Meeting minutes
<gb> Issue 39 WebML WG/CG F2F Agenda - TPAC 2026 (Dublin, Ireland) (by anssiko) [Agenda+]
Repository: webmachinelearning/webmcp
Intros & agenda bashing
Anssi: Welcome everyone to the WebML CG F2F TPAC 2026 meeting
… I'm Anssi, your chair for this meeting
… this is a big group, so I'd ask folks to introduce themselves when you queue up to speak
… to get set up, please join IRC:
https://
Anssi: and type Present+ Yourfirst_lastname
… type "q+" on IRC to queue up to speak, this will make sure our agent that helps with minute-taking is able to correlate your name with what you say
WebMCP since Kobe, and lately
Anssi: I want to start by acknowledging major spec progress since Kobe
Dominic: WebMCP has grown from its initial imperative API into a broader design covering cross-document use, elicitation, security, and declarative work
Dominic: Origin Trial feedback and framework adoption show that cross-document continuity and predictable tool availability are key developer needs
Dominic: The security model distinguishes what sites expose from the capabilities and policy of the agent consuming those tools
<Philippe> are folks still seeing network issues in the room?
<Philippe> Anssi ^
API shape & multi-stakeholder support
Anssi: in this first session our goal is to converge on the API direction and some key design considerations
Anssi: generic communication tool issue #236 and API design questions #237 are feedback from Mozilla
<gb> Issue 237 Various questions & feedback (by jakearchibald)
<gb> Issue 236 Reposition feature as a generic communication tool? (by jakearchibald)
Anssi: these issues are also of interest to the TAG
<Zakim> jyasskin, you wanted to suggest that we try to unify the LLM-supporting attributes between JS tools and HTML
<Kenji> search_faq_unless_its_about_pricing_or_maybe_delivery_time_because_those_are_too_hard_questions_thanks(yourquestion)
<Kenji> +1
<Jeffrey> I like Liad's idea of a "describeWorkflow()" kind of function, that explains to agents how they might compose a group of tools.
Jake: The group should define a generic cross-environment tool-calling primitive and justify any agent-specific API separately
Dominic: A unified primitive can retain agent-oriented metadata while allowing explicit controls over which consumers can discover a tool
Johann: The API should begin from a safe exposure default and require sites to opt into broader consumers
Liad: Agent use differs because a model interprets natural-language descriptions, while deterministic callers need machine-readable contracts
Yoav: The group should document the concrete differences between generic RPC and agent tool calling rather than assume separate APIs are necessary
Jeffrey: Schemas are useful for ordinary cross-origin calls as well as agents, and descriptions can remain optional when names and types are sufficient
Jeffrey: Exposure should describe purposes such as testing or workflows instead of depending only on an implementation label such as agent
Benjamin: A generic primitive could also support form actions and other declarative integrations
Ilya: Descriptions and input and output schemas materially improve current agent behavior, especially with code-mode agents
Jake: The design should account for model capabilities changing over time and avoid permanent restrictions that are only needed by current models
Sushanth: A generic RPC layer may need values beyond JSON and clearer execution semantics for main-thread and worker tools
Kenji: Concrete generic communication use cases are needed to evaluate how much demand exists outside agent use
Mark: Publicly exposed tools should use stronger types and agent-compatible descriptions, while same-origin tools may need less metadata
Dominic: One API with explicit exposure properties is preferable unless examples demonstrate that the combined model harms agents or developer experience
<Jake> +1
<Sushanth> +1
<Alex> +1
<Benjamin> +1
<Yoav> +1
<Jake> Show that the combined model does reduce the API's usefullness for agents, or create a DX issue
<Dominic> +1
<Benjamin> +1
<Johann> +1
<Jake> +1
<ChrisH> +1
<Samuel> +1
<Sushanth> +!
<Jessica> +1
<Nathan> +1
<Julia> +1
<Sushanth> +1
<Khushal> +1
<BenL> +1
<Yunhao> +1
RESOLUTION: The group agrees to explore the direction to make the API a more generic communication tool with focus on adequate security bar, tool registration defaults, infrastructure primitives. Next step to deliver API shape sketches for group review. Show that the combined model does not reduce the API's usefulness for agents, or create a DX issue. (issues #236 #237)
<gb> Issue 236 Reposition feature as a generic communication tool? (by jakearchibald)
<gb> Issue 237 Various questions & feedback (by jakearchibald)
Tool discovery & cross-origin exposure
Anssi: these issues highlight the need to agree on the discovery and exposure model
Discovery beyond one traversable navigable
Anssi: issue #227
<gb> Issue 227 Tool discovery should not be limited to a single traversable navigable (by domfarolino)
Anssi: we resolved this as "Draft a concrete proposal for review"
Anssi: Kushal provided a concrete proposal for review and discussion
Dominic: The proposed ModelContextClient lets a caller request tools from a specific Window or environment rather than implicitly traversing an entire frame tree
Khushal: Provider state belongs to the Document lifecycle, while the client belongs on Window because it addresses the currently committed document
Jake: Worker support should use a consistent global shape; the group should avoid repeating historical Window-versus-Document placement mistakes
Liad: A per-frame API requires agents to track frame loading, so a frame-tree convenience API also serves discovery and stability needs
Johann: Both explicit per-environment access and a separately designed subtree or frame-tree convenience operation can coexist
Benjamin: A frame-tree operation must preserve iframe security boundaries and avoid presenting untrusted third-party tools as the top-level site
Mark: The top-level site may need to curate or re-expose tools so agents see site capabilities rather than incidental iframe structure
<Johann> +1
<Jake> +1
<Benjamin> +1
<Khushal> +1
<Dominic> +1
RESOLUTION: Agree on the ModelContextClient API shape (per environment tools). Iterate on a getAllFrameTreeTools()-like API separately. (issue #227)
Host-controlled iframe access
Anssi: issue #317
<gb> Issue 317 Host-controlled WebMCP access between iframes (by sparkinson) [enhancement]
Anssi: Dom extracted the gist of this issue for host-controlled iframe access, do you want to speak to this?
<Jake> Seems like URLPattern doesn't support '*', it's '*://*:*/*' which is ick
<Jake> new URLPattern({}) is like '*' though
<Dominic> (thanks for fixing typo)
<Dominic> +1
<Khushal> +1
<Benjamin> +1
<Jake> +1
Dominic: The host-controlled case lets a site explicitly expose selected tools to another window, while privileged installed agents are a distinct security case
Khushal: The motivating web-agent flow must use a clear two-sided authorization model rather than inheriting access merely because one page opened another
Johann: A wildcard exposure is powerful and needs careful treatment because hostile pages can socially engineer users into granting access
Jessica: Testing and developer tooling need broad discovery across iframes, which may justify a distinct privileged mode
Liad: URLPattern offers a more precise mechanism than an unrestricted wildcard for ordinary window-to-window communication
Benjamin: The design should distinguish ordinary web callers from browser or extension contexts that already have elevated privileges
<François> `new URLPattern()` also works for `*`
<Jake> François: nie
<Jake> nice*
<Khushal> +1
<Johann> +1
<Dominic> +1
<Jessica> +1
<Thomas> +1
<ChrisH> +1
<Yunhao> +1
<Sarah> +1
RESOLUTION: Support either URLPattern or explicit '*' in exposedTo and fromOrigins for window-to-window comms. In a separate "privileged installed agent forum", evaluate the ability to bypass origin exposure list. (issue #317)
<Jake> +1
Two-sided opt-in
Anssi: PR #326
<gb> PR 326 Document cross-origin tool exposure and two-sided opt-in in Section 6.3.4 (by jpagnucco)
Anssi: I think we can merge this PR that adds "Enforcing Same-Origin Isolation via Two-Sided Opt-In" and "Residual Risks" sections to Security and Privacy Considerations
<Khushal> +1
Julia: The two-sided opt-in text documents same-origin isolation and residual risks, and has received human review in addition to automated checks
Liad: Top-level curation avoids exposing iframe tools unless the embedding site intentionally forwards them
Johann: Tool-name collisions and provenance need to remain understandable when capabilities are re-exposed across frames
<Benjamin> +1
<Dominic> (I landed the PR)
<Jake> +1
<Dominic> +1
RESOLUTION: Group incorporates Enforcing Same-Origin Isolation via Two-Sided Opt-In and Residual Risks into S&P Considerations. Review and revise as appropriate from the perspective of the generic tool-calling approach. (merge PR #326)
<Samuel> +1
Declarative API next steps
Anssi: the January resolution to pursue declarative WebMCP was overtaken by lower-than-expected developer uptake for declarative
Anssi: as a result, the declarative surface was removed from the WebMCP specification in PR #338
<gb> MERGED PR 338 Remove declarative API from spec (by beaufortfrancois)
Anssi: Chrome team shared its plan to ship the imperative API first, and follow up with a declarative API
Anssi: it would be appropriate to continue the declarative API work in the explainer form in this same repo
https://
Anssi: and continue track declarative issues, labeling them appropriately to facilitate discussion and prioritization
Anssi: once the implementations pick this up, it will be possible to evaluate the need for further specification work
<Dominic> +1
<Khushal> +1
<Alex> +1
<ChrisH> +1
<Samuel> +1
<Sushanth> +1
Dominic: Declarative tools remain useful, but implementation experience should precede reintegration into the normative specification
Benjamin: Declarative forms and imperative tool schemas may share concepts even if the declarative work proceeds independently
Jake: Declarative exploration may also fit broader platform work and should not be blocked on the complete WebMCP API
RESOLUTION: Declarative WebMCP work continues in the explainer form within this repo, advancement to spec stage will be considered once implementation experience is available and has been evaluated with real users against real-world use cases.
Developer feedback & ecosystem
Anssi: we see strong interest from the developer ecosystem, and our community continues to actively engage with the WebMCP initiatives
… the Chrome and Edge Origin Trials have been extended
… WebMCP challenge received 5000+ submissions
… implementations are progressing, divestifying beyond traditional browser environments
… the WebMCP polyfill was adopted as an official W3C Open Source Program pilot, announcement out today
… developer feedback is crucially important, and I want us to discuss each of these points to distill actionable insights
Origin Trial feedback
Anssi: Dom reported on the feedback received from the Origin Trial participants:
https://
Anssi: Dom, please share with us what you found out from the first Origin Trial run
Dominic: Origin Trial feedback repeatedly identified cross-document continuations, stable tool readiness, and framework-friendly integration as priority gaps
Sarah: Framework and developer outreach should convert trial feedback into concrete examples and interoperable guidance
Dominic: The challenge submissions and polyfill provide a useful corpus for validating API assumptions against real applications
https://
Browser and Agent Implementation Status
https://
Anssi: the WebMCP polyfill launched as a W3C OSS Program pilot, congrats to the project team, Alex, Dom, Francois, Kasper
Anssi: this polyfill provides a reference implementation for WebMCP, helping developers understand and experiment with the API before native support is widely available
Privacy & security
Anssi: WebMCP spec has started work on Security and Privacy Considerations, Mitigations, and Best Practices
https://
Anssi: we've collaborated with W3C's Security IG on Threat Model for WebMCP, with draft available at:
https://
Anssi: the group has initiated wide review with both security and privacy groups:
<gb> Issue 135 WebMCP 2026-06-25 > 2026-08-01 (by domfarolino) [REVIEW REQUESTED] [WD]
<gb> CLOSED Issue 213 WebMCP 2026-06-25 > 2026-08-01 (by domfarolino) [WD] [REVIEW REQUESTED]
Anssi: we believe WebMCP privacy and security considerations are closely connected and should be addressed together in the context of WebMCP
… consider this a kick off for bringing us together to work on this
Anssi: privacy feedback from Nick #313 asks for concrete mitigations
<gb> Issue 313 Mitigations for profound privacy and security issues seem to be missing? (by npdoty) [Agenda+] [privacy-needs-resolution] [beyond-scope: agents-on-the-web]
<Nick> sorry if I didn't properly introduce myself, I'm Nick Doty, at Center for Democracy & Technology, and a co-chair of the Privacy Working Group
<Zakim> npdoty, you wanted to describe customized attack surface and where to describe mitigations in user agents
<Dominic> The same origin policy constrains web content. Not users.
<Nick> Dominic, I think the same origin policy also constrains user agents -- browsers implement restrictions in order to maintain the same-origin policy
<Dominic> Yes, so we should define the relationship between user agents and AIs. In the agentic web working group. Not in the webmcp spec
RRSWeb Agent, draft minutes
<Benjamin> ?
<Johann> +1
<Julia> +1
<Mark> +1
<Victor> +1
<Dominic> +1
<Sarah> +1
<Alex> +1
Nick: The specification should state concrete mitigations for the customized attack surface created when sites describe functions for agents to invoke
Julia: Some threats depend on agent behavior outside WebMCP, so agent implementer guidance is needed alongside WebMCP-specific requirements
Victor: WebMCP should address information and controls it introduces, while broader agent behavior belongs in a venue covering agent security
Benjamin: Normative WebMCP requirements and broader agent guidance are complementary; moving general guidance elsewhere must not leave specification-specific gaps
Johann: The ecosystem needs a web security model for agents, but the WebMCP specification should still capture requirements tied directly to its API
Simone: Threat modeling should use model behavior to understand attacks while keeping mitigations grounded in enforceable platform boundaries
Alex: The security and privacy section should be expanded so first-time readers can understand both the risks and where mitigations apply
Dominic: Prompt-injection risks also exist in ordinary page content; WebMCP can provide structured, constrained tools but cannot define the entire agent relationship
Mark: Shared guidance can combine lessons from WebMCP tool use and page actuation so agent implementers apply consistent protections
Nick: Testing known attacks is useful but cannot replace normative boundaries, because an evaluation-only defense will continually lag new attacks
Victor: A shared evaluation set can establish a shipping baseline while implementers continue developing stronger normative mitigations
<Benjamin> +1
<Zakim> npdoty, you wanted to comment on venue
<Julia> BTW this is what we have published already on this topic: https://
RESOLUTION: Joint group agrees it is important to give adequate focus on general "Web Security and Privacy Guidance for Agent Implementers" in a group chartered for that work.
<Julia> +1 to Victor
<Nick> +1, I think the privacy concerns are closely related to the security risks, and we should coordinate
<Khushal> 328 please
<BenG> 328
<Mark> 135
<Nick> a test set and testing against known attacks makes sense. but I would think if that's among the primary mitigations, that defense will always be catching up. if we instead defined specific normative mitigations, then security threats are about fixing violations of a known boundary, rather than just describing attacks as they come.
<Khushal> PR 327 as well
<Samuel> 165
<Johann> I agree in principle Nick, but I'm not sure you'd be entirely satisfied with today's snapshot of such a writeup
Cross-document continuations & page context
Anssi: this top-level topic covers cross-document continuations and page context handling in WebMCP
Continuation designs: tokens v continueAfterNavigation
Anssi: The 1 October resolution advanced continuation tokens design
Anssi: WebMCP continuations explainer PR #327 was merged
<gb> MERGED PR 327 Add explainer for WebMCP continuations. (by markafoltz)
Anssi: in issue #135 Sushant shared continueAfterNavigation proposal:
<gb> Issue 135 Cross-document tool response mechanism (by domfarolino) [declarative]
<Ali> what about extensions? thatgroup is here in the room now...
webmachinelearning/
<gb> Issue 135 Cross-document tool response mechanism (by domfarolino) [declarative]
Anssi: let's discuss how we reconcile continuation tokens with the continueAfterNavigation approach
webmachinelearning/
<Nick> Johann, indeed I expect I wouldn't be satisfied. one approach would be to try to define boundaries for doing it safely before shipping it extremely widely, but I understand implementers might have different priorities.
<Simeon> Anssi: The 13:15-14:00 block only covered Privacy & security. Are we going to discuss webextensions? Most of the relevant folks are still in the room, but if we don't have time it might be better for our group to join the hallway track
<Victor> Nick I think ultimately we should be approaching it from both ends - we should have an ongoing effort to figure out the right normative mitigations that can help with broad sets of issues. While having some form of a shared testing set (evals for WebMCP in this case), can raise the floor
<Victor> The testing can form baseline expectations for shipping safely
<Dominic> +1
<Liad> +1
<Mark> +1
<Sarah> +1
Mark: Continuation tokens associate a tool result across navigation without requiring the agent to infer whether the new document continues the same task
Sushanth: Developer feedback favors a mechanism that works with server-rendered applications and existing framework lifecycles without imposing Service Worker concepts
Jake: A Service Worker can provide a long-lived endpoint across navigations and may avoid a WebMCP-specific continuation mechanism
Nathan: A Service Worker approach needs a clear protocol for page readiness, errors, cancellation, and concurrent tool calls
Sarah: Framework-level helpers could hide Service Worker complexity and provide a higher-level continuation abstraction
Dominic: A Service Worker still needs correlation state between the pre-navigation and post-navigation documents; tokens make that relationship explicit
Liad: The agent can continue addressing a stable Service Worker endpoint while the worker coordinates page-specific operations
Reilly: Navigation semantics may benefit from a general way to communicate the relationship between old and new documents
Mark: The token design keeps continuation state local to the document exchange and avoids requiring every application to adopt Service Workers
Jake: The group should review a concrete Service Worker alternative with framework authors before choosing between the approaches
RESOLUTION: Review ServiceWorker alternative to PR #327 design with frameworks. (related to issue #135)
Tool stability
Anssi: issue #328
<gb> Issue 328 Signaling tool stability to agents (by domfarolino) [Agenda+] [needs-agent-feedback]
Anssi: Dom put forward a strawman proposal for signaling tool stability to agents
… don't get caught up with the API shape, it is subject to change
… I see Sam supporting the general idea
… Khushal proposed toolsReady signal
<Khushal> webmachinelearning/
<gb> Issue 328 Signaling tool stability to agents (by domfarolino) [Agenda+] [needs-agent-feedback]
Anssi: Kenneth provides support from an extension-based implementation point of view
<Simeon> Curious to hear feedback from extension implementations. Is there a link I can check?
<Khushal> webmachinelearning/
<Sushanth> nit: why does document.modelContext.waitForToolsUntil(toolsReady);, take a promise rather than return it ?
<Jake> Sushanth: It needs to be given some indication of "work started" and "work ended" or "work failed" - a promise provides those states
Khushal: The toolsReady proposal lets a page provide a promise marking when asynchronous tool registration has completed
Alex: A blocking attribute on the script that registers tools could connect readiness to existing script loading behavior
Yunhao: The client may also need an observable readiness state or event, not only a promise supplied by the provider
Jake: The platform does not expose a universal point when progressive rendering is complete, so readiness must be explicitly scoped to tools
Liad: Asynchronous registration is expected to be common and must not leave agents guessing whether a tool set is final
Samuel: A blocking=tools script could ensure registration code is fetched and evaluated before tools are reported ready
<Dominic> +1
<Samuel> +1
RESOLUTION: The group agrees the tool stability design is sound, point 2 needs more clarification and feedback from more browser vendors. (issue #328)
Human-in-the-loop & user activation
Anssi: now we enter the space of user interaction
Elicitation
<gb> PR 204 Replace requestUserInteraction with requestUserInput (by bwalderman) [Agenda+]
<gb> Issue 165 WebMCP elicitation via requestUserInteraction() (was: Human in the Loop support for non-browser clients) (by MiguelsPizza) [enhancement] [Agenda+]
Anssi: this is a long issue
… the most recent resolution is to "initially specify the "interactive" mode in PR #204"
… and then "continue discussion on "form" and "url" modes" in issue #165
<Liad> webmachinelearning/
<gb> Issue 165 WebMCP elicitation via requestUserInteraction() (was: Human in the Loop support for non-browser clients) (by MiguelsPizza) [enhancement] [Agenda+]
Samuel: The feature needs multiple interaction modes because a tool may require a simple confirmation, structured form input, or navigation to a URL
Dominic: The API should help tool authors choose an appropriate elicitation mode without forcing agents to invent unsafe interaction patterns
Ilya: The agent may be able to choose among interaction options, but the site must retain control over sensitive or domain-specific flows
Jake: Sensitive credentials and one-time passwords should remain in trusted site UI rather than flowing through the agent conversation
Liad: The three proposed modes cover distinct needs, but their exact relationship and selection semantics need more design work
Sushanth: For destructive actions such as deleting a repository, the site must decide when to show confirmation and what context the agent may carry
Khushal: The design should support explicit user delegation without treating prior approval as unlimited authority for later operations
<Jake> I don't have enough context to +1 this but it seems like there isn't time to catch me up
<Samuel> +1
<Sushanth> +1
RESOLUTION: Agreement the feature is required, further whiteboarding time needed to hash out the details. (issue #165 PR #204)
<Khushal> +1
Built-in AI APIs overview
Anssi: Built-in AI APIs in room 07 on Friday at 11:30-12:45
Jake: The remaining time is insufficient for a substantive built-in AI API discussion
Anssi: The group will request additional room time on Friday and continue the built-in AI API discussion there
<Ehsan> https://