Guide : WebMCP

TL;DR — WebMCP lets a website décrire its buttons and functions as structured outils an AI agent peut appel pendant que lune page is ouvrir. Que peut faire a réel page action—searching inventory, validating code, filtering products—plus reliable que asking an agent to guess qui contrôler to click. It fait pas faire lune page easier to rank, explorer, or découvrir. As of July 17, 2026, it is experimental: a Community Groupe draft with a Chrome 149 origin trial, pas a finished web standard.

Ce que WebMCP is

WebMCP is a proposed navigateur API que lets a web page expose structured outils to an AI agent operating in que page’s browsing context. A outil has a nom, a description, structured inputs, and code que performs an action. Evidence for this claim WebMCP is a proposed browser API through which web applications expose JavaScript-based tools to agents in a browsing context. Scope: WebMCP Draft Community Group Report dated July 10, 2026; the report is not a W3C Standard. Confidence: high · Verified: 2026-07-17 WebMCP Draft Community Group Report

Imagine a visitor opens a schema validator and demande leur navigateur agent to “validate this Product JSON-LD.” Sans WebMCP, the agent may inspect the screen, trouver a text area, paste into it, click the correct button, and interpret le résultat. With WebMCP, lune page pourrait offer a validate_schema outil with an explicit input schema and retourner a structured result—pendant que keeping the visible interface in sync.

Que is the utile mental model: WebMCP is an explicit contrôler surface pour an agent que has déjà reached lune page. It n’est pas a nouveau façon to publish prose.

Current status on July 17, 2026

WebMCP is moving quickly, so code and compatibility claims besoin dates. The current position is: Evidence for this claim As of July 17, 2026, WebMCP is a Community Group draft rather than a W3C Standard, and Chrome documents an origin trial beginning in Chrome 149. Scope: Standards and Chrome experiment status checked July 17, 2026; these are time-sensitive claims. Confidence: high · Verified: 2026-07-17 WebMCP: Status of This Document Chrome: Join the WebMCP origin trial

I trouvé aucun primary-source basis in ce research réussir to claim stable WebMCP prise en charge in Edge, Safari, or Firefox. A Microsoft editor contributing to the draft n’est pas the même chose as a navigateur shipping the fonctionnalité.

WebMCP versus MCP

Model Context Protocol and WebMCP expose outils at différent boundaries:

• MCP usually connects an AI application to a persistent server. Le serveur peut provide outils and données si or pas a particulier web page is ouvrir.
• WebMCP lets the ouvrir page expose its current frontend actions and state. Its outils peut modifier quand the visitor navigates or quand an action becomes unavailable.

Ils complement chaque autre. A store pourrait utiliser remote MCP pour a product catalog and WebMCP pour the filter, configuration, and checkout steps visible in the visitor’s current tab. Evidence for this claim WebMCP is designed for tools owned by an active page and its browser context, while remote MCP commonly connects an AI application to a persistent backend server. Scope: Architecture-selection guidance, not a rule that prevents an application from using both technologies. Confidence: high · Verified: 2026-07-17 Chrome: When to use WebMCP and MCP

Fait WebMCP aider SEO or AI citations?

Là is aucun documented evidence que ajout WebMCP improves rankings, exploration, indexation, AI citations, or discovery. The documented fonctionnalité begins après an agent reaches une page and inspects the outils disponible là. Evidence for this claim The documented Chrome WebMCP experiment requires a browsing context and a client visit to discover page tools; no documented ranking, crawling, indexing, or citation benefit was found in this research pass. Scope: The browser-context and visit requirements are direct documentation; the SEO conclusion is a bounded absence-of-evidence statement as of July 17, 2026, not a prediction about future search systems. Confidence: high · Verified: 2026-07-17 Chrome: WebMCP overview and limitations Google Search guidance for AI experiences

It pourrait améliorer the completion of an on-site task. Que may be commercially valuable, but it is a product and conversion hypothesis—pas a ranking factor. Garder doing the ordinary fonctionner que rend une page discoverable and understandable: semantic HTML, accessible formulaires, lien internes, indexable content, and appropriate données structurées.

Editorial pages ne faites pas besoin fake outils simplement to regarder “agent-ready.” Si lune page seulement nécessite to be lire, readable HTML is the correct interface.

Peut Cloudflare enable WebMCP pour my site?

Pas by itself. Cloudflare Navigateur Run currently offers experimental lab navigateur sessions que peut visit and tester pages with WebMCP outils. Votre page encore has to define and register ceux outils. Cloudflare’s documentation aussi dit its lab sessions ne sont pas pour production workloads. Evidence for this claim Cloudflare Browser Run offers experimental lab sessions that can consume and test page-provided WebMCP tools, but the page still owns tool registration and Cloudflare says lab sessions are not for production workloads. Scope: Cloudflare product documentation last updated April 23, 2026; its example API names lag current Chrome documentation and should not be used as the API authority. Confidence: high · Verified: 2026-07-17 Cloudflare Browser Run: WebMCP

Que rend Cloudflare a possible tester environment or consuming navigateur, pas a switch que converts every existing formulaire and button into WebMCP.

Devrait vous implement it now?

Pour la plupart production sites, wait. A petit private experiment peut faire sense quand tout of ces are vrai:

• lune page déjà has a bounded, utile action;
• the normal human interface remains complet sans WebMCP;
• the action is read-only or easily reversible;
• the même application logic sert les deux the UI and the outil;
• inputs, outputs, authorization, and logs recevoir a réel security examiner; and
• success peut be mesuré sans pretending the experiment affecte search visibility.

Ne faites pas utiliser an experimental navigateur API as the seulement chemin to a critical task.

TL;DR — WebMCP exposes page-owned outils from document.modelContext to an agent in the active browsing context. The imperative API registers a nom, description, JSON Schema input, annotations, and an async callback; outils peut be state-dependent and unregistered with an AbortSignal. Chrome documents a declarative formulaire couche aussi, but the July 10, 2026 Community Groupe draft encore étiquettes its declarative section TODO. The fonctionnalité is in a Chrome 149 origin trial, pas a stable cross-browser baseline. Treat it as progressive enhancement, pas SEO infrastructure, and secure it as an authenticated application surface.

How WebMCP fonctionne in le navigateur

A utile WebMCP lifecycle begins and ends with lune page. Lune page registers seulement the actions que are valid in its current state; an agent déjà operating in que browsing context discovers and invokes un; lune page executes its normal application logic, updates the human-visible interface, and renvoie a bounded result. Quand the state changements, lune page removes outils que ne … plus appliquer. Evidence for this claim The imperative API registers named, described, schema-constrained callbacks and supports state-aware cleanup with AbortSignal plus tool-set change notifications. Scope: Current draft and Chrome experiment; API details may change before stable release. Confidence: high · Verified: 2026-07-17 WebMCP ModelContext API Chrome: WebMCP Imperative API

Ce stateful lifecycle is un raison WebMCP ne doit pas become a static dump of every function in a JavaScript bundle. A outil que is impossible in the visible UI devrait normally be unavailable to the agent aussi.

WebMCP vs MCP: two différent runtime boundaries

The noms invite confusion, but the operational boundary is différent. WebMCP lives in a document’s event loop and current navigateur session. Remote MCP lives at the AI application’s integration boundary and commonly reaches a persistent backend server. Evidence for this claim WebMCP is designed for tools owned by an active page and its browser context, while remote MCP commonly connects an AI application to a persistent backend server. Scope: Architecture-selection guidance, not a rule that prevents an application from using both technologies. Confidence: high · Verified: 2026-07-17 Chrome: When to use WebMCP and MCP

Ne faites pas choisir from the acronym. Choisir from the owner of the action. Si the action nécessite the current DOM, selection, cart, or UI state, WebMCP may fit. Si it doit run in the background, à travers nombreux sites, or sans an ouvrir tab, utiliser an API or remote MCP.

The imperative API

The imperative API registers a outil via document.modelContext.registerTool(). The outil inclut a unique nom, description, JSON Schema input, an execute callback, and optional annotations. Chrome aussi documents getTools(), executeTool() pour testing, and a toolchange event. Evidence for this claim Current Chrome documentation uses document.modelContext and says navigator.modelContext is deprecated beginning in Chrome 150. Scope: Chrome implementation guidance checked July 17, 2026; version and API-name claims expire quickly. Confidence: high · Verified: 2026-07-17 Chrome: WebMCP Imperative API

Ce illustrative candidate montre the intended shape pour a future Schema Validator pilot. It n’est pas running on ce site, and the API may modifier avant a stable release:

if (document.modelContext) {
  const registration = new AbortController();

  await document.modelContext.registerTool({
    name: 'validate_schema',
    description: 'Validate pasted JSON-LD and return bounded issues.',
    inputSchema: {
      type: 'object',
      properties: {
        markup: {
          type: 'string',
          description: 'JSON-LD markup to validate.',
          maxLength: 50000
        }
      },
      required: ['markup'],
      additionalProperties: false
    },
    annotations: {
      readOnlyHint: true,
      untrustedContentHint: true
    },
    execute: async ({ markup }) => {
      const result = await validateWithTheSameEngineAsTheUI(markup);
      renderResultInTheVisibleUI(result);
      return minimizeValidationResult(result);
    }
  }, { signal: registration.signal });

  // When this page state no longer supports validation:
  // registration.abort();
}

The important architecture n’est pas the wrapper. It is que the callback calls the même validator as the visible UI, server-side limites and authorization encore appliquer, and la réponse is deliberately minimized. Fonctionnalité detection preserves the complet human workflow in unsupported navigateurs.

Utiliser document.modelContext, pas stale exemples construit autour navigator.modelContext; Chrome marks the latter deprecated beginning in Chrome 150. Date que advice parce que the fonctionnalité remains experimental.

The declarative API has a standards mismatch

Chrome documents a declarative approach que annotates ordinary formulaires with attributes tel as toolname, tooldescription, and toolparamdescription. It aussi documents optional toolautosubmit; sinon, the utilisateur clicks Submit. SubmitEvent.agentInvoked identifies an agent-triggered submission. Evidence for this claim Chrome documents declarative WebMCP form annotations, but the July 10, 2026 Community Group draft says its Declarative WebMCP section is entirely TODO. Scope: A direct comparison between Chrome implementation documentation and the current draft; it does not imply Chrome's experimental implementation is unavailable. Confidence: high · Verified: 2026-07-17 Chrome: WebMCP Declarative API WebMCP: Declarative WebMCP

Cependant, the July 10 Community Groupe Report dit its Declarative WebMCP section is “entirely a TODO” and leaves the form-to-JSON-Schema algorithm undefined. Que ne fait pas mean Chrome’s experiment is imaginary. It signifie the implementation documentation is ahead of the normative draft. Treat declarative markup as an experimental Chrome surface, pas settled cross-browser HTML.

Pour now, ordinary semantic formulaires remain the durable base. An experimental annotation couche devrait enhance les, jamais replace étiquettes, validation, accessibility, confirmation, or server-side authorization.

Outil discovery, lifecycle, and cross-origin boundaries

The current API has several boundaries worth designing explicitly:

• Browsing context: Chrome’s experiment exige a navigateur context. The client visits le site avant it discovers le site’s outils.
• Dynamic availability: register outils quand ils are valid and abort leur registration quand state or navigation invalidates les. Observers peut listen pour toolchange.
• Même origin by par défaut: the tools Permissions Policy defaults to 'self'. Cross-origin iframes besoin explicit delegation tel as allow="tools".
• Two-sided cross-origin consent: a outil peut utiliser exposedTo to liste allowed secure origins, pendant que a caller requêtes outils from named origins with fromOrigins. Un side opting in n’est pas suffisant.
• Progressive enhancement: unsupported or disabled WebMCP doit leave the ordinary page entièrement usable.

Ces are utile platform contrôle, but ils ne faites pas turn a risky application action into a safe un.

Security: le navigateur session raises the stakes

A navigateur agent may operate à l’intérieur the user’s authenticated session. Que peut be the fonctionnalité—accès to the current cart, account, or workspace—and the danger. Chrome and the draft discuss prompt injection, misleading outil metadata, contaminated outil output, over-broad parameters, privacy leakage, cross-origin exposure, and misuse of signed-in authority. Evidence for this claim WebMCP tool hints can communicate read-only and untrusted-output intent, but they do not eliminate prompt injection, misleading metadata, privacy leakage, cross-origin risk, or misuse of authenticated browser authority. Scope: Threat-model and defensive guidance; application authorization and confirmation remain implementation responsibilities. Confidence: high · Verified: 2026-07-17 WebMCP security and privacy considerations Chrome: WebMCP tool security Chrome: Agent security considerations

Treat every outil as a public application endpoint with an unusual caller:

1. Garder the outil narrow. Un job, explicit inputs, tight enums and lengths, aucun hidden “do anything” parameter.
2. Enforce authorization in application logic. The agent ne fait pas gain plus authority que the signed-in utilisateur, and a hint n’est pas permission.
3. Separate reads from writes. readOnlyHint and untrustedContentHint communicate risk; ils ne faites pas enforce it.
4. Exiger visible confirmation pour consequences. Purchases, submissions, deletions, messages, and account changements besoin a human-understandable checkpoint.
5. Minimize output. Retourner seulement ce que the task nécessite; jamais spill session données, raw headers, secrets, or unrelated records.
6. Treat output as untrusted. A string renvoyé by un outil peut become input to plus tard model reasoning. Ne faites pas autoriser it to smuggle instructions or authority.
7. Log the boundary. Record the outil, input class, authorization decision, confirmation, result class, error, origin, and lifecycle sans logging secrets.

The safe question n’est pas “Can an agent call this?” It is “Voudrait I expose ce as a reviewed endpoint to a caller que peut misunderstand instructions and relay untrusted text?”

Tester contracts and agent behavior separately

Chrome’s eval guidance separates deterministic product tests from probabilistic agent tests. Evidence for this claim WebMCP testing should combine deterministic contract and UI-state tests with probabilistic evaluation of agent tool selection and use. Scope: Chrome's evaluation guidance; teams must define product-specific tasks, models, risks, and thresholds. Confidence: high · Verified: 2026-07-17 Chrome: Evals for WebMCP Les deux matter:

Deterministic tests devrait vérifier registration, schema rejection, valid and invalid inputs, authorization, rate limites, side effects, error shape, output minimization, UI parity, unregister behavior, and unsupported-browser fallback.

Probabilistic evals devrait mesurer si representative agents découvrir the correct outil, éviter irrelevant outils, choisir the correct parameters, demander pour clarification quand requis, respect confirmations, arrêter après success, and resist adversarial descriptions or output.

Ne faites pas utiliser un demo prompt as the release gate. A successful callback proves the code ran; it ne fait pas prove que models select it reliably or que the action is safe.

Cloudflare Navigateur Run: utile lab, pas enablement

Cloudflare documents WebMCP prise en charge in Navigateur Run’s experimental lab pool and dit lab sessions ne doit pas be utilisé pour production workloads. Its April 23 page aussi contient older Chrome-era testing noms, so utiliser que page as evidence of Cloudflare’s current product offering—pas as the authority pour the latest WebMCP API shape. Evidence for this claim Cloudflare Browser Run offers experimental lab sessions that can consume and test page-provided WebMCP tools, but the page still owns tool registration and Cloudflare says lab sessions are not for production workloads. Scope: Cloudflare product documentation last updated April 23, 2026; its example API names lag current Chrome documentation and should not be used as the API authority. Confidence: high · Verified: 2026-07-17 Cloudflare Browser Run: WebMCP

Cloudflare peut provide a navigateur session and an agent chemin que consumes outils. It ne peut pas infer the safe contract pour votre application or register page-owned outils vous did pas construire.

Proposed patrickstox.com pilot: Schema Validator

The Balisage de données structurées Validator is a bon future pilot candidate, pas a live WebMCP implementation. It déjà has a bounded pasted input, deterministic logic, structured problèmes, and a visible result. A future validate_schema outil pourrait reuse the même validation engine as the UI and the existing remote MCP outil suite.

The pilot devrait wait jusqu’à le navigateur API reaches a stable, non-experimental release and passes a fresh security examiner. The premier version devrait be read-only, feature-detected, input-limited, response-minimized, and unavailable on preview, admin, or owner-only surfaces. Ce site fait pas currently claim WebMCP prise en charge.

Decision: construire, experiment, wait, or skip

AI summary

• Definition: WebMCP is a proposed navigateur API pour exposing structured, page-owned outils to an agent in an active browsing context.
• Status on July 17, 2026: Community Groupe draft, pas W3C Standard; Chrome 149 origin trial/local flag, pas stable cross-browser prise en charge.
• Current Chrome API: document.modelContext; Chrome deprecates navigator.modelContext beginning in Chrome 150.
• Pas remote MCP: WebMCP is tab- and page-state-bound; remote MCP commonly connects an AI application to a persistent backend server.
• Pas discovery or ranking: aucun documented benefit pour exploration, indexation, rankings, AI citations, or reaching lune page in the premier placer.
• Declarative caveat: Chrome documents formulaire annotations, pendant que the July 10 Community Groupe draft encore marks que normative section TODO.
• Security: authenticated navigateur state, prompt injection, poisoned metadata or output, broad inputs, and cross-origin exposure faire least privilege, confirmation, server authorization, and logs essential.
• Cloudflare: Navigateur Run peut consume/tester WebMCP in experimental lab sessions; it ne fait pas créer outils on le site.
• Recommendation: garder the human UI complet and wait pour stable prise en charge; private experiments devrait be bounded, reversible, and measurable.
• Ce site’s candidate: a future read-only Schema Validator pilot. It is proposed, pas implemented.

Ce que the principal sources establish

Status and compatibility were vérifié July 17, 2026. Recheck ces sources avant copying code or making a production decision.

WebMCP implementation checklist

Product fit

• Lune page has un bounded action que is materially plus reliable as a outil.
• The outil solves a utilisateur task; it n’est pas being ajouté pour speculative rankings.
• A normal semantic, accessible interface remains the source of truth.
• The action belongs to the current page/session plutôt que a background API.

Contract and lifecycle

• The nom and description are short, unique, spécifique, and pas generated from untrusted content.
• The JSON Schema rejects unknown fields and limites strings, arrays, ranges, and enums.
• The callback reuses the même engine and authorization chemin as the visible UI.
• The UI visibly reflects le résultating state.
• The outil is unregistered quand navigation or state rend it invalid.
• Unsupported navigateurs retain the complet workflow.

Security and privacy

• Lire/écrire behavior is explicit; annotations are treated as hints seulement.
• Server authorization is enforced independently of the agent.
• Sensitive actions exiger visible, informed confirmation.
• Cross-origin exposure is denied unless a named utiliser cas exige it.
• Outil output is minimized and treated as untrusted downstream input.
• Secrets, cookies, local history, raw private records, and unrelated state are excluded.
• Logs prise en charge incident examiner sans storing sensitive payloads.

Release and evaluation

• Registration, validation, errors, side effects, output, UI parity, and cleanup have deterministic tests.
• Representative agents have probabilistic selection and parameterization evals.
• Adversarial prompts, metadata, output, and confused-deputy cas are testé.
• The current spec, Chrome milestone, flags/trial, and vendor docs were rechecked.
• The fonctionnalité n’est pas a production dependency pendant que prise en charge remains experimental.

WebMCP decision cheat sheet

Current syntax and status

Current object:       document.modelContext
Deprecated in C150:  navigator.modelContext
Chrome status:        149 origin trial / testing flag
Standards status:     Community Group draft, not W3C Standard
Discovery moment:     after the client visits the page
Cloudflare role:      experimental consuming/test browser

The ACTION gate

Utiliser ACTION avant exposing une page function:

• A — Action is bounded. Un objectif, explicit input, predictable output.
• C — Context belongs to lune page. The current tab or visible state is genuinely requis; sinon préférer an API or remote MCP.
• T — Trust boundary is reviewed. Authentication, origins, untrusted content, confirmation, and output minimization have owners.
• I — Interface stays complet. The ordinary UI is accessible and entièrement functional sans experimental navigateur prise en charge.
• O — Un implementation. UI, API/MCP, and WebMCP appel the même canonical engine plutôt que drifting into separate business logic.
• N — Numbers peut be mesuré. Define task completion, error, clarification, abandonment, unsafe-attempt, and fallback rates avant launch.

Si quelconque letter fails, the outil n’est pas ready pour production.

WebMCP anti-patterns

1. Ajout a outil to every article. Readable content ne fait pas become plus discoverable parce que une page registers a meaningless read_article action.
2. Treating annotations as enforcement. readOnlyHint communicates intent; it ne fait pas prevent hidden writes or guarantee safe model behavior.
3. Copying navigator.modelContext exemples sans a date. Chrome documents the déplacer to document.modelContext and deprecation beginning in Chrome 150.
4. Giving the agent a generic command runner. Broad inputs erase the benefit of a structured contract and expand prompt-injection impact.
5. Leaving stale outils registered. Si lune page disables an action, its outil ne doit pas remain callable.
6. Returning the whole application object. Minimize output; private or untrusted fields peut leak données or poison plus tard reasoning.
7. Letting Cloudflare status substitute pour page fonctionner. A beta navigateur que consumes WebMCP ne fait pas author, register, secure, or tester votre outils.
8. Replacing the formulaire with the experiment. Progressive enhancement signifie the human interface survives manquant navigateur prise en charge and API changements.
9. Appel a private demo production-ready. Contract success n’est pas agent reliability, interoperability, or a completed threat model.

Connexe outils on ce site

• Balisage de données structurées Validator: the proposed premier WebMCP pilot parce que its inputs and findings are bounded and deterministic. WebMCP n’est pas enabled on it today.
• MCP Outil Suite: le site’s live remote MCP surface and the clearest façon to comparer a persistent server integration with a future page outil.
• Agent Readiness Checker: checks several server-visible agent signals. It intentionally ne doit pas score WebMCP parce que a server-side récupérer ne peut pas reliably observe outils registered in a live document.
• Balisage de données structurées Generator: un autre exemple of a deterministic human workflow que doit remain complet regardless of agent prise en charge.

Principal sources

• WebMCP Draft Community Groupe Report — current draft, API definitions, permissions, security, and explicit standards status.
• Chrome: WebMCP overview — origin-trial entry point, limitations, and testing setup.
• Chrome: Imperative API — current object nom, registration, lifecycle, testing, events, and cross-origin contrôle.
• Chrome: Declarative API — experimental formulaire annotations and submission behavior.
• Chrome: WebMCP outil security — annotations, origin exposure, and tool-provider safeguards.
• Chrome: Agent security considerations — malicious metadata, contaminated outputs, authenticated sessions, and defense in depth.
• Chrome: WebMCP meilleur practices — outil scope, descriptions, state, and reliability.
• Chrome: WebMCP evals — deterministic testing and probabilistic agent evaluation.
• Chrome: Quand to utiliser WebMCP and MCP — frontend/page context versus persistent backend boundaries.
• Chrome: Join the WebMCP origin trial — Chrome 149 experiment announcement.
• Chrome DevTools: Debug WebMCP outils — registered and invoked outil inspection.
• Lighthouse: Registered WebMCP outils — informational registered-tool audit.
• Cloudflare Navigateur Run: WebMCP — beta lab sessions and production limitations; API exemples on que page lag current Chrome documentation.

Connexe guides

• Model Context Protocol
• Agentic Search
• llms.txt
• Balisage de données structurées pour AI
• Pour AI agents — le site’s currently disponible agent-facing fichiers and remote MCP surface; it ne fait pas claim WebMCP prise en charge.

Compatibility and status claims were verified July 17, 2026. Recheck avant implementation parce que milestones, flags, API noms, and draft text peut modifier.

Journal des modifications