statement is that from the agent's point of view, the tool can span multiple documents
… navigating to another document, or when the tool does work in one frame and wants to hand off work to another frame within the same frame tree
… currently a tool cannot span multiple documents
… this proposes a mechanism where an agent can request a token and pass the token that can be redeemed and continue working
… enables chain of tool execution between documents
… initial proposal enables tokens to be produced and consumed in same browsing context, i.e. popup or another tab cannot consume tokens from another tab
… agents can see websites as a single tool, example is a checkout flow
Anssi: any open questions?
MarkF: how to communicate the token across documents
… API shape in general, if you have app spanning pages, and you want to implement tools across them
Dominic: we would welcome web platform feedback, also agent implementer feedback, Copilot folks feedback, OpenAI feedback, this is important for agents that drive WebMCP tools
… native agents with access to browser internals, feedback from them is important
Sushanth: in our prototyping we handled passing between documents, so that tool call had an additional value in JSON
… would like to understand why you are not considering passing state across document navigation, maybe for security reasons?
MarkF: browser state we are willing to provide is original input, actuation of the tool, if there's state beyond it, it is up to the application
… we allow tool continuations to see partial outputs
Dominic: did Microsoft's prototype return a specific value?
Sushanth: yes, on the orchestration layer, after the page navigation
MarkF: earlier proposal had this re-entrant tool calling, we felt better to have this design we're proposing now
Sushanth: passing state between tool location was part of the protocol in our implementation
… in your proposal you make it an app problem, not WebMCP problem
Dominic: can consider partial state, we want to see partial state exposed to the model
… we need to figure out what the entry point is, do we force the developer to provide continuation for the previous tool
… a tool would not need to be registered on the target page
Sushanth: how does it know what methods to call to continue with the tool?
MarkF: Sushanth, please provide feedback in the issue
<domfarolino> +1
RESOLUTION: The group explores the Continuation Tokens design outlined in the explainer and welcomes further feedback. (issue #135 and PR #327)
Signaling tool stability to agents
Anssi: issue #328
<gb> Issue 328 Signaling tool stability to agents (by domfarolino) [Agenda+] [needs-agent-feedback]
Dominic: would love Microsoft's feedback
Sam: Microsoft has interest in this, there is need for this and I can take another look to provide additional feedback
Prompt API
Repository: webmachinelearning/prompt-api
Thinking mode
Anssi: PR #217
<gb> MERGED PR 217 Add Thinking mode section to the Prompt API explainer (by isaacahouma) [Agenda+]
Anssi: explainer available:
webmachinelearning/
Reilly: feedback requested for this feature